{
  "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": "1F029AC9BF4947639809E685706EAEB4",
  "traceEvents": [
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::embedding", "pid": 870, "tid": 870,
    "ts": 7429770599224.046, "dur": 116.058,
    "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": 870, "tid": 870,
    "ts": 7429770599261.501, "dur": 7.545,
    "args": {
      "External id": 2,"Record function id": 0, "finished": true, "Ev Idx": 1
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 870, "tid": 870,
    "ts": 7429770599262.657, "dur": 3.997,
    "args": {
      "External id": 3,"Record function id": 0, "finished": true, "Ev Idx": 2
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::index_select", "pid": 870, "tid": 870,
    "ts": 7429770599270.727, "dur": 63.356,
    "args": {
      "External id": 4,"Record function id": 0, "finished": true, "Ev Idx": 3
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770599286.723, "dur": 7.246,
    "args": {
      "External id": 5,"Record function id": 0, "finished": true, "Ev Idx": 4
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 870, "tid": 870,
    "ts": 7429770599297.929, "dur": 6.790,
    "args": {
      "External id": 6,"Record function id": 0, "finished": true, "Ev Idx": 5
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 870, "tid": 870,
    "ts": 7429770599334.640, "dur": 0.680,
    "args": {
      "External id": 7,"Record function id": 0, "finished": true, "Ev Idx": 6
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::arange", "pid": 870, "tid": 870,
    "ts": 7429770599370.871, "dur": 17.243,
    "args": {
      "External id": 8,"Record function id": 0, "finished": true, "Ev Idx": 7
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770599372.378, "dur": 1.081,
    "args": {
      "External id": 9,"Record function id": 0, "finished": true, "Ev Idx": 8
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::arange", "pid": 870, "tid": 870,
    "ts": 7429770599374.949, "dur": 12.891,
    "args": {
      "External id": 10,"Record function id": 0, "finished": true, "Ev Idx": 9
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 870, "tid": 870,
    "ts": 7429770599376.479, "dur": 2.022,
    "args": {
      "External id": 11,"Record function id": 0, "finished": true, "Ev Idx": 10
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770599399.913, "dur": 12.209,
    "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": 870, "tid": 870,
    "ts": 7429770599409.221, "dur": 1.259,
    "args": {
      "External id": 13,"Record function id": 0, "finished": true, "Ev Idx": 12
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::full", "pid": 870, "tid": 870,
    "ts": 7429770599451.850, "dur": 19.111,
    "args": {
      "External id": 14,"Record function id": 0, "finished": true, "Ev Idx": 13
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770599452.964, "dur": 3.077,
    "args": {
      "External id": 15,"Record function id": 0, "finished": true, "Ev Idx": 14
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 870, "tid": 870,
    "ts": 7429770599457.533, "dur": 13.140,
    "args": {
      "External id": 16,"Record function id": 0, "finished": true, "Ev Idx": 15
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::triu", "pid": 870, "tid": 870,
    "ts": 7429770599475.299, "dur": 12.571,
    "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": 870, "tid": 870,
    "ts": 7429770599494.055, "dur": 8.115,
    "args": {
      "External id": 18,"Record function id": 0, "finished": true, "Ev Idx": 17
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770599494.868, "dur": 0.724,
    "args": {
      "External id": 19,"Record function id": 0, "finished": true, "Ev Idx": 18
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::arange", "pid": 870, "tid": 870,
    "ts": 7429770599495.965, "dur": 5.964,
    "args": {
      "External id": 20,"Record function id": 0, "finished": true, "Ev Idx": 19
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 870, "tid": 870,
    "ts": 7429770599496.456, "dur": 0.950,
    "args": {
      "External id": 21,"Record function id": 0, "finished": true, "Ev Idx": 20
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770599506.338, "dur": 2.682,
    "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": 870, "tid": 870,
    "ts": 7429770599506.893, "dur": 1.933,
    "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": 870, "tid": 870,
    "ts": 7429770599511.789, "dur": 14.419,
    "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": 870, "tid": 870,
    "ts": 7429770599530.600, "dur": 13.015,
    "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": 870, "tid": 870,
    "ts": 7429770599548.393, "dur": 2.270,
    "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": 870, "tid": 870,
    "ts": 7429770599549.188, "dur": 0.889,
    "args": {
      "External id": 27,"Record function id": 0, "finished": true, "Ev Idx": 26
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770599551.203, "dur": 2.258,
    "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": 870, "tid": 870,
    "ts": 7429770599552.071, "dur": 0.625,
    "args": {
      "External id": 29,"Record function id": 0, "finished": true, "Ev Idx": 28
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770599561.822, "dur": 4.560,
    "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": 870, "tid": 870,
    "ts": 7429770599564.354, "dur": 0.460,
    "args": {
      "External id": 31,"Record function id": 0, "finished": true, "Ev Idx": 30
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770599567.284, "dur": 1.240,
    "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": 870, "tid": 870,
    "ts": 7429770599567.816, "dur": 0.249,
    "args": {
      "External id": 33,"Record function id": 0, "finished": true, "Ev Idx": 32
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770599579.466, "dur": 2.923,
    "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": 870, "tid": 870,
    "ts": 7429770599581.491, "dur": 0.299,
    "args": {
      "External id": 35,"Record function id": 0, "finished": true, "Ev Idx": 34
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 870, "tid": 870,
    "ts": 7429770599592.588, "dur": 24.320,
    "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": 870, "tid": 870,
    "ts": 7429770599594.878, "dur": 2.775,
    "args": {
      "External id": 37,"Record function id": 0, "finished": true, "Ev Idx": 36
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770599598.971, "dur": 17.620,
    "args": {
      "External id": 38,"Record function id": 0, "finished": true, "Ev Idx": 37
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770599621.312, "dur": 1.856,
    "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": 870, "tid": 870,
    "ts": 7429770599621.899, "dur": 0.729,
    "args": {
      "External id": 40,"Record function id": 0, "finished": true, "Ev Idx": 39
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770599623.835, "dur": 1.161,
    "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": 870, "tid": 870,
    "ts": 7429770599624.368, "dur": 0.218,
    "args": {
      "External id": 42,"Record function id": 0, "finished": true, "Ev Idx": 41
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770599625.613, "dur": 0.946,
    "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": 870, "tid": 870,
    "ts": 7429770599626.030, "dur": 0.198,
    "args": {
      "External id": 44,"Record function id": 0, "finished": true, "Ev Idx": 43
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770599628.753, "dur": 1.215,
    "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": 870, "tid": 870,
    "ts": 7429770599629.234, "dur": 0.338,
    "args": {
      "External id": 46,"Record function id": 0, "finished": true, "Ev Idx": 45
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770599630.418, "dur": 1.458,
    "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": 870, "tid": 870,
    "ts": 7429770599630.983, "dur": 0.282,
    "args": {
      "External id": 48,"Record function id": 0, "finished": true, "Ev Idx": 47
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770599632.263, "dur": 1.133,
    "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": 870, "tid": 870,
    "ts": 7429770599632.710, "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": 870, "tid": 870,
    "ts": 7429770599633.883, "dur": 0.997,
    "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": 870, "tid": 870,
    "ts": 7429770599634.300, "dur": 0.211,
    "args": {
      "External id": 52,"Record function id": 0, "finished": true, "Ev Idx": 51
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770599640.462, "dur": 0.412,
    "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": 870, "tid": 870,
    "ts": 7429770599644.376, "dur": 13.425,
    "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": 870, "tid": 870,
    "ts": 7429770599661.876, "dur": 15.539,
    "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": 870, "tid": 870,
    "ts": 7429770599679.947, "dur": 1.428,
    "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": 870, "tid": 870,
    "ts": 7429770599680.598, "dur": 0.301,
    "args": {
      "External id": 57,"Record function id": 0, "finished": true, "Ev Idx": 56
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770599681.979, "dur": 1.350,
    "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": 870, "tid": 870,
    "ts": 7429770599682.717, "dur": 0.208,
    "args": {
      "External id": 59,"Record function id": 0, "finished": true, "Ev Idx": 58
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770599683.927, "dur": 0.971,
    "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": 870, "tid": 870,
    "ts": 7429770599684.354, "dur": 0.210,
    "args": {
      "External id": 61,"Record function id": 0, "finished": true, "Ev Idx": 60
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill", "pid": 870, "tid": 870,
    "ts": 7429770599689.989, "dur": 35.796,
    "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": 870, "tid": 870,
    "ts": 7429770599692.266, "dur": 18.503,
    "args": {
      "External id": 63,"Record function id": 0, "finished": true, "Ev Idx": 62
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 870, "tid": 870,
    "ts": 7429770599699.783, "dur": 2.772,
    "args": {
      "External id": 64,"Record function id": 0, "finished": true, "Ev Idx": 63
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770599700.603, "dur": 1.664,
    "args": {
      "External id": 65,"Record function id": 0, "finished": true, "Ev Idx": 64
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770599703.076, "dur": 7.508,
    "args": {
      "External id": 66,"Record function id": 0, "finished": true, "Ev Idx": 65
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 870, "tid": 870,
    "ts": 7429770599717.561, "dur": 7.878,
    "args": {
      "External id": 67,"Record function id": 0, "finished": true, "Ev Idx": 66
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770599728.589, "dur": 1.333,
    "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": 870, "tid": 870,
    "ts": 7429770599729.193, "dur": 0.271,
    "args": {
      "External id": 69,"Record function id": 0, "finished": true, "Ev Idx": 68
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770599730.484, "dur": 2.106,
    "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": 870, "tid": 870,
    "ts": 7429770599730.978, "dur": 1.240,
    "args": {
      "External id": 71,"Record function id": 0, "finished": true, "Ev Idx": 70
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770599733.249, "dur": 1.156,
    "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": 870, "tid": 870,
    "ts": 7429770599733.653, "dur": 0.411,
    "args": {
      "External id": 73,"Record function id": 0, "finished": true, "Ev Idx": 72
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770599735.499, "dur": 5.880,
    "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": 870, "tid": 870,
    "ts": 7429770599794.544, "dur": 1.720,
    "args": {
      "External id": 75,"Record function id": 0, "finished": true, "Ev Idx": 74
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770599795.447, "dur": 0.345,
    "args": {
      "External id": 76,"Record function id": 0, "finished": true, "Ev Idx": 75
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770599796.924, "dur": 1.402,
    "args": {
      "External id": 77,"Record function id": 0, "finished": true, "Ev Idx": 76
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770599797.514, "dur": 0.284,
    "args": {
      "External id": 78,"Record function id": 0, "finished": true, "Ev Idx": 77
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770599798.825, "dur": 1.075,
    "args": {
      "External id": 79,"Record function id": 0, "finished": true, "Ev Idx": 78
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770599799.288, "dur": 0.284,
    "args": {
      "External id": 80,"Record function id": 0, "finished": true, "Ev Idx": 79
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770599803.693, "dur": 0.291,
    "args": {
      "External id": 81,"Record function id": 0, "finished": true, "Ev Idx": 80
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770599806.702, "dur": 1.576,
    "args": {
      "External id": 82,"Record function id": 0, "finished": true, "Ev Idx": 81
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770599807.652, "dur": 0.250,
    "args": {
      "External id": 83,"Record function id": 0, "finished": true, "Ev Idx": 82
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770599810.616, "dur": 0.151,
    "args": {
      "External id": 84,"Record function id": 0, "finished": true, "Ev Idx": 83
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770599812.483, "dur": 1.207,
    "args": {
      "External id": 85,"Record function id": 0, "finished": true, "Ev Idx": 84
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770599813.022, "dur": 0.257,
    "args": {
      "External id": 86,"Record function id": 0, "finished": true, "Ev Idx": 85
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770599814.143, "dur": 1.090,
    "args": {
      "External id": 87,"Record function id": 0, "finished": true, "Ev Idx": 86
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770599814.604, "dur": 0.334,
    "args": {
      "External id": 88,"Record function id": 0, "finished": true, "Ev Idx": 87
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770599815.793, "dur": 1.111,
    "args": {
      "External id": 89,"Record function id": 0, "finished": true, "Ev Idx": 88
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770599816.259, "dur": 0.292,
    "args": {
      "External id": 90,"Record function id": 0, "finished": true, "Ev Idx": 89
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770599818.676, "dur": 15.188,
    "args": {
      "External id": 91,"Record function id": 0, "finished": true, "Ev Idx": 90
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 870, "tid": 870,
    "ts": 7429770599819.771, "dur": 13.963,
    "args": {
      "External id": 92,"Record function id": 0, "finished": true, "Ev Idx": 91
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 870, "tid": 870,
    "ts": 7429770599822.581, "dur": 2.026,
    "args": {
      "External id": 93,"Record function id": 0, "finished": true, "Ev Idx": 92
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770599825.122, "dur": 8.270,
    "args": {
      "External id": 94,"Record function id": 0, "finished": true, "Ev Idx": 93
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770599888.959, "dur": 0.145,
    "args": {
      "External id": 95,"Record function id": 0, "finished": true, "Ev Idx": 94
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770599890.247, "dur": 0.103,
    "args": {
      "External id": 96,"Record function id": 0, "finished": true, "Ev Idx": 95
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770599892.052, "dur": 75.825,
    "args": {
      "External id": 97,"Record function id": 0, "finished": true, "Ev Idx": 96
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770599893.434, "dur": 1.344,
    "args": {
      "External id": 98,"Record function id": 0, "finished": true, "Ev Idx": 97
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770599894.037, "dur": 0.367,
    "args": {
      "External id": 99,"Record function id": 0, "finished": true, "Ev Idx": 98
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770599895.288, "dur": 1.558,
    "args": {
      "External id": 100,"Record function id": 0, "finished": true, "Ev Idx": 99
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 870, "tid": 870,
    "ts": 7429770599895.628, "dur": 1.110,
    "args": {
      "External id": 101,"Record function id": 0, "finished": true, "Ev Idx": 100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770599897.389, "dur": 0.935,
    "args": {
      "External id": 102,"Record function id": 0, "finished": true, "Ev Idx": 101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770599897.822, "dur": 0.196,
    "args": {
      "External id": 103,"Record function id": 0, "finished": true, "Ev Idx": 102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770599898.608, "dur": 1.351,
    "args": {
      "External id": 104,"Record function id": 0, "finished": true, "Ev Idx": 103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 870, "tid": 870,
    "ts": 7429770599898.863, "dur": 1.022,
    "args": {
      "External id": 105,"Record function id": 0, "finished": true, "Ev Idx": 104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 870, "tid": 870,
    "ts": 7429770599901.043, "dur": 62.797,
    "args": {
      "External id": 106,"Record function id": 0, "finished": true, "Ev Idx": 105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770599925.122, "dur": 1.769,
    "args": {
      "External id": 107,"Record function id": 0, "finished": true, "Ev Idx": 106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770599965.529, "dur": 1.656,
    "args": {
      "External id": 108,"Record function id": 0, "finished": true, "Ev Idx": 107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 870, "tid": 870,
    "ts": 7429770599973.491, "dur": 3.121,
    "args": {
      "External id": 109,"Record function id": 0, "finished": true, "Ev Idx": 108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770599975.160, "dur": 0.829,
    "args": {
      "External id": 110,"Record function id": 0, "finished": true, "Ev Idx": 109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 870, "tid": 870,
    "ts": 7429770599983.047, "dur": 17.529,
    "args": {
      "External id": 111,"Record function id": 0, "finished": true, "Ev Idx": 110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cos", "pid": 870, "tid": 870,
    "ts": 7429770600005.699, "dur": 10.581,
    "args": {
      "External id": 112,"Record function id": 0, "finished": true, "Ev Idx": 111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770600021.200, "dur": 11.691,
    "args": {
      "External id": 113,"Record function id": 0, "finished": true, "Ev Idx": 112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sin", "pid": 870, "tid": 870,
    "ts": 7429770600037.424, "dur": 8.977,
    "args": {
      "External id": 114,"Record function id": 0, "finished": true, "Ev Idx": 113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770600048.358, "dur": 5.894,
    "args": {
      "External id": 115,"Record function id": 0, "finished": true, "Ev Idx": 114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770600069.503, "dur": 0.150,
    "args": {
      "External id": 116,"Record function id": 0, "finished": true, "Ev Idx": 115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770600071.348, "dur": 0.111,
    "args": {
      "External id": 117,"Record function id": 0, "finished": true, "Ev Idx": 116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770600255.305, "dur": 0.301,
    "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": 870, "tid": 870,
    "ts": 7429770600259.277, "dur": 23.215,
    "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": 870, "tid": 870,
    "ts": 7429770600267.207, "dur": 0.497,
    "args": {
      "External id": 120,"Record function id": 0, "finished": true, "Ev Idx": 119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770600268.324, "dur": 0.204,
    "args": {
      "External id": 121,"Record function id": 0, "finished": true, "Ev Idx": 120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770600295.483, "dur": 18.783,
    "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": 870, "tid": 870,
    "ts": 7429770600319.718, "dur": 11.975,
    "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": 870, "tid": 870,
    "ts": 7429770600335.630, "dur": 14.667,
    "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": 870, "tid": 870,
    "ts": 7429770600352.190, "dur": 9.738,
    "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": 870, "tid": 870,
    "ts": 7429770600365.347, "dur": 0.234,
    "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": 870, "tid": 870,
    "ts": 7429770600366.495, "dur": 8.968,
    "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": 870, "tid": 870,
    "ts": 7429770600409.442, "dur": 78.232,
    "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": 870, "tid": 870,
    "ts": 7429770600411.033, "dur": 12.848,
    "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": 870, "tid": 870,
    "ts": 7429770600413.665, "dur": 1.945,
    "args": {
      "External id": 130,"Record function id": 0, "finished": true, "Ev Idx": 129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770600414.387, "dur": 0.896,
    "args": {
      "External id": 131,"Record function id": 0, "finished": true, "Ev Idx": 130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770600424.619, "dur": 62.807,
    "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": 870, "tid": 870,
    "ts": 7429770600425.639, "dur": 2.877,
    "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": 870, "tid": 870,
    "ts": 7429770600426.098, "dur": 2.257,
    "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": 870, "tid": 870,
    "ts": 7429770600429.494, "dur": 54.185,
    "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": 870, "tid": 870,
    "ts": 7429770600454.908, "dur": 2.042,
    "args": {
      "External id": 136,"Record function id": 0, "finished": true, "Ev Idx": 135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770600484.720, "dur": 2.057,
    "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": 870, "tid": 870,
    "ts": 7429770600492.950, "dur": 1.820,
    "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": 870, "tid": 870,
    "ts": 7429770600496.830, "dur": 3.086,
    "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": 870, "tid": 870,
    "ts": 7429770600498.296, "dur": 0.739,
    "args": {
      "External id": 140,"Record function id": 0, "finished": true, "Ev Idx": 139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770600510.637, "dur": 39.484,
    "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": 870, "tid": 870,
    "ts": 7429770600511.077, "dur": 3.421,
    "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": 870, "tid": 870,
    "ts": 7429770600512.262, "dur": 1.699,
    "args": {
      "External id": 143,"Record function id": 0, "finished": true, "Ev Idx": 142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770600513.473, "dur": 0.371,
    "args": {
      "External id": 144,"Record function id": 0, "finished": true, "Ev Idx": 143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770600514.888, "dur": 35.087,
    "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": 870, "tid": 870,
    "ts": 7429770600515.293, "dur": 1.817,
    "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": 870, "tid": 870,
    "ts": 7429770600515.544, "dur": 1.444,
    "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": 870, "tid": 870,
    "ts": 7429770600517.508, "dur": 30.028,
    "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": 870, "tid": 870,
    "ts": 7429770600529.458, "dur": 1.418,
    "args": {
      "External id": 149,"Record function id": 0, "finished": true, "Ev Idx": 148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770600548.351, "dur": 1.278,
    "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": 870, "tid": 870,
    "ts": 7429770600552.730, "dur": 1.270,
    "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": 870, "tid": 870,
    "ts": 7429770600555.911, "dur": 2.010,
    "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": 870, "tid": 870,
    "ts": 7429770600556.755, "dur": 0.622,
    "args": {
      "External id": 153,"Record function id": 0, "finished": true, "Ev Idx": 152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770600566.346, "dur": 32.757,
    "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": 870, "tid": 870,
    "ts": 7429770600566.681, "dur": 2.460,
    "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": 870, "tid": 870,
    "ts": 7429770600567.564, "dur": 1.064,
    "args": {
      "External id": 156,"Record function id": 0, "finished": true, "Ev Idx": 155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770600568.264, "dur": 0.269,
    "args": {
      "External id": 157,"Record function id": 0, "finished": true, "Ev Idx": 156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770600569.434, "dur": 29.570,
    "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": 870, "tid": 870,
    "ts": 7429770600570.029, "dur": 1.401,
    "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": 870, "tid": 870,
    "ts": 7429770600570.296, "dur": 1.001,
    "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": 870, "tid": 870,
    "ts": 7429770600571.764, "dur": 25.129,
    "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": 870, "tid": 870,
    "ts": 7429770600582.297, "dur": 1.078,
    "args": {
      "External id": 162,"Record function id": 0, "finished": true, "Ev Idx": 161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770600597.573, "dur": 1.174,
    "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": 870, "tid": 870,
    "ts": 7429770600601.548, "dur": 1.091,
    "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": 870, "tid": 870,
    "ts": 7429770600604.320, "dur": 1.814,
    "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": 870, "tid": 870,
    "ts": 7429770600605.109, "dur": 0.484,
    "args": {
      "External id": 166,"Record function id": 0, "finished": true, "Ev Idx": 165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770600610.715, "dur": 2.101,
    "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": 870, "tid": 870,
    "ts": 7429770600611.585, "dur": 0.712,
    "args": {
      "External id": 168,"Record function id": 0, "finished": true, "Ev Idx": 167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770600614.293, "dur": 1.326,
    "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": 870, "tid": 870,
    "ts": 7429770600614.762, "dur": 0.334,
    "args": {
      "External id": 170,"Record function id": 0, "finished": true, "Ev Idx": 169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770600616.743, "dur": 11.631,
    "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": 870, "tid": 870,
    "ts": 7429770600635.260, "dur": 4.917,
    "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": 870, "tid": 870,
    "ts": 7429770600638.321, "dur": 0.449,
    "args": {
      "External id": 173,"Record function id": 0, "finished": true, "Ev Idx": 172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770600642.405, "dur": 2.749,
    "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": 870, "tid": 870,
    "ts": 7429770600643.884, "dur": 0.555,
    "args": {
      "External id": 175,"Record function id": 0, "finished": true, "Ev Idx": 174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770600647.385, "dur": 10.953,
    "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": 870, "tid": 870,
    "ts": 7429770600662.654, "dur": 18.602,
    "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": 870, "tid": 870,
    "ts": 7429770600684.474, "dur": 9.493,
    "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": 870, "tid": 870,
    "ts": 7429770600696.121, "dur": 9.412,
    "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": 870, "tid": 870,
    "ts": 7429770600707.906, "dur": 7.972,
    "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": 870, "tid": 870,
    "ts": 7429770600718.343, "dur": 2.773,
    "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": 870, "tid": 870,
    "ts": 7429770600720.007, "dur": 0.442,
    "args": {
      "External id": 182,"Record function id": 0, "finished": true, "Ev Idx": 181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770600722.756, "dur": 2.726,
    "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": 870, "tid": 870,
    "ts": 7429770600724.403, "dur": 0.525,
    "args": {
      "External id": 184,"Record function id": 0, "finished": true, "Ev Idx": 183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770600726.506, "dur": 6.514,
    "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": 870, "tid": 870,
    "ts": 7429770600735.163, "dur": 10.389,
    "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": 870, "tid": 870,
    "ts": 7429770600748.095, "dur": 7.531,
    "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": 870, "tid": 870,
    "ts": 7429770600757.092, "dur": 5.865,
    "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": 870, "tid": 870,
    "ts": 7429770600796.281, "dur": 3.275,
    "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": 870, "tid": 870,
    "ts": 7429770600798.127, "dur": 0.571,
    "args": {
      "External id": 190,"Record function id": 0, "finished": true, "Ev Idx": 189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770600800.327, "dur": 2.587,
    "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": 870, "tid": 870,
    "ts": 7429770600801.876, "dur": 0.468,
    "args": {
      "External id": 192,"Record function id": 0, "finished": true, "Ev Idx": 191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770600803.690, "dur": 3.166,
    "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": 870, "tid": 870,
    "ts": 7429770600804.823, "dur": 0.823,
    "args": {
      "External id": 194,"Record function id": 0, "finished": true, "Ev Idx": 193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770600807.485, "dur": 2.158,
    "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": 870, "tid": 870,
    "ts": 7429770600808.591, "dur": 0.498,
    "args": {
      "External id": 196,"Record function id": 0, "finished": true, "Ev Idx": 195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770600810.259, "dur": 2.369,
    "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": 870, "tid": 870,
    "ts": 7429770600811.691, "dur": 0.407,
    "args": {
      "External id": 198,"Record function id": 0, "finished": true, "Ev Idx": 197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770600815.926, "dur": 3.600,
    "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": 870, "tid": 870,
    "ts": 7429770600818.215, "dur": 0.572,
    "args": {
      "External id": 200,"Record function id": 0, "finished": true, "Ev Idx": 199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770600822.147, "dur": 23.537,
    "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": 870, "tid": 870,
    "ts": 7429770600825.297, "dur": 17.696,
    "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": 870, "tid": 870,
    "ts": 7429770600827.784, "dur": 2.956,
    "args": {
      "External id": 203,"Record function id": 0, "finished": true, "Ev Idx": 202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770600828.660, "dur": 1.771,
    "args": {
      "External id": 204,"Record function id": 0, "finished": true, "Ev Idx": 203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770600831.727, "dur": 10.249,
    "args": {
      "External id": 205,"Record function id": 0, "finished": true, "Ev Idx": 204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770600843.581, "dur": 1.794,
    "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": 870, "tid": 870,
    "ts": 7429770600849.078, "dur": 2.034,
    "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": 870, "tid": 870,
    "ts": 7429770600849.894, "dur": 0.647,
    "args": {
      "External id": 208,"Record function id": 0, "finished": true, "Ev Idx": 207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770600851.726, "dur": 1.699,
    "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": 870, "tid": 870,
    "ts": 7429770600852.577, "dur": 0.380,
    "args": {
      "External id": 210,"Record function id": 0, "finished": true, "Ev Idx": 209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770600853.926, "dur": 1.752,
    "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": 870, "tid": 870,
    "ts": 7429770600854.862, "dur": 0.353,
    "args": {
      "External id": 212,"Record function id": 0, "finished": true, "Ev Idx": 211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770600856.239, "dur": 2.266,
    "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": 870, "tid": 870,
    "ts": 7429770600857.566, "dur": 0.407,
    "args": {
      "External id": 214,"Record function id": 0, "finished": true, "Ev Idx": 213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770600859.143, "dur": 1.760,
    "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": 870, "tid": 870,
    "ts": 7429770600860.158, "dur": 0.278,
    "args": {
      "External id": 216,"Record function id": 0, "finished": true, "Ev Idx": 215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770600862.531, "dur": 1.803,
    "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": 870, "tid": 870,
    "ts": 7429770600863.568, "dur": 0.327,
    "args": {
      "External id": 218,"Record function id": 0, "finished": true, "Ev Idx": 217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770600865.922, "dur": 10.748,
    "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": 870, "tid": 870,
    "ts": 7429770600866.916, "dur": 8.102,
    "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": 870, "tid": 870,
    "ts": 7429770600867.544, "dur": 1.892,
    "args": {
      "External id": 221,"Record function id": 0, "finished": true, "Ev Idx": 220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770600868.065, "dur": 1.241,
    "args": {
      "External id": 222,"Record function id": 0, "finished": true, "Ev Idx": 221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770600869.744, "dur": 4.960,
    "args": {
      "External id": 223,"Record function id": 0, "finished": true, "Ev Idx": 222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770600875.364, "dur": 1.141,
    "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": 870, "tid": 870,
    "ts": 7429770600879.197, "dur": 2.156,
    "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": 870, "tid": 870,
    "ts": 7429770600880.316, "dur": 0.408,
    "args": {
      "External id": 226,"Record function id": 0, "finished": true, "Ev Idx": 225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770600884.469, "dur": 58.549,
    "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": 870, "tid": 870,
    "ts": 7429770600885.184, "dur": 2.238,
    "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": 870, "tid": 870,
    "ts": 7429770600886.516, "dur": 0.464,
    "args": {
      "External id": 229,"Record function id": 0, "finished": true, "Ev Idx": 228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770600887.741, "dur": 12.138,
    "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": 870, "tid": 870,
    "ts": 7429770600897.308, "dur": 2.417,
    "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": 870, "tid": 870,
    "ts": 7429770600900.442, "dur": 1.370,
    "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": 870, "tid": 870,
    "ts": 7429770600901.111, "dur": 0.299,
    "args": {
      "External id": 233,"Record function id": 0, "finished": true, "Ev Idx": 232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770600902.111, "dur": 2.096,
    "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": 870, "tid": 870,
    "ts": 7429770600903.056, "dur": 1.004,
    "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": 870, "tid": 870,
    "ts": 7429770600905.080, "dur": 35.432,
    "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": 870, "tid": 870,
    "ts": 7429770600919.432, "dur": 1.431,
    "args": {
      "External id": 237,"Record function id": 0, "finished": true, "Ev Idx": 236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770600941.168, "dur": 1.483,
    "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": 870, "tid": 870,
    "ts": 7429770600946.292, "dur": 9.352,
    "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": 870, "tid": 870,
    "ts": 7429770600959.179, "dur": 2.045,
    "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": 870, "tid": 870,
    "ts": 7429770600960.048, "dur": 0.477,
    "args": {
      "External id": 241,"Record function id": 0, "finished": true, "Ev Idx": 240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770600961.887, "dur": 1.739,
    "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": 870, "tid": 870,
    "ts": 7429770600962.424, "dur": 0.823,
    "args": {
      "External id": 243,"Record function id": 0, "finished": true, "Ev Idx": 242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770600964.346, "dur": 1.247,
    "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": 870, "tid": 870,
    "ts": 7429770600964.917, "dur": 0.343,
    "args": {
      "External id": 245,"Record function id": 0, "finished": true, "Ev Idx": 244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770600967.239, "dur": 7.780,
    "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": 870, "tid": 870,
    "ts": 7429770600985.633, "dur": 14.989,
    "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": 870, "tid": 870,
    "ts": 7429770600986.882, "dur": 0.341,
    "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": 870, "tid": 870,
    "ts": 7429770600988.741, "dur": 11.595,
    "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": 870, "tid": 870,
    "ts": 7429770601003.666, "dur": 0.225,
    "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": 870, "tid": 870,
    "ts": 7429770601020.387, "dur": 0.615,
    "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": 870, "tid": 870,
    "ts": 7429770601023.330, "dur": 44.837,
    "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": 870, "tid": 870,
    "ts": 7429770601023.965, "dur": 2.122,
    "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": 870, "tid": 870,
    "ts": 7429770601025.063, "dur": 0.441,
    "args": {
      "External id": 254,"Record function id": 0, "finished": true, "Ev Idx": 253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770601026.446, "dur": 2.768,
    "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": 870, "tid": 870,
    "ts": 7429770601026.874, "dur": 2.221,
    "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": 870, "tid": 870,
    "ts": 7429770601029.675, "dur": 1.920,
    "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": 870, "tid": 870,
    "ts": 7429770601030.896, "dur": 0.329,
    "args": {
      "External id": 258,"Record function id": 0, "finished": true, "Ev Idx": 257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770601031.888, "dur": 1.434,
    "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": 870, "tid": 870,
    "ts": 7429770601032.270, "dur": 0.970,
    "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": 870, "tid": 870,
    "ts": 7429770601033.664, "dur": 32.417,
    "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": 870, "tid": 870,
    "ts": 7429770601046.305, "dur": 1.438,
    "args": {
      "External id": 262,"Record function id": 0, "finished": true, "Ev Idx": 261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770601066.729, "dur": 1.154,
    "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": 870, "tid": 870,
    "ts": 7429770601070.382, "dur": 2.036,
    "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": 870, "tid": 870,
    "ts": 7429770601071.382, "dur": 0.430,
    "args": {
      "External id": 265,"Record function id": 0, "finished": true, "Ev Idx": 264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770601076.941, "dur": 12.185,
    "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": 870, "tid": 870,
    "ts": 7429770601077.480, "dur": 11.515,
    "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": 870, "tid": 870,
    "ts": 7429770601078.688, "dur": 1.970,
    "args": {
      "External id": 268,"Record function id": 0, "finished": true, "Ev Idx": 267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770601079.088, "dur": 1.394,
    "args": {
      "External id": 269,"Record function id": 0, "finished": true, "Ev Idx": 268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770601081.007, "dur": 7.395,
    "args": {
      "External id": 270,"Record function id": 0, "finished": true, "Ev Idx": 269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770601095.260, "dur": 1.977,
    "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": 870, "tid": 870,
    "ts": 7429770601095.606, "dur": 1.526,
    "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": 870, "tid": 870,
    "ts": 7429770601114.551, "dur": 40.918,
    "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": 870, "tid": 870,
    "ts": 7429770601115.405, "dur": 3.629,
    "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": 870, "tid": 870,
    "ts": 7429770601117.237, "dur": 1.224,
    "args": {
      "External id": 275,"Record function id": 0, "finished": true, "Ev Idx": 274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770601117.855, "dur": 0.458,
    "args": {
      "External id": 276,"Record function id": 0, "finished": true, "Ev Idx": 275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770601119.444, "dur": 35.851,
    "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": 870, "tid": 870,
    "ts": 7429770601119.947, "dur": 2.635,
    "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": 870, "tid": 870,
    "ts": 7429770601120.297, "dur": 2.207,
    "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": 870, "tid": 870,
    "ts": 7429770601122.973, "dur": 30.153,
    "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": 870, "tid": 870,
    "ts": 7429770601136.304, "dur": 1.619,
    "args": {
      "External id": 281,"Record function id": 0, "finished": true, "Ev Idx": 280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770601153.775, "dur": 1.214,
    "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": 870, "tid": 870,
    "ts": 7429770601161.056, "dur": 8.052,
    "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": 870, "tid": 870,
    "ts": 7429770601179.998, "dur": 0.182,
    "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": 870, "tid": 870,
    "ts": 7429770601182.541, "dur": 12.389,
    "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": 870, "tid": 870,
    "ts": 7429770601185.204, "dur": 0.348,
    "args": {
      "External id": 286,"Record function id": 0, "finished": true, "Ev Idx": 285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770601185.919, "dur": 0.081,
    "args": {
      "External id": 287,"Record function id": 0, "finished": true, "Ev Idx": 286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770601198.107, "dur": 12.879,
    "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": 870, "tid": 870,
    "ts": 7429770601214.796, "dur": 8.610,
    "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": 870, "tid": 870,
    "ts": 7429770601225.627, "dur": 10.113,
    "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": 870, "tid": 870,
    "ts": 7429770601237.590, "dur": 7.740,
    "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": 870, "tid": 870,
    "ts": 7429770601248.284, "dur": 0.118,
    "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": 870, "tid": 870,
    "ts": 7429770601249.369, "dur": 8.426,
    "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": 870, "tid": 870,
    "ts": 7429770601278.476, "dur": 42.468,
    "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": 870, "tid": 870,
    "ts": 7429770601278.955, "dur": 3.203,
    "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": 870, "tid": 870,
    "ts": 7429770601280.102, "dur": 1.352,
    "args": {
      "External id": 296,"Record function id": 0, "finished": true, "Ev Idx": 295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770601280.764, "dur": 0.521,
    "args": {
      "External id": 297,"Record function id": 0, "finished": true, "Ev Idx": 296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770601282.500, "dur": 38.307,
    "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": 870, "tid": 870,
    "ts": 7429770601283.034, "dur": 2.975,
    "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": 870, "tid": 870,
    "ts": 7429770601283.695, "dur": 2.192,
    "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": 870, "tid": 870,
    "ts": 7429770601286.445, "dur": 32.298,
    "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": 870, "tid": 870,
    "ts": 7429770601300.298, "dur": 1.961,
    "args": {
      "External id": 302,"Record function id": 0, "finished": true, "Ev Idx": 301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770601319.379, "dur": 1.120,
    "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": 870, "tid": 870,
    "ts": 7429770601335.305, "dur": 10.288,
    "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": 870, "tid": 870,
    "ts": 7429770601353.490, "dur": 33.566,
    "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": 870, "tid": 870,
    "ts": 7429770601353.901, "dur": 3.289,
    "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": 870, "tid": 870,
    "ts": 7429770601354.945, "dur": 1.182,
    "args": {
      "External id": 307,"Record function id": 0, "finished": true, "Ev Idx": 306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770601355.502, "dur": 0.487,
    "args": {
      "External id": 308,"Record function id": 0, "finished": true, "Ev Idx": 307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770601357.512, "dur": 29.408,
    "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": 870, "tid": 870,
    "ts": 7429770601357.945, "dur": 2.426,
    "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": 870, "tid": 870,
    "ts": 7429770601358.366, "dur": 1.871,
    "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": 870, "tid": 870,
    "ts": 7429770601360.677, "dur": 24.352,
    "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": 870, "tid": 870,
    "ts": 7429770601370.715, "dur": 1.220,
    "args": {
      "External id": 313,"Record function id": 0, "finished": true, "Ev Idx": 312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770601385.643, "dur": 1.010,
    "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": 870, "tid": 870,
    "ts": 7429770601388.603, "dur": 7.988,
    "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": 870, "tid": 870,
    "ts": 7429770601406.683, "dur": 45.939,
    "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": 870, "tid": 870,
    "ts": 7429770601407.044, "dur": 2.532,
    "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": 870, "tid": 870,
    "ts": 7429770601408.266, "dur": 0.857,
    "args": {
      "External id": 318,"Record function id": 0, "finished": true, "Ev Idx": 317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770601408.678, "dur": 0.293,
    "args": {
      "External id": 319,"Record function id": 0, "finished": true, "Ev Idx": 318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770601409.900, "dur": 42.595,
    "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": 870, "tid": 870,
    "ts": 7429770601410.320, "dur": 1.707,
    "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": 870, "tid": 870,
    "ts": 7429770601410.808, "dur": 1.114,
    "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": 870, "tid": 870,
    "ts": 7429770601412.328, "dur": 38.163,
    "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": 870, "tid": 870,
    "ts": 7429770601422.760, "dur": 1.301,
    "args": {
      "External id": 324,"Record function id": 0, "finished": true, "Ev Idx": 323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770601451.146, "dur": 1.083,
    "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": 870, "tid": 870,
    "ts": 7429770601455.236, "dur": 7.307,
    "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": 870, "tid": 870,
    "ts": 7429770601483.198, "dur": 0.248,
    "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": 870, "tid": 870,
    "ts": 7429770601485.228, "dur": 9.572,
    "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": 870, "tid": 870,
    "ts": 7429770601487.160, "dur": 0.221,
    "args": {
      "External id": 329,"Record function id": 0, "finished": true, "Ev Idx": 328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770601487.717, "dur": 0.085,
    "args": {
      "External id": 330,"Record function id": 0, "finished": true, "Ev Idx": 329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770601497.183, "dur": 12.136,
    "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": 870, "tid": 870,
    "ts": 7429770601512.347, "dur": 14.219,
    "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": 870, "tid": 870,
    "ts": 7429770601528.287, "dur": 8.489,
    "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": 870, "tid": 870,
    "ts": 7429770601538.421, "dur": 10.002,
    "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": 870, "tid": 870,
    "ts": 7429770601550.874, "dur": 0.105,
    "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": 870, "tid": 870,
    "ts": 7429770601551.811, "dur": 8.140,
    "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": 870, "tid": 870,
    "ts": 7429770601578.848, "dur": 39.598,
    "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": 870, "tid": 870,
    "ts": 7429770601579.484, "dur": 4.444,
    "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": 870, "tid": 870,
    "ts": 7429770601581.807, "dur": 1.300,
    "args": {
      "External id": 339,"Record function id": 0, "finished": true, "Ev Idx": 338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770601582.365, "dur": 0.580,
    "args": {
      "External id": 340,"Record function id": 0, "finished": true, "Ev Idx": 339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770601584.367, "dur": 33.962,
    "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": 870, "tid": 870,
    "ts": 7429770601584.962, "dur": 1.714,
    "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": 870, "tid": 870,
    "ts": 7429770601585.446, "dur": 1.091,
    "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": 870, "tid": 870,
    "ts": 7429770601587.081, "dur": 29.131,
    "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": 870, "tid": 870,
    "ts": 7429770601597.826, "dur": 1.297,
    "args": {
      "External id": 345,"Record function id": 0, "finished": true, "Ev Idx": 344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770601616.893, "dur": 1.099,
    "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": 870, "tid": 870,
    "ts": 7429770601621.500, "dur": 1.407,
    "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": 870, "tid": 870,
    "ts": 7429770601624.701, "dur": 2.415,
    "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": 870, "tid": 870,
    "ts": 7429770601625.749, "dur": 0.584,
    "args": {
      "External id": 349,"Record function id": 0, "finished": true, "Ev Idx": 348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770601634.958, "dur": 33.680,
    "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": 870, "tid": 870,
    "ts": 7429770601635.228, "dur": 2.513,
    "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": 870, "tid": 870,
    "ts": 7429770601636.214, "dur": 1.038,
    "args": {
      "External id": 352,"Record function id": 0, "finished": true, "Ev Idx": 351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770601636.870, "dur": 0.267,
    "args": {
      "External id": 353,"Record function id": 0, "finished": true, "Ev Idx": 352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770601638.192, "dur": 30.349,
    "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": 870, "tid": 870,
    "ts": 7429770601638.714, "dur": 1.440,
    "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": 870, "tid": 870,
    "ts": 7429770601639.102, "dur": 0.956,
    "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": 870, "tid": 870,
    "ts": 7429770601640.479, "dur": 26.460,
    "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": 870, "tid": 870,
    "ts": 7429770601650.950, "dur": 1.226,
    "args": {
      "External id": 358,"Record function id": 0, "finished": true, "Ev Idx": 357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770601667.497, "dur": 0.788,
    "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": 870, "tid": 870,
    "ts": 7429770601670.868, "dur": 1.321,
    "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": 870, "tid": 870,
    "ts": 7429770601673.698, "dur": 1.722,
    "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": 870, "tid": 870,
    "ts": 7429770601674.474, "dur": 0.410,
    "args": {
      "External id": 362,"Record function id": 0, "finished": true, "Ev Idx": 361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770601682.365, "dur": 33.375,
    "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": 870, "tid": 870,
    "ts": 7429770601682.719, "dur": 2.346,
    "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": 870, "tid": 870,
    "ts": 7429770601683.668, "dur": 0.944,
    "args": {
      "External id": 365,"Record function id": 0, "finished": true, "Ev Idx": 364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770601684.197, "dur": 0.284,
    "args": {
      "External id": 366,"Record function id": 0, "finished": true, "Ev Idx": 365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770601685.366, "dur": 30.297,
    "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": 870, "tid": 870,
    "ts": 7429770601685.787, "dur": 1.457,
    "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": 870, "tid": 870,
    "ts": 7429770601686.281, "dur": 0.860,
    "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": 870, "tid": 870,
    "ts": 7429770601687.569, "dur": 26.349,
    "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": 870, "tid": 870,
    "ts": 7429770601699.004, "dur": 1.206,
    "args": {
      "External id": 371,"Record function id": 0, "finished": true, "Ev Idx": 370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770601714.497, "dur": 0.902,
    "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": 870, "tid": 870,
    "ts": 7429770601718.074, "dur": 1.250,
    "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": 870, "tid": 870,
    "ts": 7429770601720.757, "dur": 1.731,
    "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": 870, "tid": 870,
    "ts": 7429770601721.591, "dur": 0.428,
    "args": {
      "External id": 375,"Record function id": 0, "finished": true, "Ev Idx": 374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770601725.284, "dur": 2.285,
    "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": 870, "tid": 870,
    "ts": 7429770601726.737, "dur": 0.337,
    "args": {
      "External id": 377,"Record function id": 0, "finished": true, "Ev Idx": 376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770601728.743, "dur": 0.991,
    "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": 870, "tid": 870,
    "ts": 7429770601729.147, "dur": 0.273,
    "args": {
      "External id": 379,"Record function id": 0, "finished": true, "Ev Idx": 378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770601730.483, "dur": 9.202,
    "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": 870, "tid": 870,
    "ts": 7429770601744.878, "dur": 3.896,
    "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": 870, "tid": 870,
    "ts": 7429770601746.999, "dur": 0.500,
    "args": {
      "External id": 382,"Record function id": 0, "finished": true, "Ev Idx": 381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770601750.830, "dur": 3.171,
    "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": 870, "tid": 870,
    "ts": 7429770601752.709, "dur": 0.675,
    "args": {
      "External id": 384,"Record function id": 0, "finished": true, "Ev Idx": 383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770601755.664, "dur": 9.659,
    "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": 870, "tid": 870,
    "ts": 7429770601769.352, "dur": 16.762,
    "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": 870, "tid": 870,
    "ts": 7429770601788.626, "dur": 9.594,
    "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": 870, "tid": 870,
    "ts": 7429770601799.797, "dur": 8.707,
    "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": 870, "tid": 870,
    "ts": 7429770601810.691, "dur": 8.184,
    "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": 870, "tid": 870,
    "ts": 7429770601821.227, "dur": 2.062,
    "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": 870, "tid": 870,
    "ts": 7429770601822.184, "dur": 0.434,
    "args": {
      "External id": 391,"Record function id": 0, "finished": true, "Ev Idx": 390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770601824.848, "dur": 1.670,
    "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": 870, "tid": 870,
    "ts": 7429770601825.687, "dur": 0.299,
    "args": {
      "External id": 393,"Record function id": 0, "finished": true, "Ev Idx": 392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770601827.507, "dur": 8.946,
    "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": 870, "tid": 870,
    "ts": 7429770601838.362, "dur": 10.288,
    "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": 870, "tid": 870,
    "ts": 7429770601851.152, "dur": 6.798,
    "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": 870, "tid": 870,
    "ts": 7429770601859.332, "dur": 6.875,
    "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": 870, "tid": 870,
    "ts": 7429770601888.687, "dur": 2.526,
    "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": 870, "tid": 870,
    "ts": 7429770601889.949, "dur": 0.441,
    "args": {
      "External id": 399,"Record function id": 0, "finished": true, "Ev Idx": 398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770601891.874, "dur": 1.720,
    "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": 870, "tid": 870,
    "ts": 7429770601892.752, "dur": 0.322,
    "args": {
      "External id": 401,"Record function id": 0, "finished": true, "Ev Idx": 400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770601894.346, "dur": 3.454,
    "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": 870, "tid": 870,
    "ts": 7429770601896.330, "dur": 0.798,
    "args": {
      "External id": 403,"Record function id": 0, "finished": true, "Ev Idx": 402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770601898.529, "dur": 3.246,
    "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": 870, "tid": 870,
    "ts": 7429770601900.729, "dur": 0.552,
    "args": {
      "External id": 405,"Record function id": 0, "finished": true, "Ev Idx": 404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770601902.409, "dur": 2.201,
    "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": 870, "tid": 870,
    "ts": 7429770601903.800, "dur": 0.363,
    "args": {
      "External id": 407,"Record function id": 0, "finished": true, "Ev Idx": 406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770601907.214, "dur": 2.881,
    "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": 870, "tid": 870,
    "ts": 7429770601908.981, "dur": 0.490,
    "args": {
      "External id": 409,"Record function id": 0, "finished": true, "Ev Idx": 408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770601912.342, "dur": 18.362,
    "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": 870, "tid": 870,
    "ts": 7429770601914.630, "dur": 13.940,
    "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": 870, "tid": 870,
    "ts": 7429770601915.889, "dur": 2.736,
    "args": {
      "External id": 412,"Record function id": 0, "finished": true, "Ev Idx": 411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770601916.709, "dur": 1.728,
    "args": {
      "External id": 413,"Record function id": 0, "finished": true, "Ev Idx": 412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770601919.301, "dur": 8.833,
    "args": {
      "External id": 414,"Record function id": 0, "finished": true, "Ev Idx": 413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770601929.057, "dur": 1.402,
    "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": 870, "tid": 870,
    "ts": 7429770601933.827, "dur": 2.939,
    "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": 870, "tid": 870,
    "ts": 7429770601935.717, "dur": 0.478,
    "args": {
      "External id": 417,"Record function id": 0, "finished": true, "Ev Idx": 416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770601937.385, "dur": 2.468,
    "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": 870, "tid": 870,
    "ts": 7429770601939.111, "dur": 0.277,
    "args": {
      "External id": 419,"Record function id": 0, "finished": true, "Ev Idx": 418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770601940.384, "dur": 1.426,
    "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": 870, "tid": 870,
    "ts": 7429770601941.106, "dur": 0.312,
    "args": {
      "External id": 421,"Record function id": 0, "finished": true, "Ev Idx": 420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770601942.427, "dur": 1.587,
    "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": 870, "tid": 870,
    "ts": 7429770601943.256, "dur": 0.302,
    "args": {
      "External id": 423,"Record function id": 0, "finished": true, "Ev Idx": 422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770601944.633, "dur": 1.483,
    "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": 870, "tid": 870,
    "ts": 7429770601945.391, "dur": 0.282,
    "args": {
      "External id": 425,"Record function id": 0, "finished": true, "Ev Idx": 424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770601947.643, "dur": 1.534,
    "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": 870, "tid": 870,
    "ts": 7429770601948.411, "dur": 0.352,
    "args": {
      "External id": 427,"Record function id": 0, "finished": true, "Ev Idx": 426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770601950.611, "dur": 13.030,
    "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": 870, "tid": 870,
    "ts": 7429770601951.546, "dur": 10.492,
    "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": 870, "tid": 870,
    "ts": 7429770601953.401, "dur": 1.845,
    "args": {
      "External id": 430,"Record function id": 0, "finished": true, "Ev Idx": 429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770601953.937, "dur": 1.186,
    "args": {
      "External id": 431,"Record function id": 0, "finished": true, "Ev Idx": 430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770601955.551, "dur": 6.149,
    "args": {
      "External id": 432,"Record function id": 0, "finished": true, "Ev Idx": 431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770601962.523, "dur": 0.955,
    "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": 870, "tid": 870,
    "ts": 7429770601965.737, "dur": 1.757,
    "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": 870, "tid": 870,
    "ts": 7429770601966.610, "dur": 0.342,
    "args": {
      "External id": 435,"Record function id": 0, "finished": true, "Ev Idx": 434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770601968.976, "dur": 46.175,
    "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": 870, "tid": 870,
    "ts": 7429770601969.614, "dur": 1.565,
    "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": 870, "tid": 870,
    "ts": 7429770601970.390, "dur": 0.330,
    "args": {
      "External id": 438,"Record function id": 0, "finished": true, "Ev Idx": 437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770601971.533, "dur": 3.917,
    "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": 870, "tid": 870,
    "ts": 7429770601973.623, "dur": 1.693,
    "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": 870, "tid": 870,
    "ts": 7429770601975.901, "dur": 1.955,
    "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": 870, "tid": 870,
    "ts": 7429770601977.152, "dur": 0.302,
    "args": {
      "External id": 442,"Record function id": 0, "finished": true, "Ev Idx": 441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770601978.153, "dur": 1.681,
    "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": 870, "tid": 870,
    "ts": 7429770601978.943, "dur": 0.816,
    "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": 870, "tid": 870,
    "ts": 7429770601980.493, "dur": 32.683,
    "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": 870, "tid": 870,
    "ts": 7429770601994.574, "dur": 1.212,
    "args": {
      "External id": 446,"Record function id": 0, "finished": true, "Ev Idx": 445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770602013.783, "dur": 1.046,
    "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": 870, "tid": 870,
    "ts": 7429770602018.262, "dur": 9.374,
    "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": 870, "tid": 870,
    "ts": 7429770602030.898, "dur": 1.890,
    "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": 870, "tid": 870,
    "ts": 7429770602031.690, "dur": 0.434,
    "args": {
      "External id": 450,"Record function id": 0, "finished": true, "Ev Idx": 449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770602033.375, "dur": 1.252,
    "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": 870, "tid": 870,
    "ts": 7429770602033.958, "dur": 0.289,
    "args": {
      "External id": 452,"Record function id": 0, "finished": true, "Ev Idx": 451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770602035.299, "dur": 1.015,
    "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": 870, "tid": 870,
    "ts": 7429770602035.754, "dur": 0.208,
    "args": {
      "External id": 454,"Record function id": 0, "finished": true, "Ev Idx": 453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770602037.761, "dur": 7.726,
    "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": 870, "tid": 870,
    "ts": 7429770602050.870, "dur": 16.419,
    "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": 870, "tid": 870,
    "ts": 7429770602052.187, "dur": 0.332,
    "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": 870, "tid": 870,
    "ts": 7429770602053.943, "dur": 13.075,
    "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": 870, "tid": 870,
    "ts": 7429770602088.800, "dur": 0.389,
    "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": 870, "tid": 870,
    "ts": 7429770602102.981, "dur": 0.749,
    "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": 870, "tid": 870,
    "ts": 7429770602105.573, "dur": 50.773,
    "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": 870, "tid": 870,
    "ts": 7429770602106.479, "dur": 3.350,
    "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": 870, "tid": 870,
    "ts": 7429770602108.095, "dur": 0.619,
    "args": {
      "External id": 463,"Record function id": 0, "finished": true, "Ev Idx": 462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770602110.267, "dur": 2.194,
    "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": 870, "tid": 870,
    "ts": 7429770602110.666, "dur": 1.690,
    "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": 870, "tid": 870,
    "ts": 7429770602112.939, "dur": 2.718,
    "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": 870, "tid": 870,
    "ts": 7429770602114.871, "dur": 0.412,
    "args": {
      "External id": 467,"Record function id": 0, "finished": true, "Ev Idx": 466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770602115.958, "dur": 2.949,
    "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": 870, "tid": 870,
    "ts": 7429770602116.443, "dur": 2.391,
    "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": 870, "tid": 870,
    "ts": 7429770602119.336, "dur": 34.865,
    "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": 870, "tid": 870,
    "ts": 7429770602133.277, "dur": 1.479,
    "args": {
      "External id": 471,"Record function id": 0, "finished": true, "Ev Idx": 470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770602154.821, "dur": 1.193,
    "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": 870, "tid": 870,
    "ts": 7429770602158.584, "dur": 1.994,
    "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": 870, "tid": 870,
    "ts": 7429770602159.636, "dur": 0.399,
    "args": {
      "External id": 474,"Record function id": 0, "finished": true, "Ev Idx": 473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770602163.406, "dur": 13.517,
    "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": 870, "tid": 870,
    "ts": 7429770602164.035, "dur": 12.745,
    "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": 870, "tid": 870,
    "ts": 7429770602165.128, "dur": 2.581,
    "args": {
      "External id": 477,"Record function id": 0, "finished": true, "Ev Idx": 476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770602165.605, "dur": 1.950,
    "args": {
      "External id": 478,"Record function id": 0, "finished": true, "Ev Idx": 477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770602168.085, "dur": 8.282,
    "args": {
      "External id": 479,"Record function id": 0, "finished": true, "Ev Idx": 478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770602182.824, "dur": 2.549,
    "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": 870, "tid": 870,
    "ts": 7429770602183.341, "dur": 1.935,
    "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": 870, "tid": 870,
    "ts": 7429770602200.919, "dur": 38.664,
    "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": 870, "tid": 870,
    "ts": 7429770602201.814, "dur": 3.562,
    "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": 870, "tid": 870,
    "ts": 7429770602203.733, "dur": 1.085,
    "args": {
      "External id": 484,"Record function id": 0, "finished": true, "Ev Idx": 483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770602204.307, "dur": 0.341,
    "args": {
      "External id": 485,"Record function id": 0, "finished": true, "Ev Idx": 484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770602205.841, "dur": 33.593,
    "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": 870, "tid": 870,
    "ts": 7429770602206.428, "dur": 1.777,
    "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": 870, "tid": 870,
    "ts": 7429770602206.799, "dur": 1.304,
    "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": 870, "tid": 870,
    "ts": 7429770602208.983, "dur": 28.575,
    "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": 870, "tid": 870,
    "ts": 7429770602220.563, "dur": 1.152,
    "args": {
      "External id": 490,"Record function id": 0, "finished": true, "Ev Idx": 489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770602238.146, "dur": 1.016,
    "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": 870, "tid": 870,
    "ts": 7429770602244.615, "dur": 9.160,
    "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": 870, "tid": 870,
    "ts": 7429770602263.217, "dur": 0.191,
    "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": 870, "tid": 870,
    "ts": 7429770602265.548, "dur": 13.333,
    "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": 870, "tid": 870,
    "ts": 7429770602268.401, "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": 870, "tid": 870,
    "ts": 7429770602269.417, "dur": 0.208,
    "args": {
      "External id": 496,"Record function id": 0, "finished": true, "Ev Idx": 495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770602281.629, "dur": 14.441,
    "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": 870, "tid": 870,
    "ts": 7429770602304.876, "dur": 13.424,
    "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": 870, "tid": 870,
    "ts": 7429770602321.052, "dur": 9.837,
    "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": 870, "tid": 870,
    "ts": 7429770602332.896, "dur": 11.923,
    "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": 870, "tid": 870,
    "ts": 7429770602348.045, "dur": 0.139,
    "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": 870, "tid": 870,
    "ts": 7429770602349.069, "dur": 9.188,
    "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": 870, "tid": 870,
    "ts": 7429770602374.610, "dur": 42.665,
    "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": 870, "tid": 870,
    "ts": 7429770602375.239, "dur": 4.692,
    "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": 870, "tid": 870,
    "ts": 7429770602377.706, "dur": 1.448,
    "args": {
      "External id": 505,"Record function id": 0, "finished": true, "Ev Idx": 504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770602378.473, "dur": 0.555,
    "args": {
      "External id": 506,"Record function id": 0, "finished": true, "Ev Idx": 505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770602380.302, "dur": 36.818,
    "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": 870, "tid": 870,
    "ts": 7429770602381.112, "dur": 1.947,
    "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": 870, "tid": 870,
    "ts": 7429770602381.518, "dur": 1.403,
    "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": 870, "tid": 870,
    "ts": 7429770602383.454, "dur": 31.551,
    "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": 870, "tid": 870,
    "ts": 7429770602396.908, "dur": 1.403,
    "args": {
      "External id": 511,"Record function id": 0, "finished": true, "Ev Idx": 510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770602415.665, "dur": 1.111,
    "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": 870, "tid": 870,
    "ts": 7429770602458.133, "dur": 10.916,
    "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": 870, "tid": 870,
    "ts": 7429770602476.944, "dur": 31.679,
    "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": 870, "tid": 870,
    "ts": 7429770602477.333, "dur": 2.476,
    "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": 870, "tid": 870,
    "ts": 7429770602478.407, "dur": 0.890,
    "args": {
      "External id": 516,"Record function id": 0, "finished": true, "Ev Idx": 515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770602478.817, "dur": 0.360,
    "args": {
      "External id": 517,"Record function id": 0, "finished": true, "Ev Idx": 516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770602480.152, "dur": 28.356,
    "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": 870, "tid": 870,
    "ts": 7429770602480.531, "dur": 1.402,
    "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": 870, "tid": 870,
    "ts": 7429770602480.828, "dur": 0.998,
    "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": 870, "tid": 870,
    "ts": 7429770602482.244, "dur": 24.348,
    "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": 870, "tid": 870,
    "ts": 7429770602490.872, "dur": 1.315,
    "args": {
      "External id": 522,"Record function id": 0, "finished": true, "Ev Idx": 521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770602507.215, "dur": 0.941,
    "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": 870, "tid": 870,
    "ts": 7429770602510.190, "dur": 8.530,
    "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": 870, "tid": 870,
    "ts": 7429770602525.145, "dur": 49.275,
    "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": 870, "tid": 870,
    "ts": 7429770602525.541, "dur": 2.220,
    "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": 870, "tid": 870,
    "ts": 7429770602526.385, "dur": 0.903,
    "args": {
      "External id": 527,"Record function id": 0, "finished": true, "Ev Idx": 526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770602526.818, "dur": 0.326,
    "args": {
      "External id": 528,"Record function id": 0, "finished": true, "Ev Idx": 527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770602528.072, "dur": 46.229,
    "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": 870, "tid": 870,
    "ts": 7429770602528.506, "dur": 1.335,
    "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": 870, "tid": 870,
    "ts": 7429770602528.797, "dur": 0.939,
    "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": 870, "tid": 870,
    "ts": 7429770602530.165, "dur": 42.330,
    "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": 870, "tid": 870,
    "ts": 7429770602542.720, "dur": 1.159,
    "args": {
      "External id": 533,"Record function id": 0, "finished": true, "Ev Idx": 532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770602573.100, "dur": 0.921,
    "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": 870, "tid": 870,
    "ts": 7429770602577.008, "dur": 7.801,
    "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": 870, "tid": 870,
    "ts": 7429770602605.252, "dur": 0.180,
    "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": 870, "tid": 870,
    "ts": 7429770602607.408, "dur": 10.674,
    "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": 870, "tid": 870,
    "ts": 7429770602610.034, "dur": 0.387,
    "args": {
      "External id": 538,"Record function id": 0, "finished": true, "Ev Idx": 537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770602610.775, "dur": 0.072,
    "args": {
      "External id": 539,"Record function id": 0, "finished": true, "Ev Idx": 538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770602620.561, "dur": 13.970,
    "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": 870, "tid": 870,
    "ts": 7429770602637.877, "dur": 10.283,
    "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": 870, "tid": 870,
    "ts": 7429770602649.739, "dur": 9.406,
    "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": 870, "tid": 870,
    "ts": 7429770602660.608, "dur": 7.216,
    "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": 870, "tid": 870,
    "ts": 7429770602670.158, "dur": 0.113,
    "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": 870, "tid": 870,
    "ts": 7429770602671.050, "dur": 8.840,
    "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": 870, "tid": 870,
    "ts": 7429770602704.343, "dur": 36.160,
    "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": 870, "tid": 870,
    "ts": 7429770602704.759, "dur": 2.655,
    "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": 870, "tid": 870,
    "ts": 7429770602705.742, "dur": 1.070,
    "args": {
      "External id": 548,"Record function id": 0, "finished": true, "Ev Idx": 547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770602706.250, "dur": 0.404,
    "args": {
      "External id": 549,"Record function id": 0, "finished": true, "Ev Idx": 548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770602707.777, "dur": 32.609,
    "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": 870, "tid": 870,
    "ts": 7429770602708.310, "dur": 1.421,
    "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": 870, "tid": 870,
    "ts": 7429770602708.621, "dur": 1.013,
    "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": 870, "tid": 870,
    "ts": 7429770602710.069, "dur": 28.425,
    "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": 870, "tid": 870,
    "ts": 7429770602721.319, "dur": 1.221,
    "args": {
      "External id": 554,"Record function id": 0, "finished": true, "Ev Idx": 553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770602739.108, "dur": 1.024,
    "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": 870, "tid": 870,
    "ts": 7429770602743.612, "dur": 1.322,
    "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": 870, "tid": 870,
    "ts": 7429770602746.562, "dur": 2.144,
    "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": 870, "tid": 870,
    "ts": 7429770602747.568, "dur": 0.450,
    "args": {
      "External id": 558,"Record function id": 0, "finished": true, "Ev Idx": 557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770602756.340, "dur": 31.150,
    "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": 870, "tid": 870,
    "ts": 7429770602756.672, "dur": 2.091,
    "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": 870, "tid": 870,
    "ts": 7429770602757.492, "dur": 0.786,
    "args": {
      "External id": 561,"Record function id": 0, "finished": true, "Ev Idx": 560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770602757.899, "dur": 0.289,
    "args": {
      "External id": 562,"Record function id": 0, "finished": true, "Ev Idx": 561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770602759.071, "dur": 28.321,
    "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": 870, "tid": 870,
    "ts": 7429770602759.471, "dur": 1.439,
    "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": 870, "tid": 870,
    "ts": 7429770602759.777, "dur": 1.037,
    "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": 870, "tid": 870,
    "ts": 7429770602761.245, "dur": 24.354,
    "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": 870, "tid": 870,
    "ts": 7429770602770.204, "dur": 1.266,
    "args": {
      "External id": 567,"Record function id": 0, "finished": true, "Ev Idx": 566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770602786.147, "dur": 0.974,
    "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": 870, "tid": 870,
    "ts": 7429770602789.758, "dur": 1.100,
    "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": 870, "tid": 870,
    "ts": 7429770602792.363, "dur": 1.643,
    "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": 870, "tid": 870,
    "ts": 7429770602793.027, "dur": 0.508,
    "args": {
      "External id": 571,"Record function id": 0, "finished": true, "Ev Idx": 570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770602800.687, "dur": 31.398,
    "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": 870, "tid": 870,
    "ts": 7429770602801.009, "dur": 1.942,
    "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": 870, "tid": 870,
    "ts": 7429770602801.795, "dur": 0.747,
    "args": {
      "External id": 574,"Record function id": 0, "finished": true, "Ev Idx": 573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770602802.154, "dur": 0.271,
    "args": {
      "External id": 575,"Record function id": 0, "finished": true, "Ev Idx": 574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770602803.221, "dur": 28.767,
    "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": 870, "tid": 870,
    "ts": 7429770602803.633, "dur": 1.199,
    "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": 870, "tid": 870,
    "ts": 7429770602803.897, "dur": 0.846,
    "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": 870, "tid": 870,
    "ts": 7429770602805.147, "dur": 25.170,
    "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": 870, "tid": 870,
    "ts": 7429770602815.664, "dur": 1.063,
    "args": {
      "External id": 580,"Record function id": 0, "finished": true, "Ev Idx": 579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770602830.842, "dur": 0.897,
    "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": 870, "tid": 870,
    "ts": 7429770602834.370, "dur": 1.137,
    "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": 870, "tid": 870,
    "ts": 7429770602837.064, "dur": 1.569,
    "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": 870, "tid": 870,
    "ts": 7429770602837.804, "dur": 0.359,
    "args": {
      "External id": 584,"Record function id": 0, "finished": true, "Ev Idx": 583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770602841.598, "dur": 2.171,
    "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": 870, "tid": 870,
    "ts": 7429770602842.994, "dur": 0.289,
    "args": {
      "External id": 586,"Record function id": 0, "finished": true, "Ev Idx": 585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770602844.885, "dur": 1.086,
    "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": 870, "tid": 870,
    "ts": 7429770602845.370, "dur": 0.295,
    "args": {
      "External id": 588,"Record function id": 0, "finished": true, "Ev Idx": 587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770602846.771, "dur": 9.661,
    "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": 870, "tid": 870,
    "ts": 7429770602862.081, "dur": 5.675,
    "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": 870, "tid": 870,
    "ts": 7429770602864.931, "dur": 0.914,
    "args": {
      "External id": 591,"Record function id": 0, "finished": true, "Ev Idx": 590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770602870.154, "dur": 1.737,
    "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": 870, "tid": 870,
    "ts": 7429770602870.978, "dur": 0.352,
    "args": {
      "External id": 593,"Record function id": 0, "finished": true, "Ev Idx": 592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770602873.735, "dur": 11.069,
    "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": 870, "tid": 870,
    "ts": 7429770602889.120, "dur": 16.330,
    "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": 870, "tid": 870,
    "ts": 7429770602908.443, "dur": 7.818,
    "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": 870, "tid": 870,
    "ts": 7429770602917.882, "dur": 10.877,
    "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": 870, "tid": 870,
    "ts": 7429770602930.779, "dur": 7.858,
    "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": 870, "tid": 870,
    "ts": 7429770602941.051, "dur": 2.098,
    "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": 870, "tid": 870,
    "ts": 7429770602942.088, "dur": 0.398,
    "args": {
      "External id": 600,"Record function id": 0, "finished": true, "Ev Idx": 599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770602944.777, "dur": 1.741,
    "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": 870, "tid": 870,
    "ts": 7429770602945.666, "dur": 0.309,
    "args": {
      "External id": 602,"Record function id": 0, "finished": true, "Ev Idx": 601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770602947.461, "dur": 8.383,
    "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": 870, "tid": 870,
    "ts": 7429770602957.714, "dur": 9.424,
    "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": 870, "tid": 870,
    "ts": 7429770602969.219, "dur": 8.066,
    "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": 870, "tid": 870,
    "ts": 7429770602978.706, "dur": 6.554,
    "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": 870, "tid": 870,
    "ts": 7429770603008.394, "dur": 2.332,
    "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": 870, "tid": 870,
    "ts": 7429770603009.538, "dur": 0.410,
    "args": {
      "External id": 608,"Record function id": 0, "finished": true, "Ev Idx": 607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770603011.418, "dur": 2.927,
    "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": 870, "tid": 870,
    "ts": 7429770603013.172, "dur": 0.610,
    "args": {
      "External id": 610,"Record function id": 0, "finished": true, "Ev Idx": 609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770603015.093, "dur": 3.187,
    "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": 870, "tid": 870,
    "ts": 7429770603016.950, "dur": 0.703,
    "args": {
      "External id": 612,"Record function id": 0, "finished": true, "Ev Idx": 611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770603018.930, "dur": 1.630,
    "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": 870, "tid": 870,
    "ts": 7429770603019.676, "dur": 0.375,
    "args": {
      "External id": 614,"Record function id": 0, "finished": true, "Ev Idx": 613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770603021.191, "dur": 1.551,
    "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": 870, "tid": 870,
    "ts": 7429770603021.920, "dur": 0.322,
    "args": {
      "External id": 616,"Record function id": 0, "finished": true, "Ev Idx": 615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770603025.485, "dur": 2.819,
    "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": 870, "tid": 870,
    "ts": 7429770603027.155, "dur": 0.416,
    "args": {
      "External id": 618,"Record function id": 0, "finished": true, "Ev Idx": 617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770603030.700, "dur": 21.472,
    "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": 870, "tid": 870,
    "ts": 7429770603033.088, "dur": 16.181,
    "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": 870, "tid": 870,
    "ts": 7429770603034.494, "dur": 2.567,
    "args": {
      "External id": 621,"Record function id": 0, "finished": true, "Ev Idx": 620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770603035.228, "dur": 1.629,
    "args": {
      "External id": 622,"Record function id": 0, "finished": true, "Ev Idx": 621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770603037.724, "dur": 10.989,
    "args": {
      "External id": 623,"Record function id": 0, "finished": true, "Ev Idx": 622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770603049.789, "dur": 2.116,
    "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": 870, "tid": 870,
    "ts": 7429770603055.375, "dur": 1.931,
    "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": 870, "tid": 870,
    "ts": 7429770603056.412, "dur": 0.332,
    "args": {
      "External id": 626,"Record function id": 0, "finished": true, "Ev Idx": 625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770603057.956, "dur": 1.433,
    "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": 870, "tid": 870,
    "ts": 7429770603058.638, "dur": 0.277,
    "args": {
      "External id": 628,"Record function id": 0, "finished": true, "Ev Idx": 627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770603059.884, "dur": 1.593,
    "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": 870, "tid": 870,
    "ts": 7429770603060.730, "dur": 0.325,
    "args": {
      "External id": 630,"Record function id": 0, "finished": true, "Ev Idx": 629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770603062.060, "dur": 1.448,
    "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": 870, "tid": 870,
    "ts": 7429770603062.717, "dur": 0.299,
    "args": {
      "External id": 632,"Record function id": 0, "finished": true, "Ev Idx": 631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770603064.075, "dur": 2.725,
    "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": 870, "tid": 870,
    "ts": 7429770603065.910, "dur": 0.426,
    "args": {
      "External id": 634,"Record function id": 0, "finished": true, "Ev Idx": 633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770603068.358, "dur": 1.860,
    "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": 870, "tid": 870,
    "ts": 7429770603069.448, "dur": 0.337,
    "args": {
      "External id": 636,"Record function id": 0, "finished": true, "Ev Idx": 635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770603071.756, "dur": 11.584,
    "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": 870, "tid": 870,
    "ts": 7429770603072.700, "dur": 9.213,
    "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": 870, "tid": 870,
    "ts": 7429770603073.313, "dur": 1.695,
    "args": {
      "External id": 639,"Record function id": 0, "finished": true, "Ev Idx": 638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770603073.642, "dur": 1.242,
    "args": {
      "External id": 640,"Record function id": 0, "finished": true, "Ev Idx": 639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770603075.321, "dur": 6.248,
    "args": {
      "External id": 641,"Record function id": 0, "finished": true, "Ev Idx": 640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770603082.302, "dur": 0.877,
    "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": 870, "tid": 870,
    "ts": 7429770603085.307, "dur": 1.600,
    "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": 870, "tid": 870,
    "ts": 7429770603086.068, "dur": 0.349,
    "args": {
      "External id": 644,"Record function id": 0, "finished": true, "Ev Idx": 643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770603088.374, "dur": 47.233,
    "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": 870, "tid": 870,
    "ts": 7429770603089.022, "dur": 2.600,
    "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": 870, "tid": 870,
    "ts": 7429770603090.686, "dur": 0.490,
    "args": {
      "External id": 647,"Record function id": 0, "finished": true, "Ev Idx": 646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770603091.976, "dur": 4.849,
    "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": 870, "tid": 870,
    "ts": 7429770603094.326, "dur": 2.370,
    "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": 870, "tid": 870,
    "ts": 7429770603097.271, "dur": 1.654,
    "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": 870, "tid": 870,
    "ts": 7429770603098.103, "dur": 0.429,
    "args": {
      "External id": 651,"Record function id": 0, "finished": true, "Ev Idx": 650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770603099.230, "dur": 1.715,
    "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": 870, "tid": 870,
    "ts": 7429770603100.040, "dur": 0.815,
    "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": 870, "tid": 870,
    "ts": 7429770603101.599, "dur": 31.995,
    "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": 870, "tid": 870,
    "ts": 7429770603114.445, "dur": 1.137,
    "args": {
      "External id": 655,"Record function id": 0, "finished": true, "Ev Idx": 654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770603134.207, "dur": 1.064,
    "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": 870, "tid": 870,
    "ts": 7429770603138.708, "dur": 9.940,
    "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": 870, "tid": 870,
    "ts": 7429770603151.850, "dur": 2.046,
    "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": 870, "tid": 870,
    "ts": 7429770603152.599, "dur": 0.635,
    "args": {
      "External id": 659,"Record function id": 0, "finished": true, "Ev Idx": 658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770603154.709, "dur": 1.475,
    "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": 870, "tid": 870,
    "ts": 7429770603155.447, "dur": 0.324,
    "args": {
      "External id": 661,"Record function id": 0, "finished": true, "Ev Idx": 660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770603156.870, "dur": 1.023,
    "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": 870, "tid": 870,
    "ts": 7429770603157.305, "dur": 0.240,
    "args": {
      "External id": 663,"Record function id": 0, "finished": true, "Ev Idx": 662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770603159.477, "dur": 8.320,
    "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": 870, "tid": 870,
    "ts": 7429770603174.038, "dur": 15.760,
    "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": 870, "tid": 870,
    "ts": 7429770603175.268, "dur": 0.497,
    "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": 870, "tid": 870,
    "ts": 7429770603177.156, "dur": 12.400,
    "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": 870, "tid": 870,
    "ts": 7429770603192.386, "dur": 0.155,
    "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": 870, "tid": 870,
    "ts": 7429770603203.627, "dur": 0.438,
    "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": 870, "tid": 870,
    "ts": 7429770603205.820, "dur": 44.122,
    "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": 870, "tid": 870,
    "ts": 7429770603206.314, "dur": 2.157,
    "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": 870, "tid": 870,
    "ts": 7429770603207.433, "dur": 0.359,
    "args": {
      "External id": 672,"Record function id": 0, "finished": true, "Ev Idx": 671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770603208.857, "dur": 2.650,
    "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": 870, "tid": 870,
    "ts": 7429770603209.423, "dur": 1.957,
    "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": 870, "tid": 870,
    "ts": 7429770603212.103, "dur": 1.618,
    "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": 870, "tid": 870,
    "ts": 7429770603212.971, "dur": 0.396,
    "args": {
      "External id": 676,"Record function id": 0, "finished": true, "Ev Idx": 675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770603214.150, "dur": 1.916,
    "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": 870, "tid": 870,
    "ts": 7429770603214.550, "dur": 1.442,
    "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": 870, "tid": 870,
    "ts": 7429770603216.432, "dur": 31.574,
    "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": 870, "tid": 870,
    "ts": 7429770603228.792, "dur": 1.221,
    "args": {
      "External id": 680,"Record function id": 0, "finished": true, "Ev Idx": 679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770603248.583, "dur": 1.068,
    "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": 870, "tid": 870,
    "ts": 7429770603251.885, "dur": 1.964,
    "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": 870, "tid": 870,
    "ts": 7429770603252.961, "dur": 0.344,
    "args": {
      "External id": 683,"Record function id": 0, "finished": true, "Ev Idx": 682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770603256.215, "dur": 12.217,
    "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": 870, "tid": 870,
    "ts": 7429770603256.720, "dur": 11.577,
    "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": 870, "tid": 870,
    "ts": 7429770603257.759, "dur": 1.865,
    "args": {
      "External id": 686,"Record function id": 0, "finished": true, "Ev Idx": 685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770603258.171, "dur": 1.307,
    "args": {
      "External id": 687,"Record function id": 0, "finished": true, "Ev Idx": 686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770603260.022, "dur": 7.851,
    "args": {
      "External id": 688,"Record function id": 0, "finished": true, "Ev Idx": 687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770603274.256, "dur": 1.621,
    "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": 870, "tid": 870,
    "ts": 7429770603274.560, "dur": 1.231,
    "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": 870, "tid": 870,
    "ts": 7429770603291.580, "dur": 36.312,
    "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": 870, "tid": 870,
    "ts": 7429770603292.155, "dur": 3.267,
    "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": 870, "tid": 870,
    "ts": 7429770603293.998, "dur": 0.914,
    "args": {
      "External id": 693,"Record function id": 0, "finished": true, "Ev Idx": 692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770603294.423, "dur": 0.359,
    "args": {
      "External id": 694,"Record function id": 0, "finished": true, "Ev Idx": 693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770603295.820, "dur": 31.906,
    "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": 870, "tid": 870,
    "ts": 7429770603296.301, "dur": 2.047,
    "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": 870, "tid": 870,
    "ts": 7429770603296.579, "dur": 1.657,
    "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": 870, "tid": 870,
    "ts": 7429770603298.866, "dur": 26.771,
    "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": 870, "tid": 870,
    "ts": 7429770603309.387, "dur": 1.111,
    "args": {
      "External id": 699,"Record function id": 0, "finished": true, "Ev Idx": 698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770603326.347, "dur": 1.060,
    "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": 870, "tid": 870,
    "ts": 7429770603332.842, "dur": 8.625,
    "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": 870, "tid": 870,
    "ts": 7429770603350.896, "dur": 0.170,
    "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": 870, "tid": 870,
    "ts": 7429770603353.032, "dur": 12.818,
    "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": 870, "tid": 870,
    "ts": 7429770603355.234, "dur": 0.534,
    "args": {
      "External id": 704,"Record function id": 0, "finished": true, "Ev Idx": 703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770603356.287, "dur": 0.078,
    "args": {
      "External id": 705,"Record function id": 0, "finished": true, "Ev Idx": 704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770603368.740, "dur": 11.929,
    "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": 870, "tid": 870,
    "ts": 7429770603383.915, "dur": 9.938,
    "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": 870, "tid": 870,
    "ts": 7429770603395.747, "dur": 11.533,
    "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": 870, "tid": 870,
    "ts": 7429770603409.123, "dur": 8.108,
    "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": 870, "tid": 870,
    "ts": 7429770603419.732, "dur": 0.326,
    "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": 870, "tid": 870,
    "ts": 7429770603420.920, "dur": 9.077,
    "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": 870, "tid": 870,
    "ts": 7429770603446.513, "dur": 37.678,
    "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": 870, "tid": 870,
    "ts": 7429770603447.064, "dur": 3.006,
    "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": 870, "tid": 870,
    "ts": 7429770603448.208, "dur": 1.248,
    "args": {
      "External id": 714,"Record function id": 0, "finished": true, "Ev Idx": 713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770603448.903, "dur": 0.406,
    "args": {
      "External id": 715,"Record function id": 0, "finished": true, "Ev Idx": 714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770603450.411, "dur": 33.638,
    "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": 870, "tid": 870,
    "ts": 7429770603450.870, "dur": 3.377,
    "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": 870, "tid": 870,
    "ts": 7429770603451.218, "dur": 2.918,
    "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": 870, "tid": 870,
    "ts": 7429770603454.590, "dur": 27.302,
    "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": 870, "tid": 870,
    "ts": 7429770603464.745, "dur": 1.718,
    "args": {
      "External id": 720,"Record function id": 0, "finished": true, "Ev Idx": 719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770603482.547, "dur": 1.162,
    "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": 870, "tid": 870,
    "ts": 7429770603493.014, "dur": 9.756,
    "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": 870, "tid": 870,
    "ts": 7429770603511.034, "dur": 35.720,
    "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": 870, "tid": 870,
    "ts": 7429770603511.347, "dur": 2.283,
    "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": 870, "tid": 870,
    "ts": 7429770603512.172, "dur": 0.964,
    "args": {
      "External id": 725,"Record function id": 0, "finished": true, "Ev Idx": 724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770603512.608, "dur": 0.415,
    "args": {
      "External id": 726,"Record function id": 0, "finished": true, "Ev Idx": 725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770603513.939, "dur": 32.710,
    "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": 870, "tid": 870,
    "ts": 7429770603514.365, "dur": 2.464,
    "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": 870, "tid": 870,
    "ts": 7429770603515.015, "dur": 1.694,
    "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": 870, "tid": 870,
    "ts": 7429770603517.128, "dur": 27.546,
    "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": 870, "tid": 870,
    "ts": 7429770603527.590, "dur": 1.314,
    "args": {
      "External id": 731,"Record function id": 0, "finished": true, "Ev Idx": 730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770603545.396, "dur": 0.991,
    "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": 870, "tid": 870,
    "ts": 7429770603548.349, "dur": 9.346,
    "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": 870, "tid": 870,
    "ts": 7429770603564.691, "dur": 46.068,
    "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": 870, "tid": 870,
    "ts": 7429770603565.054, "dur": 2.333,
    "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": 870, "tid": 870,
    "ts": 7429770603565.991, "dur": 0.956,
    "args": {
      "External id": 736,"Record function id": 0, "finished": true, "Ev Idx": 735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770603566.397, "dur": 0.423,
    "args": {
      "External id": 737,"Record function id": 0, "finished": true, "Ev Idx": 736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770603567.681, "dur": 42.978,
    "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": 870, "tid": 870,
    "ts": 7429770603568.117, "dur": 1.482,
    "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": 870, "tid": 870,
    "ts": 7429770603568.366, "dur": 1.111,
    "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": 870, "tid": 870,
    "ts": 7429770603569.900, "dur": 38.865,
    "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": 870, "tid": 870,
    "ts": 7429770603581.331, "dur": 1.288,
    "args": {
      "External id": 742,"Record function id": 0, "finished": true, "Ev Idx": 741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770603609.375, "dur": 1.004,
    "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": 870, "tid": 870,
    "ts": 7429770603613.372, "dur": 7.353,
    "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": 870, "tid": 870,
    "ts": 7429770603643.573, "dur": 0.170,
    "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": 870, "tid": 870,
    "ts": 7429770603645.457, "dur": 9.320,
    "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": 870, "tid": 870,
    "ts": 7429770603647.013, "dur": 0.233,
    "args": {
      "External id": 747,"Record function id": 0, "finished": true, "Ev Idx": 746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770603647.543, "dur": 0.075,
    "args": {
      "External id": 748,"Record function id": 0, "finished": true, "Ev Idx": 747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770603656.999, "dur": 10.565,
    "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": 870, "tid": 870,
    "ts": 7429770603670.615, "dur": 8.076,
    "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": 870, "tid": 870,
    "ts": 7429770603680.188, "dur": 7.442,
    "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": 870, "tid": 870,
    "ts": 7429770603688.977, "dur": 9.927,
    "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": 870, "tid": 870,
    "ts": 7429770603701.371, "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": 870, "tid": 870,
    "ts": 7429770603702.307, "dur": 7.665,
    "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": 870, "tid": 870,
    "ts": 7429770603728.035, "dur": 39.040,
    "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": 870, "tid": 870,
    "ts": 7429770603728.597, "dur": 3.652,
    "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": 870, "tid": 870,
    "ts": 7429770603730.427, "dur": 1.200,
    "args": {
      "External id": 757,"Record function id": 0, "finished": true, "Ev Idx": 756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770603730.958, "dur": 0.419,
    "args": {
      "External id": 758,"Record function id": 0, "finished": true, "Ev Idx": 757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770603732.621, "dur": 34.357,
    "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": 870, "tid": 870,
    "ts": 7429770603733.618, "dur": 1.624,
    "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": 870, "tid": 870,
    "ts": 7429770603733.936, "dur": 1.180,
    "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": 870, "tid": 870,
    "ts": 7429770603735.621, "dur": 29.512,
    "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": 870, "tid": 870,
    "ts": 7429770603749.048, "dur": 1.218,
    "args": {
      "External id": 763,"Record function id": 0, "finished": true, "Ev Idx": 762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770603765.759, "dur": 0.957,
    "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": 870, "tid": 870,
    "ts": 7429770603770.058, "dur": 1.333,
    "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": 870, "tid": 870,
    "ts": 7429770603773.037, "dur": 1.862,
    "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": 870, "tid": 870,
    "ts": 7429770603773.879, "dur": 0.419,
    "args": {
      "External id": 767,"Record function id": 0, "finished": true, "Ev Idx": 766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770603781.989, "dur": 34.932,
    "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": 870, "tid": 870,
    "ts": 7429770603782.320, "dur": 1.926,
    "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": 870, "tid": 870,
    "ts": 7429770603783.095, "dur": 0.735,
    "args": {
      "External id": 770,"Record function id": 0, "finished": true, "Ev Idx": 769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770603783.467, "dur": 0.264,
    "args": {
      "External id": 771,"Record function id": 0, "finished": true, "Ev Idx": 770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770603784.547, "dur": 32.278,
    "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": 870, "tid": 870,
    "ts": 7429770603784.959, "dur": 1.914,
    "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": 870, "tid": 870,
    "ts": 7429770603785.210, "dur": 1.539,
    "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": 870, "tid": 870,
    "ts": 7429770603787.182, "dur": 27.932,
    "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": 870, "tid": 870,
    "ts": 7429770603799.573, "dur": 1.145,
    "args": {
      "External id": 776,"Record function id": 0, "finished": true, "Ev Idx": 775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770603815.631, "dur": 0.927,
    "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": 870, "tid": 870,
    "ts": 7429770603819.093, "dur": 1.200,
    "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": 870, "tid": 870,
    "ts": 7429770603821.544, "dur": 1.554,
    "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": 870, "tid": 870,
    "ts": 7429770603822.268, "dur": 0.349,
    "args": {
      "External id": 780,"Record function id": 0, "finished": true, "Ev Idx": 779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770603829.564, "dur": 35.007,
    "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": 870, "tid": 870,
    "ts": 7429770603830.035, "dur": 2.208,
    "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": 870, "tid": 870,
    "ts": 7429770603831.006, "dur": 0.865,
    "args": {
      "External id": 783,"Record function id": 0, "finished": true, "Ev Idx": 782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770603831.516, "dur": 0.257,
    "args": {
      "External id": 784,"Record function id": 0, "finished": true, "Ev Idx": 783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770603832.682, "dur": 31.791,
    "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": 870, "tid": 870,
    "ts": 7429770603833.266, "dur": 2.583,
    "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": 870, "tid": 870,
    "ts": 7429770603833.695, "dur": 2.033,
    "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": 870, "tid": 870,
    "ts": 7429770603836.155, "dur": 26.625,
    "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": 870, "tid": 870,
    "ts": 7429770603847.796, "dur": 1.105,
    "args": {
      "External id": 789,"Record function id": 0, "finished": true, "Ev Idx": 788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770603863.311, "dur": 0.913,
    "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": 870, "tid": 870,
    "ts": 7429770603866.807, "dur": 1.121,
    "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": 870, "tid": 870,
    "ts": 7429770603869.396, "dur": 1.477,
    "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": 870, "tid": 870,
    "ts": 7429770603870.107, "dur": 0.332,
    "args": {
      "External id": 793,"Record function id": 0, "finished": true, "Ev Idx": 792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770603873.626, "dur": 1.869,
    "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": 870, "tid": 870,
    "ts": 7429770603874.682, "dur": 0.293,
    "args": {
      "External id": 795,"Record function id": 0, "finished": true, "Ev Idx": 794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770603876.569, "dur": 1.003,
    "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": 870, "tid": 870,
    "ts": 7429770603876.990, "dur": 0.288,
    "args": {
      "External id": 797,"Record function id": 0, "finished": true, "Ev Idx": 796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770603878.537, "dur": 10.173,
    "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": 870, "tid": 870,
    "ts": 7429770603893.333, "dur": 3.737,
    "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": 870, "tid": 870,
    "ts": 7429770603895.419, "dur": 0.454,
    "args": {
      "External id": 800,"Record function id": 0, "finished": true, "Ev Idx": 799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770603899.136, "dur": 1.793,
    "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": 870, "tid": 870,
    "ts": 7429770603899.964, "dur": 0.311,
    "args": {
      "External id": 802,"Record function id": 0, "finished": true, "Ev Idx": 801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770603902.540, "dur": 11.477,
    "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": 870, "tid": 870,
    "ts": 7429770603917.544, "dur": 15.585,
    "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": 870, "tid": 870,
    "ts": 7429770603935.992, "dur": 9.554,
    "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": 870, "tid": 870,
    "ts": 7429770603947.284, "dur": 8.092,
    "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": 870, "tid": 870,
    "ts": 7429770603957.414, "dur": 8.387,
    "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": 870, "tid": 870,
    "ts": 7429770603968.157, "dur": 2.812,
    "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": 870, "tid": 870,
    "ts": 7429770603969.924, "dur": 0.359,
    "args": {
      "External id": 809,"Record function id": 0, "finished": true, "Ev Idx": 808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770603972.564, "dur": 2.141,
    "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": 870, "tid": 870,
    "ts": 7429770603973.860, "dur": 0.317,
    "args": {
      "External id": 811,"Record function id": 0, "finished": true, "Ev Idx": 810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770603975.613, "dur": 7.002,
    "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": 870, "tid": 870,
    "ts": 7429770603984.522, "dur": 12.809,
    "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": 870, "tid": 870,
    "ts": 7429770603999.418, "dur": 10.043,
    "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": 870, "tid": 870,
    "ts": 7429770604011.278, "dur": 6.759,
    "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": 870, "tid": 870,
    "ts": 7429770604039.701, "dur": 2.784,
    "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": 870, "tid": 870,
    "ts": 7429770604041.078, "dur": 0.405,
    "args": {
      "External id": 817,"Record function id": 0, "finished": true, "Ev Idx": 816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770604043.209, "dur": 1.645,
    "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": 870, "tid": 870,
    "ts": 7429770604043.988, "dur": 0.319,
    "args": {
      "External id": 819,"Record function id": 0, "finished": true, "Ev Idx": 818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770604045.534, "dur": 2.203,
    "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": 870, "tid": 870,
    "ts": 7429770604046.430, "dur": 0.639,
    "args": {
      "External id": 821,"Record function id": 0, "finished": true, "Ev Idx": 820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770604048.354, "dur": 1.729,
    "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": 870, "tid": 870,
    "ts": 7429770604049.151, "dur": 0.398,
    "args": {
      "External id": 823,"Record function id": 0, "finished": true, "Ev Idx": 822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770604050.659, "dur": 1.676,
    "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": 870, "tid": 870,
    "ts": 7429770604051.570, "dur": 0.302,
    "args": {
      "External id": 825,"Record function id": 0, "finished": true, "Ev Idx": 824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770604055.047, "dur": 2.896,
    "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": 870, "tid": 870,
    "ts": 7429770604056.937, "dur": 0.420,
    "args": {
      "External id": 827,"Record function id": 0, "finished": true, "Ev Idx": 826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770604060.233, "dur": 20.590,
    "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": 870, "tid": 870,
    "ts": 7429770604062.117, "dur": 15.215,
    "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": 870, "tid": 870,
    "ts": 7429770604063.484, "dur": 2.639,
    "args": {
      "External id": 830,"Record function id": 0, "finished": true, "Ev Idx": 829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770604064.261, "dur": 1.664,
    "args": {
      "External id": 831,"Record function id": 0, "finished": true, "Ev Idx": 830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770604066.635, "dur": 10.220,
    "args": {
      "External id": 832,"Record function id": 0, "finished": true, "Ev Idx": 831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770604077.834, "dur": 2.724,
    "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": 870, "tid": 870,
    "ts": 7429770604084.179, "dur": 2.197,
    "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": 870, "tid": 870,
    "ts": 7429770604085.446, "dur": 0.354,
    "args": {
      "External id": 835,"Record function id": 0, "finished": true, "Ev Idx": 834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770604086.974, "dur": 1.463,
    "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": 870, "tid": 870,
    "ts": 7429770604087.708, "dur": 0.276,
    "args": {
      "External id": 837,"Record function id": 0, "finished": true, "Ev Idx": 836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770604088.939, "dur": 1.495,
    "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": 870, "tid": 870,
    "ts": 7429770604089.689, "dur": 0.334,
    "args": {
      "External id": 839,"Record function id": 0, "finished": true, "Ev Idx": 838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770604091.022, "dur": 1.464,
    "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": 870, "tid": 870,
    "ts": 7429770604091.695, "dur": 0.308,
    "args": {
      "External id": 841,"Record function id": 0, "finished": true, "Ev Idx": 840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770604093.066, "dur": 1.595,
    "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": 870, "tid": 870,
    "ts": 7429770604093.919, "dur": 0.271,
    "args": {
      "External id": 843,"Record function id": 0, "finished": true, "Ev Idx": 842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770604096.338, "dur": 2.848,
    "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": 870, "tid": 870,
    "ts": 7429770604098.395, "dur": 0.343,
    "args": {
      "External id": 845,"Record function id": 0, "finished": true, "Ev Idx": 844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770604100.720, "dur": 12.539,
    "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": 870, "tid": 870,
    "ts": 7429770604101.608, "dur": 10.145,
    "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": 870, "tid": 870,
    "ts": 7429770604102.995, "dur": 1.850,
    "args": {
      "External id": 848,"Record function id": 0, "finished": true, "Ev Idx": 847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770604103.343, "dur": 1.364,
    "args": {
      "External id": 849,"Record function id": 0, "finished": true, "Ev Idx": 848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770604105.176, "dur": 6.281,
    "args": {
      "External id": 850,"Record function id": 0, "finished": true, "Ev Idx": 849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770604112.131, "dur": 0.942,
    "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": 870, "tid": 870,
    "ts": 7429770604115.322, "dur": 1.617,
    "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": 870, "tid": 870,
    "ts": 7429770604116.141, "dur": 0.347,
    "args": {
      "External id": 853,"Record function id": 0, "finished": true, "Ev Idx": 852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770604118.558, "dur": 42.023,
    "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": 870, "tid": 870,
    "ts": 7429770604119.064, "dur": 1.945,
    "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": 870, "tid": 870,
    "ts": 7429770604120.104, "dur": 0.477,
    "args": {
      "External id": 856,"Record function id": 0, "finished": true, "Ev Idx": 855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770604121.377, "dur": 3.804,
    "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": 870, "tid": 870,
    "ts": 7429770604122.852, "dur": 2.183,
    "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": 870, "tid": 870,
    "ts": 7429770604125.616, "dur": 1.336,
    "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": 870, "tid": 870,
    "ts": 7429770604126.299, "dur": 0.266,
    "args": {
      "External id": 860,"Record function id": 0, "finished": true, "Ev Idx": 859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770604127.249, "dur": 1.661,
    "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": 870, "tid": 870,
    "ts": 7429770604128.012, "dur": 0.792,
    "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": 870, "tid": 870,
    "ts": 7429770604129.435, "dur": 29.248,
    "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": 870, "tid": 870,
    "ts": 7429770604141.002, "dur": 1.119,
    "args": {
      "External id": 864,"Record function id": 0, "finished": true, "Ev Idx": 863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770604159.305, "dur": 0.983,
    "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": 870, "tid": 870,
    "ts": 7429770604163.291, "dur": 9.244,
    "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": 870, "tid": 870,
    "ts": 7429770604175.635, "dur": 1.983,
    "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": 870, "tid": 870,
    "ts": 7429770604176.482, "dur": 0.549,
    "args": {
      "External id": 868,"Record function id": 0, "finished": true, "Ev Idx": 867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770604178.301, "dur": 1.356,
    "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": 870, "tid": 870,
    "ts": 7429770604178.925, "dur": 0.309,
    "args": {
      "External id": 870,"Record function id": 0, "finished": true, "Ev Idx": 869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770604180.344, "dur": 1.185,
    "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": 870, "tid": 870,
    "ts": 7429770604180.926, "dur": 0.264,
    "args": {
      "External id": 872,"Record function id": 0, "finished": true, "Ev Idx": 871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770604183.106, "dur": 7.494,
    "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": 870, "tid": 870,
    "ts": 7429770604195.188, "dur": 12.171,
    "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": 870, "tid": 870,
    "ts": 7429770604196.050, "dur": 0.279,
    "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": 870, "tid": 870,
    "ts": 7429770604196.968, "dur": 10.162,
    "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": 870, "tid": 870,
    "ts": 7429770604209.884, "dur": 0.203,
    "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": 870, "tid": 870,
    "ts": 7429770604220.043, "dur": 0.336,
    "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": 870, "tid": 870,
    "ts": 7429770604222.006, "dur": 40.108,
    "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": 870, "tid": 870,
    "ts": 7429770604222.552, "dur": 2.389,
    "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": 870, "tid": 870,
    "ts": 7429770604224.093, "dur": 0.350,
    "args": {
      "External id": 881,"Record function id": 0, "finished": true, "Ev Idx": 880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770604225.311, "dur": 3.018,
    "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": 870, "tid": 870,
    "ts": 7429770604225.614, "dur": 2.590,
    "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": 870, "tid": 870,
    "ts": 7429770604228.731, "dur": 2.747,
    "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": 870, "tid": 870,
    "ts": 7429770604230.848, "dur": 0.239,
    "args": {
      "External id": 885,"Record function id": 0, "finished": true, "Ev Idx": 884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770604232.060, "dur": 1.876,
    "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": 870, "tid": 870,
    "ts": 7429770604232.470, "dur": 1.393,
    "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": 870, "tid": 870,
    "ts": 7429770604234.344, "dur": 25.995,
    "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": 870, "tid": 870,
    "ts": 7429770604244.458, "dur": 1.097,
    "args": {
      "External id": 889,"Record function id": 0, "finished": true, "Ev Idx": 888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770604260.940, "dur": 0.917,
    "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": 870, "tid": 870,
    "ts": 7429770604264.035, "dur": 1.692,
    "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": 870, "tid": 870,
    "ts": 7429770604264.906, "dur": 0.335,
    "args": {
      "External id": 892,"Record function id": 0, "finished": true, "Ev Idx": 891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770604267.744, "dur": 11.294,
    "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": 870, "tid": 870,
    "ts": 7429770604268.171, "dur": 10.724,
    "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": 870, "tid": 870,
    "ts": 7429770604269.065, "dur": 1.880,
    "args": {
      "External id": 895,"Record function id": 0, "finished": true, "Ev Idx": 894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770604269.455, "dur": 1.323,
    "args": {
      "External id": 896,"Record function id": 0, "finished": true, "Ev Idx": 895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770604271.312, "dur": 7.219,
    "args": {
      "External id": 897,"Record function id": 0, "finished": true, "Ev Idx": 896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770604284.268, "dur": 1.640,
    "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": 870, "tid": 870,
    "ts": 7429770604284.560, "dur": 1.247,
    "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": 870, "tid": 870,
    "ts": 7429770604299.764, "dur": 36.765,
    "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": 870, "tid": 870,
    "ts": 7429770604300.342, "dur": 2.796,
    "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": 870, "tid": 870,
    "ts": 7429770604301.780, "dur": 0.861,
    "args": {
      "External id": 902,"Record function id": 0, "finished": true, "Ev Idx": 901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770604302.191, "dur": 0.324,
    "args": {
      "External id": 903,"Record function id": 0, "finished": true, "Ev Idx": 902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770604303.468, "dur": 32.933,
    "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": 870, "tid": 870,
    "ts": 7429770604303.910, "dur": 1.614,
    "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": 870, "tid": 870,
    "ts": 7429770604304.200, "dur": 1.231,
    "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": 870, "tid": 870,
    "ts": 7429770604306.053, "dur": 28.550,
    "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": 870, "tid": 870,
    "ts": 7429770604318.715, "dur": 1.100,
    "args": {
      "External id": 908,"Record function id": 0, "finished": true, "Ev Idx": 907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770604335.223, "dur": 0.895,
    "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": 870, "tid": 870,
    "ts": 7429770604341.616, "dur": 8.472,
    "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": 870, "tid": 870,
    "ts": 7429770604358.150, "dur": 0.181,
    "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": 870, "tid": 870,
    "ts": 7429770604360.250, "dur": 10.411,
    "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": 870, "tid": 870,
    "ts": 7429770604362.039, "dur": 0.379,
    "args": {
      "External id": 913,"Record function id": 0, "finished": true, "Ev Idx": 912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770604362.813, "dur": 0.111,
    "args": {
      "External id": 914,"Record function id": 0, "finished": true, "Ev Idx": 913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770604373.114, "dur": 12.599,
    "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": 870, "tid": 870,
    "ts": 7429770604388.976, "dur": 8.571,
    "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": 870, "tid": 870,
    "ts": 7429770604399.305, "dur": 8.992,
    "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": 870, "tid": 870,
    "ts": 7429770604409.896, "dur": 8.791,
    "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": 870, "tid": 870,
    "ts": 7429770604421.311, "dur": 0.283,
    "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": 870, "tid": 870,
    "ts": 7429770604422.446, "dur": 9.762,
    "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": 870, "tid": 870,
    "ts": 7429770604446.329, "dur": 38.935,
    "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": 870, "tid": 870,
    "ts": 7429770604446.729, "dur": 2.817,
    "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": 870, "tid": 870,
    "ts": 7429770604447.642, "dur": 1.282,
    "args": {
      "External id": 923,"Record function id": 0, "finished": true, "Ev Idx": 922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770604448.382, "dur": 0.405,
    "args": {
      "External id": 924,"Record function id": 0, "finished": true, "Ev Idx": 923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770604449.913, "dur": 35.229,
    "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": 870, "tid": 870,
    "ts": 7429770604450.407, "dur": 1.534,
    "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": 870, "tid": 870,
    "ts": 7429770604450.735, "dur": 1.108,
    "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": 870, "tid": 870,
    "ts": 7429770604452.301, "dur": 30.934,
    "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": 870, "tid": 870,
    "ts": 7429770604466.955, "dur": 1.144,
    "args": {
      "External id": 929,"Record function id": 0, "finished": true, "Ev Idx": 928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770604483.952, "dur": 0.902,
    "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": 870, "tid": 870,
    "ts": 7429770604493.273, "dur": 9.073,
    "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": 870, "tid": 870,
    "ts": 7429770604510.429, "dur": 33.418,
    "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": 870, "tid": 870,
    "ts": 7429770604510.761, "dur": 2.181,
    "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": 870, "tid": 870,
    "ts": 7429770604511.528, "dur": 0.858,
    "args": {
      "External id": 934,"Record function id": 0, "finished": true, "Ev Idx": 933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770604511.914, "dur": 0.370,
    "args": {
      "External id": 935,"Record function id": 0, "finished": true, "Ev Idx": 934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770604513.309, "dur": 30.455,
    "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": 870, "tid": 870,
    "ts": 7429770604513.883, "dur": 2.979,
    "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": 870, "tid": 870,
    "ts": 7429770604514.308, "dur": 2.431,
    "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": 870, "tid": 870,
    "ts": 7429770604517.180, "dur": 24.769,
    "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": 870, "tid": 870,
    "ts": 7429770604526.599, "dur": 1.109,
    "args": {
      "External id": 940,"Record function id": 0, "finished": true, "Ev Idx": 939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770604542.719, "dur": 0.807,
    "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": 870, "tid": 870,
    "ts": 7429770604545.381, "dur": 8.182,
    "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": 870, "tid": 870,
    "ts": 7429770604559.915, "dur": 39.925,
    "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": 870, "tid": 870,
    "ts": 7429770604560.422, "dur": 2.501,
    "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": 870, "tid": 870,
    "ts": 7429770604561.398, "dur": 1.047,
    "args": {
      "External id": 945,"Record function id": 0, "finished": true, "Ev Idx": 944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770604562.022, "dur": 0.300,
    "args": {
      "External id": 946,"Record function id": 0, "finished": true, "Ev Idx": 945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770604563.383, "dur": 36.353,
    "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": 870, "tid": 870,
    "ts": 7429770604563.789, "dur": 1.480,
    "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": 870, "tid": 870,
    "ts": 7429770604564.054, "dur": 1.087,
    "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": 870, "tid": 870,
    "ts": 7429770604565.555, "dur": 32.454,
    "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": 870, "tid": 870,
    "ts": 7429770604573.757, "dur": 1.087,
    "args": {
      "External id": 951,"Record function id": 0, "finished": true, "Ev Idx": 950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770604598.590, "dur": 0.909,
    "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": 870, "tid": 870,
    "ts": 7429770604602.257, "dur": 7.435,
    "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": 870, "tid": 870,
    "ts": 7429770604626.686, "dur": 0.143,
    "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": 870, "tid": 870,
    "ts": 7429770604628.449, "dur": 8.401,
    "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": 870, "tid": 870,
    "ts": 7429770604629.694, "dur": 0.253,
    "args": {
      "External id": 956,"Record function id": 0, "finished": true, "Ev Idx": 955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770604630.239, "dur": 0.073,
    "args": {
      "External id": 957,"Record function id": 0, "finished": true, "Ev Idx": 956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770604638.894, "dur": 13.168,
    "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": 870, "tid": 870,
    "ts": 7429770604654.842, "dur": 8.362,
    "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": 870, "tid": 870,
    "ts": 7429770604664.636, "dur": 9.669,
    "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": 870, "tid": 870,
    "ts": 7429770604675.881, "dur": 8.568,
    "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": 870, "tid": 870,
    "ts": 7429770604686.674, "dur": 0.101,
    "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": 870, "tid": 870,
    "ts": 7429770604687.534, "dur": 8.742,
    "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": 870, "tid": 870,
    "ts": 7429770604712.701, "dur": 175.434,
    "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": 870, "tid": 870,
    "ts": 7429770604713.103, "dur": 2.658,
    "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": 870, "tid": 870,
    "ts": 7429770604714.069, "dur": 1.114,
    "args": {
      "External id": 966,"Record function id": 0, "finished": true, "Ev Idx": 965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770604714.566, "dur": 0.481,
    "args": {
      "External id": 967,"Record function id": 0, "finished": true, "Ev Idx": 966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770604716.098, "dur": 171.859,
    "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": 870, "tid": 870,
    "ts": 7429770604716.656, "dur": 1.578,
    "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": 870, "tid": 870,
    "ts": 7429770604717.120, "dur": 0.993,
    "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": 870, "tid": 870,
    "ts": 7429770604718.712, "dur": 165.552,
    "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": 870, "tid": 870,
    "ts": 7429770604732.292, "dur": 127.845,
    "args": {
      "External id": 972,"Record function id": 0, "finished": true, "Ev Idx": 971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770604885.554, "dur": 1.930,
    "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": 870, "tid": 870,
    "ts": 7429770604893.826, "dur": 1.838,
    "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": 870, "tid": 870,
    "ts": 7429770604897.852, "dur": 3.257,
    "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": 870, "tid": 870,
    "ts": 7429770604899.489, "dur": 0.709,
    "args": {
      "External id": 976,"Record function id": 0, "finished": true, "Ev Idx": 975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770604913.558, "dur": 44.166,
    "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": 870, "tid": 870,
    "ts": 7429770604914.160, "dur": 3.207,
    "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": 870, "tid": 870,
    "ts": 7429770604915.683, "dur": 1.164,
    "args": {
      "External id": 979,"Record function id": 0, "finished": true, "Ev Idx": 978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770604916.357, "dur": 0.338,
    "args": {
      "External id": 980,"Record function id": 0, "finished": true, "Ev Idx": 979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770604917.738, "dur": 39.870,
    "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": 870, "tid": 870,
    "ts": 7429770604918.353, "dur": 2.118,
    "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": 870, "tid": 870,
    "ts": 7429770604918.858, "dur": 1.500,
    "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": 870, "tid": 870,
    "ts": 7429770604920.898, "dur": 34.697,
    "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": 870, "tid": 870,
    "ts": 7429770604936.898, "dur": 1.856,
    "args": {
      "External id": 985,"Record function id": 0, "finished": true, "Ev Idx": 984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770604956.226, "dur": 1.054,
    "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": 870, "tid": 870,
    "ts": 7429770604960.213, "dur": 1.210,
    "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": 870, "tid": 870,
    "ts": 7429770604963.173, "dur": 1.925,
    "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": 870, "tid": 870,
    "ts": 7429770604964.185, "dur": 0.417,
    "args": {
      "External id": 989,"Record function id": 0, "finished": true, "Ev Idx": 988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770604973.920, "dur": 33.795,
    "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": 870, "tid": 870,
    "ts": 7429770604974.438, "dur": 2.694,
    "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": 870, "tid": 870,
    "ts": 7429770604975.578, "dur": 1.063,
    "args": {
      "External id": 992,"Record function id": 0, "finished": true, "Ev Idx": 991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770604976.150, "dur": 0.391,
    "args": {
      "External id": 993,"Record function id": 0, "finished": true, "Ev Idx": 992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770604977.584, "dur": 30.044,
    "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": 870, "tid": 870,
    "ts": 7429770604978.154, "dur": 1.398,
    "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": 870, "tid": 870,
    "ts": 7429770604978.529, "dur": 0.928,
    "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": 870, "tid": 870,
    "ts": 7429770604979.848, "dur": 26.008,
    "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": 870, "tid": 870,
    "ts": 7429770604989.706, "dur": 1.185,
    "args": {
      "External id": 998,"Record function id": 0, "finished": true, "Ev Idx": 997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770605006.442, "dur": 0.894,
    "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": 870, "tid": 870,
    "ts": 7429770605010.569, "dur": 1.123,
    "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": 870, "tid": 870,
    "ts": 7429770605013.342, "dur": 1.526,
    "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": 870, "tid": 870,
    "ts": 7429770605014.038, "dur": 0.330,
    "args": {
      "External id": 1002,"Record function id": 0, "finished": true, "Ev Idx": 1001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770605018.911, "dur": 2.008,
    "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": 870, "tid": 870,
    "ts": 7429770605020.076, "dur": 0.322,
    "args": {
      "External id": 1004,"Record function id": 0, "finished": true, "Ev Idx": 1003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770605021.971, "dur": 1.104,
    "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": 870, "tid": 870,
    "ts": 7429770605022.398, "dur": 0.283,
    "args": {
      "External id": 1006,"Record function id": 0, "finished": true, "Ev Idx": 1005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770605024.141, "dur": 13.004,
    "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": 870, "tid": 870,
    "ts": 7429770605042.886, "dur": 5.860,
    "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": 870, "tid": 870,
    "ts": 7429770605045.561, "dur": 1.390,
    "args": {
      "External id": 1009,"Record function id": 0, "finished": true, "Ev Idx": 1008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770605051.210, "dur": 2.619,
    "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": 870, "tid": 870,
    "ts": 7429770605052.846, "dur": 0.330,
    "args": {
      "External id": 1011,"Record function id": 0, "finished": true, "Ev Idx": 1010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770605055.981, "dur": 10.541,
    "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": 870, "tid": 870,
    "ts": 7429770605072.837, "dur": 20.230,
    "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": 870, "tid": 870,
    "ts": 7429770605095.973, "dur": 8.179,
    "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": 870, "tid": 870,
    "ts": 7429770605106.941, "dur": 9.834,
    "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": 870, "tid": 870,
    "ts": 7429770605119.028, "dur": 8.401,
    "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": 870, "tid": 870,
    "ts": 7429770605129.972, "dur": 2.199,
    "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": 870, "tid": 870,
    "ts": 7429770605131.104, "dur": 0.377,
    "args": {
      "External id": 1018,"Record function id": 0, "finished": true, "Ev Idx": 1017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770605133.847, "dur": 1.704,
    "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": 870, "tid": 870,
    "ts": 7429770605134.691, "dur": 0.313,
    "args": {
      "External id": 1020,"Record function id": 0, "finished": true, "Ev Idx": 1019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770605136.569, "dur": 8.292,
    "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": 870, "tid": 870,
    "ts": 7429770605146.857, "dur": 8.858,
    "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": 870, "tid": 870,
    "ts": 7429770605158.115, "dur": 7.396,
    "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": 870, "tid": 870,
    "ts": 7429770605167.010, "dur": 5.935,
    "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": 870, "tid": 870,
    "ts": 7429770605233.797, "dur": 2.472,
    "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": 870, "tid": 870,
    "ts": 7429770605234.990, "dur": 0.425,
    "args": {
      "External id": 1026,"Record function id": 0, "finished": true, "Ev Idx": 1025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770605236.945, "dur": 1.653,
    "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": 870, "tid": 870,
    "ts": 7429770605237.753, "dur": 0.317,
    "args": {
      "External id": 1028,"Record function id": 0, "finished": true, "Ev Idx": 1027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770605239.368, "dur": 3.394,
    "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": 870, "tid": 870,
    "ts": 7429770605241.367, "dur": 0.748,
    "args": {
      "External id": 1030,"Record function id": 0, "finished": true, "Ev Idx": 1029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770605243.443, "dur": 2.584,
    "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": 870, "tid": 870,
    "ts": 7429770605245.120, "dur": 0.380,
    "args": {
      "External id": 1032,"Record function id": 0, "finished": true, "Ev Idx": 1031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770605246.663, "dur": 1.508,
    "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": 870, "tid": 870,
    "ts": 7429770605247.389, "dur": 0.326,
    "args": {
      "External id": 1034,"Record function id": 0, "finished": true, "Ev Idx": 1033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770605251.053, "dur": 3.100,
    "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": 870, "tid": 870,
    "ts": 7429770605253.049, "dur": 0.402,
    "args": {
      "External id": 1036,"Record function id": 0, "finished": true, "Ev Idx": 1035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770605256.493, "dur": 22.052,
    "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": 870, "tid": 870,
    "ts": 7429770605259.202, "dur": 16.033,
    "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": 870, "tid": 870,
    "ts": 7429770605260.944, "dur": 2.892,
    "args": {
      "External id": 1039,"Record function id": 0, "finished": true, "Ev Idx": 1038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770605261.824, "dur": 1.794,
    "args": {
      "External id": 1040,"Record function id": 0, "finished": true, "Ev Idx": 1039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770605264.760, "dur": 9.809,
    "args": {
      "External id": 1041,"Record function id": 0, "finished": true, "Ev Idx": 1040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770605275.801, "dur": 2.441,
    "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": 870, "tid": 870,
    "ts": 7429770605281.883, "dur": 2.419,
    "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": 870, "tid": 870,
    "ts": 7429770605283.313, "dur": 0.389,
    "args": {
      "External id": 1044,"Record function id": 0, "finished": true, "Ev Idx": 1043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770605284.951, "dur": 1.501,
    "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": 870, "tid": 870,
    "ts": 7429770605285.708, "dur": 0.272,
    "args": {
      "External id": 1046,"Record function id": 0, "finished": true, "Ev Idx": 1045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770605286.942, "dur": 1.508,
    "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": 870, "tid": 870,
    "ts": 7429770605287.667, "dur": 0.327,
    "args": {
      "External id": 1048,"Record function id": 0, "finished": true, "Ev Idx": 1047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770605289.060, "dur": 1.456,
    "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": 870, "tid": 870,
    "ts": 7429770605289.713, "dur": 0.307,
    "args": {
      "External id": 1050,"Record function id": 0, "finished": true, "Ev Idx": 1049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770605291.124, "dur": 1.390,
    "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": 870, "tid": 870,
    "ts": 7429770605291.795, "dur": 0.275,
    "args": {
      "External id": 1052,"Record function id": 0, "finished": true, "Ev Idx": 1051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770605294.071, "dur": 2.507,
    "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": 870, "tid": 870,
    "ts": 7429770605295.765, "dur": 0.357,
    "args": {
      "External id": 1054,"Record function id": 0, "finished": true, "Ev Idx": 1053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770605298.118, "dur": 12.031,
    "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": 870, "tid": 870,
    "ts": 7429770605299.021, "dur": 9.579,
    "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": 870, "tid": 870,
    "ts": 7429770605300.181, "dur": 1.695,
    "args": {
      "External id": 1057,"Record function id": 0, "finished": true, "Ev Idx": 1056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770605300.533, "dur": 1.216,
    "args": {
      "External id": 1058,"Record function id": 0, "finished": true, "Ev Idx": 1057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770605302.196, "dur": 6.074,
    "args": {
      "External id": 1059,"Record function id": 0, "finished": true, "Ev Idx": 1058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770605308.991, "dur": 0.975,
    "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": 870, "tid": 870,
    "ts": 7429770605312.225, "dur": 1.711,
    "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": 870, "tid": 870,
    "ts": 7429770605313.068, "dur": 0.357,
    "args": {
      "External id": 1062,"Record function id": 0, "finished": true, "Ev Idx": 1061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770605315.565, "dur": 49.133,
    "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": 870, "tid": 870,
    "ts": 7429770605316.205, "dur": 1.600,
    "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": 870, "tid": 870,
    "ts": 7429770605317.026, "dur": 0.346,
    "args": {
      "External id": 1065,"Record function id": 0, "finished": true, "Ev Idx": 1064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770605318.172, "dur": 5.109,
    "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": 870, "tid": 870,
    "ts": 7429770605320.253, "dur": 2.896,
    "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": 870, "tid": 870,
    "ts": 7429770605323.758, "dur": 2.145,
    "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": 870, "tid": 870,
    "ts": 7429770605325.234, "dur": 0.297,
    "args": {
      "External id": 1069,"Record function id": 0, "finished": true, "Ev Idx": 1068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770605326.184, "dur": 1.805,
    "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": 870, "tid": 870,
    "ts": 7429770605327.003, "dur": 0.885,
    "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": 870, "tid": 870,
    "ts": 7429770605328.863, "dur": 33.681,
    "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": 870, "tid": 870,
    "ts": 7429770605342.799, "dur": 1.172,
    "args": {
      "External id": 1073,"Record function id": 0, "finished": true, "Ev Idx": 1072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770605363.192, "dur": 1.111,
    "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": 870, "tid": 870,
    "ts": 7429770605368.830, "dur": 9.901,
    "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": 870, "tid": 870,
    "ts": 7429770605382.104, "dur": 1.823,
    "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": 870, "tid": 870,
    "ts": 7429770605382.914, "dur": 0.375,
    "args": {
      "External id": 1077,"Record function id": 0, "finished": true, "Ev Idx": 1076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770605384.588, "dur": 1.335,
    "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": 870, "tid": 870,
    "ts": 7429770605385.188, "dur": 0.302,
    "args": {
      "External id": 1079,"Record function id": 0, "finished": true, "Ev Idx": 1078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770605386.560, "dur": 1.039,
    "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": 870, "tid": 870,
    "ts": 7429770605387.022, "dur": 0.243,
    "args": {
      "External id": 1081,"Record function id": 0, "finished": true, "Ev Idx": 1080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770605389.303, "dur": 7.729,
    "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": 870, "tid": 870,
    "ts": 7429770605403.274, "dur": 13.826,
    "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": 870, "tid": 870,
    "ts": 7429770605404.645, "dur": 0.456,
    "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": 870, "tid": 870,
    "ts": 7429770605405.843, "dur": 11.015,
    "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": 870, "tid": 870,
    "ts": 7429770605420.545, "dur": 0.260,
    "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": 870, "tid": 870,
    "ts": 7429770605433.410, "dur": 0.463,
    "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": 870, "tid": 870,
    "ts": 7429770605435.685, "dur": 43.065,
    "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": 870, "tid": 870,
    "ts": 7429770605436.146, "dur": 2.558,
    "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": 870, "tid": 870,
    "ts": 7429770605437.803, "dur": 0.384,
    "args": {
      "External id": 1090,"Record function id": 0, "finished": true, "Ev Idx": 1089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770605439.179, "dur": 3.258,
    "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": 870, "tid": 870,
    "ts": 7429770605439.645, "dur": 2.662,
    "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": 870, "tid": 870,
    "ts": 7429770605443.076, "dur": 2.031,
    "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": 870, "tid": 870,
    "ts": 7429770605444.479, "dur": 0.260,
    "args": {
      "External id": 1094,"Record function id": 0, "finished": true, "Ev Idx": 1093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770605445.403, "dur": 1.206,
    "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": 870, "tid": 870,
    "ts": 7429770605445.673, "dur": 0.850,
    "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": 870, "tid": 870,
    "ts": 7429770605446.994, "dur": 29.709,
    "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": 870, "tid": 870,
    "ts": 7429770605458.494, "dur": 1.123,
    "args": {
      "External id": 1098,"Record function id": 0, "finished": true, "Ev Idx": 1097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770605477.389, "dur": 1.070,
    "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": 870, "tid": 870,
    "ts": 7429770605480.777, "dur": 1.794,
    "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": 870, "tid": 870,
    "ts": 7429770605481.695, "dur": 0.339,
    "args": {
      "External id": 1101,"Record function id": 0, "finished": true, "Ev Idx": 1100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770605485.467, "dur": 12.001,
    "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": 870, "tid": 870,
    "ts": 7429770605485.968, "dur": 11.348,
    "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": 870, "tid": 870,
    "ts": 7429770605486.960, "dur": 1.908,
    "args": {
      "External id": 1104,"Record function id": 0, "finished": true, "Ev Idx": 1103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770605487.368, "dur": 1.339,
    "args": {
      "External id": 1105,"Record function id": 0, "finished": true, "Ev Idx": 1104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770605489.244, "dur": 7.637,
    "args": {
      "External id": 1106,"Record function id": 0, "finished": true, "Ev Idx": 1105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770605503.371, "dur": 1.619,
    "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": 870, "tid": 870,
    "ts": 7429770605503.748, "dur": 1.139,
    "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": 870, "tid": 870,
    "ts": 7429770605520.215, "dur": 37.524,
    "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": 870, "tid": 870,
    "ts": 7429770605520.905, "dur": 3.141,
    "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": 870, "tid": 870,
    "ts": 7429770605522.428, "dur": 1.038,
    "args": {
      "External id": 1111,"Record function id": 0, "finished": true, "Ev Idx": 1110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770605523.027, "dur": 0.331,
    "args": {
      "External id": 1112,"Record function id": 0, "finished": true, "Ev Idx": 1111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770605524.457, "dur": 33.119,
    "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": 870, "tid": 870,
    "ts": 7429770605525.047, "dur": 2.166,
    "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": 870, "tid": 870,
    "ts": 7429770605525.503, "dur": 1.601,
    "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": 870, "tid": 870,
    "ts": 7429770605527.791, "dur": 27.790,
    "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": 870, "tid": 870,
    "ts": 7429770605540.565, "dur": 1.093,
    "args": {
      "External id": 1117,"Record function id": 0, "finished": true, "Ev Idx": 1116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770605556.234, "dur": 1.036,
    "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": 870, "tid": 870,
    "ts": 7429770605564.797, "dur": 10.349,
    "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": 870, "tid": 870,
    "ts": 7429770605586.055, "dur": 0.156,
    "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": 870, "tid": 870,
    "ts": 7429770605588.921, "dur": 14.350,
    "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": 870, "tid": 870,
    "ts": 7429770605591.057, "dur": 0.443,
    "args": {
      "External id": 1122,"Record function id": 0, "finished": true, "Ev Idx": 1121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770605592.001, "dur": 0.080,
    "args": {
      "External id": 1123,"Record function id": 0, "finished": true, "Ev Idx": 1122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770605606.618, "dur": 16.178,
    "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": 870, "tid": 870,
    "ts": 7429770605626.676, "dur": 9.491,
    "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": 870, "tid": 870,
    "ts": 7429770605638.355, "dur": 11.217,
    "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": 870, "tid": 870,
    "ts": 7429770605651.336, "dur": 9.074,
    "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": 870, "tid": 870,
    "ts": 7429770605663.179, "dur": 0.105,
    "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": 870, "tid": 870,
    "ts": 7429770605664.193, "dur": 9.714,
    "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": 870, "tid": 870,
    "ts": 7429770605694.618, "dur": 35.636,
    "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": 870, "tid": 870,
    "ts": 7429770605695.101, "dur": 2.938,
    "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": 870, "tid": 870,
    "ts": 7429770605696.145, "dur": 1.213,
    "args": {
      "External id": 1132,"Record function id": 0, "finished": true, "Ev Idx": 1131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770605696.688, "dur": 0.451,
    "args": {
      "External id": 1133,"Record function id": 0, "finished": true, "Ev Idx": 1132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770605698.393, "dur": 31.726,
    "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": 870, "tid": 870,
    "ts": 7429770605698.939, "dur": 1.468,
    "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": 870, "tid": 870,
    "ts": 7429770605699.234, "dur": 1.053,
    "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": 870, "tid": 870,
    "ts": 7429770605700.746, "dur": 27.451,
    "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": 870, "tid": 870,
    "ts": 7429770605711.336, "dur": 1.225,
    "args": {
      "External id": 1138,"Record function id": 0, "finished": true, "Ev Idx": 1137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770605728.825, "dur": 1.028,
    "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": 870, "tid": 870,
    "ts": 7429770605738.834, "dur": 11.271,
    "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": 870, "tid": 870,
    "ts": 7429770605758.199, "dur": 34.472,
    "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": 870, "tid": 870,
    "ts": 7429770605758.502, "dur": 2.450,
    "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": 870, "tid": 870,
    "ts": 7429770605759.525, "dur": 0.915,
    "args": {
      "External id": 1143,"Record function id": 0, "finished": true, "Ev Idx": 1142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770605759.969, "dur": 0.350,
    "args": {
      "External id": 1144,"Record function id": 0, "finished": true, "Ev Idx": 1143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770605761.274, "dur": 31.305,
    "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": 870, "tid": 870,
    "ts": 7429770605761.662, "dur": 1.581,
    "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": 870, "tid": 870,
    "ts": 7429770605761.948, "dur": 1.184,
    "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": 870, "tid": 870,
    "ts": 7429770605763.529, "dur": 27.196,
    "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": 870, "tid": 870,
    "ts": 7429770605775.428, "dur": 1.140,
    "args": {
      "External id": 1149,"Record function id": 0, "finished": true, "Ev Idx": 1148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770605791.350, "dur": 0.974,
    "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": 870, "tid": 870,
    "ts": 7429770605794.140, "dur": 9.110,
    "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": 870, "tid": 870,
    "ts": 7429770605809.703, "dur": 46.987,
    "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": 870, "tid": 870,
    "ts": 7429770605810.068, "dur": 2.380,
    "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": 870, "tid": 870,
    "ts": 7429770605811.075, "dur": 0.886,
    "args": {
      "External id": 1154,"Record function id": 0, "finished": true, "Ev Idx": 1153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770605811.457, "dur": 0.337,
    "args": {
      "External id": 1155,"Record function id": 0, "finished": true, "Ev Idx": 1154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770605812.743, "dur": 43.845,
    "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": 870, "tid": 870,
    "ts": 7429770605813.093, "dur": 1.510,
    "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": 870, "tid": 870,
    "ts": 7429770605813.382, "dur": 1.101,
    "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": 870, "tid": 870,
    "ts": 7429770605814.904, "dur": 39.829,
    "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": 870, "tid": 870,
    "ts": 7429770605825.817, "dur": 1.072,
    "args": {
      "External id": 1160,"Record function id": 0, "finished": true, "Ev Idx": 1159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770605855.372, "dur": 0.947,
    "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": 870, "tid": 870,
    "ts": 7429770605859.222, "dur": 7.707,
    "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": 870, "tid": 870,
    "ts": 7429770605887.083, "dur": 0.141,
    "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": 870, "tid": 870,
    "ts": 7429770605888.958, "dur": 9.282,
    "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": 870, "tid": 870,
    "ts": 7429770605890.359, "dur": 0.255,
    "args": {
      "External id": 1165,"Record function id": 0, "finished": true, "Ev Idx": 1164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770605890.906, "dur": 0.073,
    "args": {
      "External id": 1166,"Record function id": 0, "finished": true, "Ev Idx": 1165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770605900.413, "dur": 10.712,
    "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": 870, "tid": 870,
    "ts": 7429770605914.206, "dur": 8.208,
    "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": 870, "tid": 870,
    "ts": 7429770605923.984, "dur": 9.062,
    "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": 870, "tid": 870,
    "ts": 7429770605934.530, "dur": 7.634,
    "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": 870, "tid": 870,
    "ts": 7429770605944.559, "dur": 0.087,
    "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": 870, "tid": 870,
    "ts": 7429770605945.281, "dur": 9.005,
    "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": 870, "tid": 870,
    "ts": 7429770605972.796, "dur": 36.878,
    "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": 870, "tid": 870,
    "ts": 7429770605973.184, "dur": 2.907,
    "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": 870, "tid": 870,
    "ts": 7429770605974.398, "dur": 1.113,
    "args": {
      "External id": 1175,"Record function id": 0, "finished": true, "Ev Idx": 1174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770605974.925, "dur": 0.434,
    "args": {
      "External id": 1176,"Record function id": 0, "finished": true, "Ev Idx": 1175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770605976.429, "dur": 33.109,
    "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": 870, "tid": 870,
    "ts": 7429770605976.912, "dur": 1.459,
    "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": 870, "tid": 870,
    "ts": 7429770605977.217, "dur": 1.037,
    "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": 870, "tid": 870,
    "ts": 7429770605978.731, "dur": 28.856,
    "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": 870, "tid": 870,
    "ts": 7429770605991.376, "dur": 1.170,
    "args": {
      "External id": 1181,"Record function id": 0, "finished": true, "Ev Idx": 1180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770606008.263, "dur": 1.016,
    "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": 870, "tid": 870,
    "ts": 7429770606012.464, "dur": 1.241,
    "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": 870, "tid": 870,
    "ts": 7429770606015.961, "dur": 1.919,
    "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": 870, "tid": 870,
    "ts": 7429770606016.819, "dur": 0.433,
    "args": {
      "External id": 1185,"Record function id": 0, "finished": true, "Ev Idx": 1184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770606025.233, "dur": 30.761,
    "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": 870, "tid": 870,
    "ts": 7429770606025.538, "dur": 2.122,
    "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": 870, "tid": 870,
    "ts": 7429770606026.479, "dur": 0.737,
    "args": {
      "External id": 1188,"Record function id": 0, "finished": true, "Ev Idx": 1187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770606026.838, "dur": 0.251,
    "args": {
      "External id": 1189,"Record function id": 0, "finished": true, "Ev Idx": 1188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770606027.983, "dur": 27.914,
    "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": 870, "tid": 870,
    "ts": 7429770606028.401, "dur": 1.941,
    "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": 870, "tid": 870,
    "ts": 7429770606028.693, "dur": 1.534,
    "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": 870, "tid": 870,
    "ts": 7429770606030.666, "dur": 23.510,
    "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": 870, "tid": 870,
    "ts": 7429770606038.594, "dur": 1.109,
    "args": {
      "External id": 1194,"Record function id": 0, "finished": true, "Ev Idx": 1193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770606054.736, "dur": 0.909,
    "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": 870, "tid": 870,
    "ts": 7429770606058.314, "dur": 1.103,
    "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": 870, "tid": 870,
    "ts": 7429770606060.894, "dur": 1.714,
    "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": 870, "tid": 870,
    "ts": 7429770606061.640, "dur": 0.447,
    "args": {
      "External id": 1198,"Record function id": 0, "finished": true, "Ev Idx": 1197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770606089.456, "dur": 39.845,
    "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": 870, "tid": 870,
    "ts": 7429770606090.032, "dur": 4.078,
    "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": 870, "tid": 870,
    "ts": 7429770606091.746, "dur": 1.653,
    "args": {
      "External id": 1201,"Record function id": 0, "finished": true, "Ev Idx": 1200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770606092.662, "dur": 0.496,
    "args": {
      "External id": 1202,"Record function id": 0, "finished": true, "Ev Idx": 1201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770606094.457, "dur": 34.749,
    "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": 870, "tid": 870,
    "ts": 7429770606095.033, "dur": 2.327,
    "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": 870, "tid": 870,
    "ts": 7429770606095.365, "dur": 1.910,
    "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": 870, "tid": 870,
    "ts": 7429770606097.672, "dur": 29.650,
    "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": 870, "tid": 870,
    "ts": 7429770606108.446, "dur": 1.278,
    "args": {
      "External id": 1207,"Record function id": 0, "finished": true, "Ev Idx": 1206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770606127.910, "dur": 0.959,
    "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": 870, "tid": 870,
    "ts": 7429770606132.983, "dur": 1.191,
    "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": 870, "tid": 870,
    "ts": 7429770606135.898, "dur": 1.556,
    "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": 870, "tid": 870,
    "ts": 7429770606136.609, "dur": 0.347,
    "args": {
      "External id": 1211,"Record function id": 0, "finished": true, "Ev Idx": 1210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770606140.347, "dur": 2.073,
    "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": 870, "tid": 870,
    "ts": 7429770606141.623, "dur": 0.329,
    "args": {
      "External id": 1213,"Record function id": 0, "finished": true, "Ev Idx": 1212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770606143.415, "dur": 1.138,
    "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": 870, "tid": 870,
    "ts": 7429770606143.830, "dur": 0.442,
    "args": {
      "External id": 1215,"Record function id": 0, "finished": true, "Ev Idx": 1214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770606145.508, "dur": 10.822,
    "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": 870, "tid": 870,
    "ts": 7429770606161.517, "dur": 3.938,
    "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": 870, "tid": 870,
    "ts": 7429770606163.666, "dur": 0.454,
    "args": {
      "External id": 1218,"Record function id": 0, "finished": true, "Ev Idx": 1217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770606167.657, "dur": 3.080,
    "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": 870, "tid": 870,
    "ts": 7429770606169.490, "dur": 0.449,
    "args": {
      "External id": 1220,"Record function id": 0, "finished": true, "Ev Idx": 1219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770606172.522, "dur": 11.163,
    "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": 870, "tid": 870,
    "ts": 7429770606187.705, "dur": 16.738,
    "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": 870, "tid": 870,
    "ts": 7429770606207.515, "dur": 8.706,
    "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": 870, "tid": 870,
    "ts": 7429770606217.912, "dur": 9.390,
    "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": 870, "tid": 870,
    "ts": 7429770606229.512, "dur": 8.254,
    "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": 870, "tid": 870,
    "ts": 7429770606240.171, "dur": 2.865,
    "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": 870, "tid": 870,
    "ts": 7429770606241.923, "dur": 0.380,
    "args": {
      "External id": 1227,"Record function id": 0, "finished": true, "Ev Idx": 1226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770606244.532, "dur": 1.818,
    "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": 870, "tid": 870,
    "ts": 7429770606245.514, "dur": 0.306,
    "args": {
      "External id": 1229,"Record function id": 0, "finished": true, "Ev Idx": 1228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770606247.332, "dur": 7.768,
    "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": 870, "tid": 870,
    "ts": 7429770606256.921, "dur": 11.586,
    "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": 870, "tid": 870,
    "ts": 7429770606270.608, "dur": 6.744,
    "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": 870, "tid": 870,
    "ts": 7429770606278.871, "dur": 6.910,
    "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": 870, "tid": 870,
    "ts": 7429770606308.884, "dur": 2.323,
    "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": 870, "tid": 870,
    "ts": 7429770606310.005, "dur": 0.428,
    "args": {
      "External id": 1235,"Record function id": 0, "finished": true, "Ev Idx": 1234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770606311.927, "dur": 1.935,
    "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": 870, "tid": 870,
    "ts": 7429770606313.034, "dur": 0.308,
    "args": {
      "External id": 1237,"Record function id": 0, "finished": true, "Ev Idx": 1236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770606314.598, "dur": 2.948,
    "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": 870, "tid": 870,
    "ts": 7429770606316.251, "dur": 0.695,
    "args": {
      "External id": 1239,"Record function id": 0, "finished": true, "Ev Idx": 1238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770606318.222, "dur": 2.191,
    "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": 870, "tid": 870,
    "ts": 7429770606319.492, "dur": 0.402,
    "args": {
      "External id": 1241,"Record function id": 0, "finished": true, "Ev Idx": 1240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770606321.057, "dur": 2.500,
    "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": 870, "tid": 870,
    "ts": 7429770606322.616, "dur": 0.483,
    "args": {
      "External id": 1243,"Record function id": 0, "finished": true, "Ev Idx": 1242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770606326.448, "dur": 3.655,
    "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": 870, "tid": 870,
    "ts": 7429770606328.828, "dur": 0.600,
    "args": {
      "External id": 1245,"Record function id": 0, "finished": true, "Ev Idx": 1244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770606332.468, "dur": 19.870,
    "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": 870, "tid": 870,
    "ts": 7429770606334.753, "dur": 15.472,
    "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": 870, "tid": 870,
    "ts": 7429770606336.894, "dur": 2.987,
    "args": {
      "External id": 1248,"Record function id": 0, "finished": true, "Ev Idx": 1247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770606337.780, "dur": 1.826,
    "args": {
      "External id": 1249,"Record function id": 0, "finished": true, "Ev Idx": 1248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770606340.742, "dur": 8.985,
    "args": {
      "External id": 1250,"Record function id": 0, "finished": true, "Ev Idx": 1249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770606350.849, "dur": 1.250,
    "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": 870, "tid": 870,
    "ts": 7429770606355.606, "dur": 2.566,
    "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": 870, "tid": 870,
    "ts": 7429770606357.166, "dur": 0.447,
    "args": {
      "External id": 1253,"Record function id": 0, "finished": true, "Ev Idx": 1252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770606358.806, "dur": 2.495,
    "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": 870, "tid": 870,
    "ts": 7429770606360.525, "dur": 0.276,
    "args": {
      "External id": 1255,"Record function id": 0, "finished": true, "Ev Idx": 1254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770606361.795, "dur": 2.392,
    "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": 870, "tid": 870,
    "ts": 7429770606363.409, "dur": 0.333,
    "args": {
      "External id": 1257,"Record function id": 0, "finished": true, "Ev Idx": 1256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770606364.779, "dur": 2.010,
    "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": 870, "tid": 870,
    "ts": 7429770606365.814, "dur": 0.445,
    "args": {
      "External id": 1259,"Record function id": 0, "finished": true, "Ev Idx": 1258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770606367.382, "dur": 1.447,
    "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": 870, "tid": 870,
    "ts": 7429770606368.083, "dur": 0.280,
    "args": {
      "External id": 1261,"Record function id": 0, "finished": true, "Ev Idx": 1260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770606370.333, "dur": 1.571,
    "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": 870, "tid": 870,
    "ts": 7429770606371.130, "dur": 0.335,
    "args": {
      "External id": 1263,"Record function id": 0, "finished": true, "Ev Idx": 1262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770606373.273, "dur": 13.904,
    "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": 870, "tid": 870,
    "ts": 7429770606374.170, "dur": 10.827,
    "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": 870, "tid": 870,
    "ts": 7429770606374.828, "dur": 1.613,
    "args": {
      "External id": 1266,"Record function id": 0, "finished": true, "Ev Idx": 1265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770606375.182, "dur": 1.156,
    "args": {
      "External id": 1267,"Record function id": 0, "finished": true, "Ev Idx": 1266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770606376.738, "dur": 7.916,
    "args": {
      "External id": 1268,"Record function id": 0, "finished": true, "Ev Idx": 1267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770606385.369, "dur": 1.642,
    "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": 870, "tid": 870,
    "ts": 7429770606389.225, "dur": 2.036,
    "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": 870, "tid": 870,
    "ts": 7429770606390.426, "dur": 0.352,
    "args": {
      "External id": 1271,"Record function id": 0, "finished": true, "Ev Idx": 1270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770606392.792, "dur": 47.247,
    "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": 870, "tid": 870,
    "ts": 7429770606393.488, "dur": 1.644,
    "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": 870, "tid": 870,
    "ts": 7429770606394.357, "dur": 0.326,
    "args": {
      "External id": 1274,"Record function id": 0, "finished": true, "Ev Idx": 1273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770606395.460, "dur": 4.000,
    "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": 870, "tid": 870,
    "ts": 7429770606397.294, "dur": 2.029,
    "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": 870, "tid": 870,
    "ts": 7429770606399.907, "dur": 1.674,
    "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": 870, "tid": 870,
    "ts": 7429770606400.920, "dur": 0.299,
    "args": {
      "External id": 1278,"Record function id": 0, "finished": true, "Ev Idx": 1277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770606402.011, "dur": 2.691,
    "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": 870, "tid": 870,
    "ts": 7429770606402.942, "dur": 1.651,
    "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": 870, "tid": 870,
    "ts": 7429770606405.540, "dur": 32.645,
    "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": 870, "tid": 870,
    "ts": 7429770606419.099, "dur": 1.088,
    "args": {
      "External id": 1282,"Record function id": 0, "finished": true, "Ev Idx": 1281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770606438.820, "dur": 0.917,
    "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": 870, "tid": 870,
    "ts": 7429770606443.029, "dur": 9.852,
    "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": 870, "tid": 870,
    "ts": 7429770606455.985, "dur": 1.713,
    "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": 870, "tid": 870,
    "ts": 7429770606456.692, "dur": 0.375,
    "args": {
      "External id": 1286,"Record function id": 0, "finished": true, "Ev Idx": 1285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770606458.282, "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": 870, "tid": 870,
    "ts": 7429770606458.856, "dur": 0.291,
    "args": {
      "External id": 1288,"Record function id": 0, "finished": true, "Ev Idx": 1287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770606460.134, "dur": 1.048,
    "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": 870, "tid": 870,
    "ts": 7429770606460.600, "dur": 0.232,
    "args": {
      "External id": 1290,"Record function id": 0, "finished": true, "Ev Idx": 1289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770606462.748, "dur": 7.938,
    "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": 870, "tid": 870,
    "ts": 7429770606476.157, "dur": 15.105,
    "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": 870, "tid": 870,
    "ts": 7429770606477.389, "dur": 0.413,
    "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": 870, "tid": 870,
    "ts": 7429770606478.789, "dur": 12.191,
    "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": 870, "tid": 870,
    "ts": 7429770606494.111, "dur": 0.210,
    "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": 870, "tid": 870,
    "ts": 7429770606505.808, "dur": 0.443,
    "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": 870, "tid": 870,
    "ts": 7429770606507.954, "dur": 41.424,
    "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": 870, "tid": 870,
    "ts": 7429770606508.739, "dur": 2.278,
    "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": 870, "tid": 870,
    "ts": 7429770606510.095, "dur": 0.382,
    "args": {
      "External id": 1299,"Record function id": 0, "finished": true, "Ev Idx": 1298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770606511.536, "dur": 2.629,
    "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": 870, "tid": 870,
    "ts": 7429770606512.037, "dur": 1.996,
    "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": 870, "tid": 870,
    "ts": 7429770606514.564, "dur": 2.125,
    "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": 870, "tid": 870,
    "ts": 7429770606516.091, "dur": 0.227,
    "args": {
      "External id": 1303,"Record function id": 0, "finished": true, "Ev Idx": 1302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770606516.988, "dur": 2.710,
    "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": 870, "tid": 870,
    "ts": 7429770606517.346, "dur": 2.285,
    "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": 870, "tid": 870,
    "ts": 7429770606520.234, "dur": 27.169,
    "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": 870, "tid": 870,
    "ts": 7429770606530.401, "dur": 1.145,
    "args": {
      "External id": 1307,"Record function id": 0, "finished": true, "Ev Idx": 1306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770606548.105, "dur": 0.985,
    "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": 870, "tid": 870,
    "ts": 7429770606551.285, "dur": 1.899,
    "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": 870, "tid": 870,
    "ts": 7429770606552.317, "dur": 0.338,
    "args": {
      "External id": 1310,"Record function id": 0, "finished": true, "Ev Idx": 1309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770606555.842, "dur": 12.119,
    "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": 870, "tid": 870,
    "ts": 7429770606556.535, "dur": 11.289,
    "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": 870, "tid": 870,
    "ts": 7429770606557.494, "dur": 1.741,
    "args": {
      "External id": 1313,"Record function id": 0, "finished": true, "Ev Idx": 1312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770606557.884, "dur": 1.219,
    "args": {
      "External id": 1314,"Record function id": 0, "finished": true, "Ev Idx": 1313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770606559.588, "dur": 7.744,
    "args": {
      "External id": 1315,"Record function id": 0, "finished": true, "Ev Idx": 1314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770606573.098, "dur": 1.472,
    "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": 870, "tid": 870,
    "ts": 7429770606573.401, "dur": 1.071,
    "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": 870, "tid": 870,
    "ts": 7429770606589.188, "dur": 35.553,
    "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": 870, "tid": 870,
    "ts": 7429770606589.855, "dur": 2.954,
    "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": 870, "tid": 870,
    "ts": 7429770606591.264, "dur": 1.006,
    "args": {
      "External id": 1320,"Record function id": 0, "finished": true, "Ev Idx": 1319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770606591.823, "dur": 0.306,
    "args": {
      "External id": 1321,"Record function id": 0, "finished": true, "Ev Idx": 1320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770606593.167, "dur": 31.424,
    "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": 870, "tid": 870,
    "ts": 7429770606593.781, "dur": 1.515,
    "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": 870, "tid": 870,
    "ts": 7429770606594.181, "dur": 1.030,
    "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": 870, "tid": 870,
    "ts": 7429770606595.776, "dur": 26.910,
    "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": 870, "tid": 870,
    "ts": 7429770606607.076, "dur": 1.126,
    "args": {
      "External id": 1326,"Record function id": 0, "finished": true, "Ev Idx": 1325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770606623.316, "dur": 1.016,
    "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": 870, "tid": 870,
    "ts": 7429770606629.712, "dur": 8.004,
    "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": 870, "tid": 870,
    "ts": 7429770606647.571, "dur": 0.162,
    "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": 870, "tid": 870,
    "ts": 7429770606649.633, "dur": 11.153,
    "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": 870, "tid": 870,
    "ts": 7429770606651.495, "dur": 0.409,
    "args": {
      "External id": 1331,"Record function id": 0, "finished": true, "Ev Idx": 1330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770606652.310, "dur": 0.125,
    "args": {
      "External id": 1332,"Record function id": 0, "finished": true, "Ev Idx": 1331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770606663.271, "dur": 13.794,
    "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": 870, "tid": 870,
    "ts": 7429770606680.451, "dur": 8.729,
    "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": 870, "tid": 870,
    "ts": 7429770606691.112, "dur": 10.531,
    "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": 870, "tid": 870,
    "ts": 7429770606703.438, "dur": 8.583,
    "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": 870, "tid": 870,
    "ts": 7429770606714.421, "dur": 0.249,
    "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": 870, "tid": 870,
    "ts": 7429770606715.442, "dur": 9.248,
    "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": 870, "tid": 870,
    "ts": 7429770606739.248, "dur": 34.943,
    "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": 870, "tid": 870,
    "ts": 7429770606739.636, "dur": 2.950,
    "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": 870, "tid": 870,
    "ts": 7429770606740.648, "dur": 1.326,
    "args": {
      "External id": 1341,"Record function id": 0, "finished": true, "Ev Idx": 1340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770606741.278, "dur": 0.517,
    "args": {
      "External id": 1342,"Record function id": 0, "finished": true, "Ev Idx": 1341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770606742.929, "dur": 31.126,
    "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": 870, "tid": 870,
    "ts": 7429770606743.448, "dur": 1.853,
    "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": 870, "tid": 870,
    "ts": 7429770606744.100, "dur": 1.085,
    "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": 870, "tid": 870,
    "ts": 7429770606745.626, "dur": 26.455,
    "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": 870, "tid": 870,
    "ts": 7429770606756.445, "dur": 1.156,
    "args": {
      "External id": 1347,"Record function id": 0, "finished": true, "Ev Idx": 1346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770606772.661, "dur": 1.107,
    "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": 870, "tid": 870,
    "ts": 7429770606782.754, "dur": 9.634,
    "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": 870, "tid": 870,
    "ts": 7429770606800.474, "dur": 34.270,
    "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": 870, "tid": 870,
    "ts": 7429770606801.072, "dur": 2.979,
    "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": 870, "tid": 870,
    "ts": 7429770606802.261, "dur": 1.080,
    "args": {
      "External id": 1352,"Record function id": 0, "finished": true, "Ev Idx": 1351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770606802.800, "dur": 0.368,
    "args": {
      "External id": 1353,"Record function id": 0, "finished": true, "Ev Idx": 1352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770606804.556, "dur": 30.090,
    "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": 870, "tid": 870,
    "ts": 7429770606804.951, "dur": 2.583,
    "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": 870, "tid": 870,
    "ts": 7429770606805.237, "dur": 2.215,
    "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": 870, "tid": 870,
    "ts": 7429770606807.836, "dur": 25.042,
    "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": 870, "tid": 870,
    "ts": 7429770606817.397, "dur": 1.143,
    "args": {
      "External id": 1358,"Record function id": 0, "finished": true, "Ev Idx": 1357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770606833.527, "dur": 0.882,
    "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": 870, "tid": 870,
    "ts": 7429770606836.461, "dur": 9.068,
    "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": 870, "tid": 870,
    "ts": 7429770606852.468, "dur": 42.124,
    "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": 870, "tid": 870,
    "ts": 7429770606852.903, "dur": 2.070,
    "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": 870, "tid": 870,
    "ts": 7429770606853.727, "dur": 0.796,
    "args": {
      "External id": 1363,"Record function id": 0, "finished": true, "Ev Idx": 1362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770606854.108, "dur": 0.296,
    "args": {
      "External id": 1364,"Record function id": 0, "finished": true, "Ev Idx": 1363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770606855.273, "dur": 39.219,
    "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": 870, "tid": 870,
    "ts": 7429770606855.685, "dur": 1.488,
    "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": 870, "tid": 870,
    "ts": 7429770606855.955, "dur": 1.114,
    "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": 870, "tid": 870,
    "ts": 7429770606857.465, "dur": 35.280,
    "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": 870, "tid": 870,
    "ts": 7429770606866.722, "dur": 1.078,
    "args": {
      "External id": 1369,"Record function id": 0, "finished": true, "Ev Idx": 1368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770606893.366, "dur": 0.864,
    "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": 870, "tid": 870,
    "ts": 7429770606897.125, "dur": 8.278,
    "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": 870, "tid": 870,
    "ts": 7429770606924.551, "dur": 0.146,
    "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": 870, "tid": 870,
    "ts": 7429770606926.284, "dur": 11.699,
    "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": 870, "tid": 870,
    "ts": 7429770606927.607, "dur": 0.207,
    "args": {
      "External id": 1374,"Record function id": 0, "finished": true, "Ev Idx": 1373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770606928.152, "dur": 0.082,
    "args": {
      "External id": 1375,"Record function id": 0, "finished": true, "Ev Idx": 1374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770606940.328, "dur": 11.012,
    "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": 870, "tid": 870,
    "ts": 7429770606954.525, "dur": 8.777,
    "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": 870, "tid": 870,
    "ts": 7429770606965.069, "dur": 7.854,
    "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": 870, "tid": 870,
    "ts": 7429770606974.477, "dur": 8.255,
    "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": 870, "tid": 870,
    "ts": 7429770606985.473, "dur": 0.283,
    "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": 870, "tid": 870,
    "ts": 7429770606986.470, "dur": 9.404,
    "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": 870, "tid": 870,
    "ts": 7429770607014.672, "dur": 41.080,
    "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": 870, "tid": 870,
    "ts": 7429770607015.049, "dur": 4.360,
    "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": 870, "tid": 870,
    "ts": 7429770607016.722, "dur": 1.594,
    "args": {
      "External id": 1384,"Record function id": 0, "finished": true, "Ev Idx": 1383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770607017.661, "dur": 0.509,
    "args": {
      "External id": 1385,"Record function id": 0, "finished": true, "Ev Idx": 1384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770607019.882, "dur": 35.764,
    "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": 870, "tid": 870,
    "ts": 7429770607020.801, "dur": 1.680,
    "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": 870, "tid": 870,
    "ts": 7429770607021.315, "dur": 1.068,
    "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": 870, "tid": 870,
    "ts": 7429770607023.094, "dur": 30.449,
    "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": 870, "tid": 870,
    "ts": 7429770607035.012, "dur": 1.248,
    "args": {
      "External id": 1390,"Record function id": 0, "finished": true, "Ev Idx": 1389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770607054.176, "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": 870, "tid": 870,
    "ts": 7429770607058.763, "dur": 1.383,
    "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": 870, "tid": 870,
    "ts": 7429770607062.265, "dur": 2.290,
    "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": 870, "tid": 870,
    "ts": 7429770607063.178, "dur": 0.679,
    "args": {
      "External id": 1394,"Record function id": 0, "finished": true, "Ev Idx": 1393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770607072.021, "dur": 31.208,
    "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": 870, "tid": 870,
    "ts": 7429770607072.348, "dur": 2.171,
    "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": 870, "tid": 870,
    "ts": 7429770607073.050, "dur": 1.032,
    "args": {
      "External id": 1397,"Record function id": 0, "finished": true, "Ev Idx": 1396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770607073.564, "dur": 0.420,
    "args": {
      "External id": 1398,"Record function id": 0, "finished": true, "Ev Idx": 1397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770607074.836, "dur": 28.291,
    "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": 870, "tid": 870,
    "ts": 7429770607075.328, "dur": 1.475,
    "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": 870, "tid": 870,
    "ts": 7429770607075.635, "dur": 1.054,
    "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": 870, "tid": 870,
    "ts": 7429770607077.127, "dur": 24.229,
    "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": 870, "tid": 870,
    "ts": 7429770607085.884, "dur": 1.178,
    "args": {
      "External id": 1403,"Record function id": 0, "finished": true, "Ev Idx": 1402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770607101.917, "dur": 0.937,
    "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": 870, "tid": 870,
    "ts": 7429770607105.472, "dur": 1.105,
    "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": 870, "tid": 870,
    "ts": 7429770607107.997, "dur": 1.563,
    "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": 870, "tid": 870,
    "ts": 7429770607108.763, "dur": 0.318,
    "args": {
      "External id": 1407,"Record function id": 0, "finished": true, "Ev Idx": 1406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770607116.167, "dur": 31.456,
    "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": 870, "tid": 870,
    "ts": 7429770607116.487, "dur": 1.785,
    "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": 870, "tid": 870,
    "ts": 7429770607117.190, "dur": 0.712,
    "args": {
      "External id": 1410,"Record function id": 0, "finished": true, "Ev Idx": 1409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770607117.534, "dur": 0.240,
    "args": {
      "External id": 1411,"Record function id": 0, "finished": true, "Ev Idx": 1410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770607118.525, "dur": 29.001,
    "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": 870, "tid": 870,
    "ts": 7429770607118.924, "dur": 1.357,
    "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": 870, "tid": 870,
    "ts": 7429770607119.326, "dur": 0.848,
    "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": 870, "tid": 870,
    "ts": 7429770607120.572, "dur": 25.306,
    "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": 870, "tid": 870,
    "ts": 7429770607130.327, "dur": 0.980,
    "args": {
      "External id": 1416,"Record function id": 0, "finished": true, "Ev Idx": 1415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770607146.376, "dur": 0.899,
    "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": 870, "tid": 870,
    "ts": 7429770607149.782, "dur": 1.062,
    "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": 870, "tid": 870,
    "ts": 7429770607152.700, "dur": 1.557,
    "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": 870, "tid": 870,
    "ts": 7429770607153.391, "dur": 0.432,
    "args": {
      "External id": 1420,"Record function id": 0, "finished": true, "Ev Idx": 1419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770607157.121, "dur": 1.682,
    "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": 870, "tid": 870,
    "ts": 7429770607158.031, "dur": 0.289,
    "args": {
      "External id": 1422,"Record function id": 0, "finished": true, "Ev Idx": 1421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770607159.829, "dur": 1.004,
    "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": 870, "tid": 870,
    "ts": 7429770607160.250, "dur": 0.299,
    "args": {
      "External id": 1424,"Record function id": 0, "finished": true, "Ev Idx": 1423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770607162.084, "dur": 9.059,
    "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": 870, "tid": 870,
    "ts": 7429770607176.028, "dur": 3.485,
    "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": 870, "tid": 870,
    "ts": 7429770607177.824, "dur": 0.495,
    "args": {
      "External id": 1427,"Record function id": 0, "finished": true, "Ev Idx": 1426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770607181.906, "dur": 2.189,
    "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": 870, "tid": 870,
    "ts": 7429770607182.762, "dur": 0.701,
    "args": {
      "External id": 1429,"Record function id": 0, "finished": true, "Ev Idx": 1428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770607185.643, "dur": 8.996,
    "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": 870, "tid": 870,
    "ts": 7429770607198.049, "dur": 14.665,
    "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": 870, "tid": 870,
    "ts": 7429770607215.398, "dur": 9.261,
    "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": 870, "tid": 870,
    "ts": 7429770607226.247, "dur": 8.392,
    "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": 870, "tid": 870,
    "ts": 7429770607236.534, "dur": 10.035,
    "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": 870, "tid": 870,
    "ts": 7429770607248.840, "dur": 2.562,
    "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": 870, "tid": 870,
    "ts": 7429770607250.275, "dur": 0.436,
    "args": {
      "External id": 1436,"Record function id": 0, "finished": true, "Ev Idx": 1435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770607252.887, "dur": 1.644,
    "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": 870, "tid": 870,
    "ts": 7429770607253.680, "dur": 0.309,
    "args": {
      "External id": 1438,"Record function id": 0, "finished": true, "Ev Idx": 1437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770607255.483, "dur": 7.723,
    "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": 870, "tid": 870,
    "ts": 7429770607265.451, "dur": 10.020,
    "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": 870, "tid": 870,
    "ts": 7429770607277.847, "dur": 7.002,
    "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": 870, "tid": 870,
    "ts": 7429770607286.452, "dur": 8.716,
    "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": 870, "tid": 870,
    "ts": 7429770607317.420, "dur": 2.494,
    "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": 870, "tid": 870,
    "ts": 7429770607318.678, "dur": 0.429,
    "args": {
      "External id": 1444,"Record function id": 0, "finished": true, "Ev Idx": 1443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770607320.598, "dur": 1.793,
    "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": 870, "tid": 870,
    "ts": 7429770607321.493, "dur": 0.349,
    "args": {
      "External id": 1446,"Record function id": 0, "finished": true, "Ev Idx": 1445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770607323.048, "dur": 2.611,
    "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": 870, "tid": 870,
    "ts": 7429770607324.367, "dur": 0.638,
    "args": {
      "External id": 1448,"Record function id": 0, "finished": true, "Ev Idx": 1447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770607326.274, "dur": 3.007,
    "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": 870, "tid": 870,
    "ts": 7429770607328.202, "dur": 0.537,
    "args": {
      "External id": 1450,"Record function id": 0, "finished": true, "Ev Idx": 1449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770607329.880, "dur": 2.861,
    "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": 870, "tid": 870,
    "ts": 7429770607331.784, "dur": 0.466,
    "args": {
      "External id": 1452,"Record function id": 0, "finished": true, "Ev Idx": 1451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770607335.358, "dur": 3.009,
    "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": 870, "tid": 870,
    "ts": 7429770607337.299, "dur": 0.413,
    "args": {
      "External id": 1454,"Record function id": 0, "finished": true, "Ev Idx": 1453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770607340.640, "dur": 17.976,
    "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": 870, "tid": 870,
    "ts": 7429770607342.685, "dur": 13.385,
    "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": 870, "tid": 870,
    "ts": 7429770607343.925, "dur": 2.561,
    "args": {
      "External id": 1457,"Record function id": 0, "finished": true, "Ev Idx": 1456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770607344.771, "dur": 1.524,
    "args": {
      "External id": 1458,"Record function id": 0, "finished": true, "Ev Idx": 1457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770607347.308, "dur": 8.340,
    "args": {
      "External id": 1459,"Record function id": 0, "finished": true, "Ev Idx": 1458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770607356.608, "dur": 1.738,
    "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": 870, "tid": 870,
    "ts": 7429770607362.201, "dur": 1.917,
    "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": 870, "tid": 870,
    "ts": 7429770607363.113, "dur": 0.395,
    "args": {
      "External id": 1462,"Record function id": 0, "finished": true, "Ev Idx": 1461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770607364.764, "dur": 2.973,
    "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": 870, "tid": 870,
    "ts": 7429770607366.804, "dur": 0.445,
    "args": {
      "External id": 1464,"Record function id": 0, "finished": true, "Ev Idx": 1463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770607368.267, "dur": 2.346,
    "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": 870, "tid": 870,
    "ts": 7429770607369.875, "dur": 0.313,
    "args": {
      "External id": 1466,"Record function id": 0, "finished": true, "Ev Idx": 1465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770607371.361, "dur": 1.870,
    "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": 870, "tid": 870,
    "ts": 7429770607372.447, "dur": 0.302,
    "args": {
      "External id": 1468,"Record function id": 0, "finished": true, "Ev Idx": 1467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770607373.853, "dur": 1.559,
    "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": 870, "tid": 870,
    "ts": 7429770607374.672, "dur": 0.276,
    "args": {
      "External id": 1470,"Record function id": 0, "finished": true, "Ev Idx": 1469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770607376.980, "dur": 1.708,
    "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": 870, "tid": 870,
    "ts": 7429770607377.758, "dur": 0.478,
    "args": {
      "External id": 1472,"Record function id": 0, "finished": true, "Ev Idx": 1471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770607380.443, "dur": 14.350,
    "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": 870, "tid": 870,
    "ts": 7429770607381.337, "dur": 11.192,
    "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": 870, "tid": 870,
    "ts": 7429770607382.116, "dur": 1.783,
    "args": {
      "External id": 1475,"Record function id": 0, "finished": true, "Ev Idx": 1474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770607382.463, "dur": 1.316,
    "args": {
      "External id": 1476,"Record function id": 0, "finished": true, "Ev Idx": 1475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770607384.201, "dur": 8.035,
    "args": {
      "External id": 1477,"Record function id": 0, "finished": true, "Ev Idx": 1476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770607392.981, "dur": 1.621,
    "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": 870, "tid": 870,
    "ts": 7429770607396.902, "dur": 1.663,
    "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": 870, "tid": 870,
    "ts": 7429770607397.740, "dur": 0.352,
    "args": {
      "External id": 1480,"Record function id": 0, "finished": true, "Ev Idx": 1479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770607399.956, "dur": 46.312,
    "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": 870, "tid": 870,
    "ts": 7429770607400.501, "dur": 1.639,
    "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": 870, "tid": 870,
    "ts": 7429770607401.363, "dur": 0.332,
    "args": {
      "External id": 1483,"Record function id": 0, "finished": true, "Ev Idx": 1482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770607402.475, "dur": 3.185,
    "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": 870, "tid": 870,
    "ts": 7429770607404.015, "dur": 1.509,
    "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": 870, "tid": 870,
    "ts": 7429770607406.146, "dur": 1.340,
    "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": 870, "tid": 870,
    "ts": 7429770607406.842, "dur": 0.287,
    "args": {
      "External id": 1487,"Record function id": 0, "finished": true, "Ev Idx": 1486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770607407.776, "dur": 2.957,
    "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": 870, "tid": 870,
    "ts": 7429770607408.593, "dur": 2.023,
    "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": 870, "tid": 870,
    "ts": 7429770607411.586, "dur": 32.637,
    "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": 870, "tid": 870,
    "ts": 7429770607424.511, "dur": 1.270,
    "args": {
      "External id": 1491,"Record function id": 0, "finished": true, "Ev Idx": 1490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770607444.821, "dur": 1.099,
    "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": 870, "tid": 870,
    "ts": 7429770607449.195, "dur": 9.429,
    "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": 870, "tid": 870,
    "ts": 7429770607462.154, "dur": 1.837,
    "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": 870, "tid": 870,
    "ts": 7429770607462.875, "dur": 0.404,
    "args": {
      "External id": 1495,"Record function id": 0, "finished": true, "Ev Idx": 1494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770607464.591, "dur": 1.249,
    "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": 870, "tid": 870,
    "ts": 7429770607465.158, "dur": 0.305,
    "args": {
      "External id": 1497,"Record function id": 0, "finished": true, "Ev Idx": 1496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770607466.503, "dur": 1.016,
    "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": 870, "tid": 870,
    "ts": 7429770607466.947, "dur": 0.222,
    "args": {
      "External id": 1499,"Record function id": 0, "finished": true, "Ev Idx": 1498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770607469.138, "dur": 8.325,
    "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": 870, "tid": 870,
    "ts": 7429770607482.873, "dur": 12.417,
    "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": 870, "tid": 870,
    "ts": 7429770607483.818, "dur": 0.282,
    "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": 870, "tid": 870,
    "ts": 7429770607484.865, "dur": 10.184,
    "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": 870, "tid": 870,
    "ts": 7429770607497.959, "dur": 0.166,
    "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": 870, "tid": 870,
    "ts": 7429770607508.322, "dur": 0.342,
    "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": 870, "tid": 870,
    "ts": 7429770607510.339, "dur": 39.722,
    "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": 870, "tid": 870,
    "ts": 7429770607510.827, "dur": 2.195,
    "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": 870, "tid": 870,
    "ts": 7429770607512.116, "dur": 0.329,
    "args": {
      "External id": 1508,"Record function id": 0, "finished": true, "Ev Idx": 1507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770607513.407, "dur": 2.372,
    "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": 870, "tid": 870,
    "ts": 7429770607513.697, "dur": 1.954,
    "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": 870, "tid": 870,
    "ts": 7429770607516.214, "dur": 2.246,
    "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": 870, "tid": 870,
    "ts": 7429770607517.661, "dur": 0.429,
    "args": {
      "External id": 1512,"Record function id": 0, "finished": true, "Ev Idx": 1511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770607518.885, "dur": 2.157,
    "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": 870, "tid": 870,
    "ts": 7429770607519.370, "dur": 1.591,
    "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": 870, "tid": 870,
    "ts": 7429770607521.664, "dur": 26.510,
    "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": 870, "tid": 870,
    "ts": 7429770607531.659, "dur": 1.065,
    "args": {
      "External id": 1516,"Record function id": 0, "finished": true, "Ev Idx": 1515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770607548.755, "dur": 1.024,
    "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": 870, "tid": 870,
    "ts": 7429770607552.062, "dur": 1.809,
    "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": 870, "tid": 870,
    "ts": 7429770607552.983, "dur": 0.340,
    "args": {
      "External id": 1519,"Record function id": 0, "finished": true, "Ev Idx": 1518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770607556.019, "dur": 11.790,
    "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": 870, "tid": 870,
    "ts": 7429770607556.526, "dur": 11.134,
    "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": 870, "tid": 870,
    "ts": 7429770607557.600, "dur": 1.926,
    "args": {
      "External id": 1522,"Record function id": 0, "finished": true, "Ev Idx": 1521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770607558.135, "dur": 1.248,
    "args": {
      "External id": 1523,"Record function id": 0, "finished": true, "Ev Idx": 1522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770607559.929, "dur": 7.303,
    "args": {
      "External id": 1524,"Record function id": 0, "finished": true, "Ev Idx": 1523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770607573.399, "dur": 1.548,
    "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": 870, "tid": 870,
    "ts": 7429770607573.691, "dur": 1.117,
    "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": 870, "tid": 870,
    "ts": 7429770607587.462, "dur": 40.298,
    "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": 870, "tid": 870,
    "ts": 7429770607588.254, "dur": 2.684,
    "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": 870, "tid": 870,
    "ts": 7429770607589.519, "dur": 0.846,
    "args": {
      "External id": 1529,"Record function id": 0, "finished": true, "Ev Idx": 1528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770607589.931, "dur": 0.312,
    "args": {
      "External id": 1530,"Record function id": 0, "finished": true, "Ev Idx": 1529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770607591.342, "dur": 36.280,
    "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": 870, "tid": 870,
    "ts": 7429770607591.814, "dur": 2.717,
    "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": 870, "tid": 870,
    "ts": 7429770607592.377, "dur": 2.045,
    "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": 870, "tid": 870,
    "ts": 7429770607595.035, "dur": 30.559,
    "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": 870, "tid": 870,
    "ts": 7429770607607.002, "dur": 1.301,
    "args": {
      "External id": 1535,"Record function id": 0, "finished": true, "Ev Idx": 1534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770607626.303, "dur": 1.053,
    "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": 870, "tid": 870,
    "ts": 7429770607663.408, "dur": 9.608,
    "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": 870, "tid": 870,
    "ts": 7429770607681.592, "dur": 0.140,
    "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": 870, "tid": 870,
    "ts": 7429770607683.584, "dur": 11.241,
    "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": 870, "tid": 870,
    "ts": 7429770607685.799, "dur": 0.373,
    "args": {
      "External id": 1540,"Record function id": 0, "finished": true, "Ev Idx": 1539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770607686.535, "dur": 0.069,
    "args": {
      "External id": 1541,"Record function id": 0, "finished": true, "Ev Idx": 1540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770607697.497, "dur": 10.834,
    "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": 870, "tid": 870,
    "ts": 7429770607711.485, "dur": 8.448,
    "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": 870, "tid": 870,
    "ts": 7429770607721.833, "dur": 9.265,
    "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": 870, "tid": 870,
    "ts": 7429770607732.708, "dur": 9.068,
    "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": 870, "tid": 870,
    "ts": 7429770607744.288, "dur": 0.099,
    "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": 870, "tid": 870,
    "ts": 7429770607745.272, "dur": 7.722,
    "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": 870, "tid": 870,
    "ts": 7429770607770.440, "dur": 126.275,
    "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": 870, "tid": 870,
    "ts": 7429770607770.818, "dur": 2.928,
    "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": 870, "tid": 870,
    "ts": 7429770607772.007, "dur": 1.023,
    "args": {
      "External id": 1550,"Record function id": 0, "finished": true, "Ev Idx": 1549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770607772.494, "dur": 0.401,
    "args": {
      "External id": 1551,"Record function id": 0, "finished": true, "Ev Idx": 1550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770607774.107, "dur": 122.415,
    "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": 870, "tid": 870,
    "ts": 7429770607774.569, "dur": 1.526,
    "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": 870, "tid": 870,
    "ts": 7429770607774.858, "dur": 1.095,
    "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": 870, "tid": 870,
    "ts": 7429770607776.442, "dur": 117.429,
    "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": 870, "tid": 870,
    "ts": 7429770607789.039, "dur": 82.166,
    "args": {
      "External id": 1556,"Record function id": 0, "finished": true, "Ev Idx": 1555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770607894.684, "dur": 1.455,
    "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": 870, "tid": 870,
    "ts": 7429770607906.790, "dur": 11.853,
    "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": 870, "tid": 870,
    "ts": 7429770607927.881, "dur": 37.822,
    "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": 870, "tid": 870,
    "ts": 7429770607928.370, "dur": 3.573,
    "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": 870, "tid": 870,
    "ts": 7429770607929.616, "dur": 1.696,
    "args": {
      "External id": 1561,"Record function id": 0, "finished": true, "Ev Idx": 1560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770607930.672, "dur": 0.464,
    "args": {
      "External id": 1562,"Record function id": 0, "finished": true, "Ev Idx": 1561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770607932.487, "dur": 33.111,
    "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": 870, "tid": 870,
    "ts": 7429770607933.092, "dur": 3.023,
    "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": 870, "tid": 870,
    "ts": 7429770607933.565, "dur": 2.409,
    "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": 870, "tid": 870,
    "ts": 7429770607936.484, "dur": 27.235,
    "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": 870, "tid": 870,
    "ts": 7429770607946.719, "dur": 1.398,
    "args": {
      "External id": 1567,"Record function id": 0, "finished": true, "Ev Idx": 1566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770607964.362, "dur": 0.943,
    "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": 870, "tid": 870,
    "ts": 7429770607967.523, "dur": 8.997,
    "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": 870, "tid": 870,
    "ts": 7429770607983.466, "dur": 46.501,
    "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": 870, "tid": 870,
    "ts": 7429770607983.837, "dur": 2.180,
    "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": 870, "tid": 870,
    "ts": 7429770607984.721, "dur": 0.854,
    "args": {
      "External id": 1572,"Record function id": 0, "finished": true, "Ev Idx": 1571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770607985.109, "dur": 0.337,
    "args": {
      "External id": 1573,"Record function id": 0, "finished": true, "Ev Idx": 1572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770607986.324, "dur": 43.551,
    "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": 870, "tid": 870,
    "ts": 7429770607986.726, "dur": 1.358,
    "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": 870, "tid": 870,
    "ts": 7429770607987.012, "dur": 0.949,
    "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": 870, "tid": 870,
    "ts": 7429770607988.383, "dur": 39.605,
    "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": 870, "tid": 870,
    "ts": 7429770607998.885, "dur": 1.582,
    "args": {
      "External id": 1578,"Record function id": 0, "finished": true, "Ev Idx": 1577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770608028.597, "dur": 1.030,
    "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": 870, "tid": 870,
    "ts": 7429770608032.967, "dur": 7.662,
    "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": 870, "tid": 870,
    "ts": 7429770608061.288, "dur": 0.240,
    "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": 870, "tid": 870,
    "ts": 7429770608063.804, "dur": 9.586,
    "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": 870, "tid": 870,
    "ts": 7429770608065.264, "dur": 0.359,
    "args": {
      "External id": 1583,"Record function id": 0, "finished": true, "Ev Idx": 1582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770608065.919, "dur": 0.077,
    "args": {
      "External id": 1584,"Record function id": 0, "finished": true, "Ev Idx": 1583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770608076.346, "dur": 12.705,
    "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": 870, "tid": 870,
    "ts": 7429770608093.161, "dur": 11.680,
    "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": 870, "tid": 870,
    "ts": 7429770608106.608, "dur": 8.607,
    "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": 870, "tid": 870,
    "ts": 7429770608116.675, "dur": 7.948,
    "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": 870, "tid": 870,
    "ts": 7429770608127.240, "dur": 0.103,
    "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": 870, "tid": 870,
    "ts": 7429770608128.275, "dur": 7.406,
    "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": 870, "tid": 870,
    "ts": 7429770608155.549, "dur": 34.423,
    "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": 870, "tid": 870,
    "ts": 7429770608156.075, "dur": 2.687,
    "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": 870, "tid": 870,
    "ts": 7429770608157.003, "dur": 1.147,
    "args": {
      "External id": 1593,"Record function id": 0, "finished": true, "Ev Idx": 1592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770608157.587, "dur": 0.418,
    "args": {
      "External id": 1594,"Record function id": 0, "finished": true, "Ev Idx": 1593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770608159.058, "dur": 30.817,
    "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": 870, "tid": 870,
    "ts": 7429770608159.519, "dur": 1.428,
    "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": 870, "tid": 870,
    "ts": 7429770608159.810, "dur": 1.041,
    "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": 870, "tid": 870,
    "ts": 7429770608161.301, "dur": 26.818,
    "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": 870, "tid": 870,
    "ts": 7429770608171.956, "dur": 1.161,
    "args": {
      "External id": 1599,"Record function id": 0, "finished": true, "Ev Idx": 1598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770608188.689, "dur": 0.940,
    "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": 870, "tid": 870,
    "ts": 7429770608193.035, "dur": 1.298,
    "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": 870, "tid": 870,
    "ts": 7429770608196.303, "dur": 2.647,
    "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": 870, "tid": 870,
    "ts": 7429770608197.566, "dur": 0.467,
    "args": {
      "External id": 1603,"Record function id": 0, "finished": true, "Ev Idx": 1602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770608207.026, "dur": 32.426,
    "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": 870, "tid": 870,
    "ts": 7429770608207.497, "dur": 2.156,
    "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": 870, "tid": 870,
    "ts": 7429770608208.290, "dur": 0.933,
    "args": {
      "External id": 1606,"Record function id": 0, "finished": true, "Ev Idx": 1605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770608208.822, "dur": 0.266,
    "args": {
      "External id": 1607,"Record function id": 0, "finished": true, "Ev Idx": 1606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770608209.994, "dur": 29.359,
    "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": 870, "tid": 870,
    "ts": 7429770608210.441, "dur": 1.859,
    "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": 870, "tid": 870,
    "ts": 7429770608210.729, "dur": 1.471,
    "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": 870, "tid": 870,
    "ts": 7429770608212.595, "dur": 24.952,
    "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": 870, "tid": 870,
    "ts": 7429770608222.039, "dur": 1.249,
    "args": {
      "External id": 1612,"Record function id": 0, "finished": true, "Ev Idx": 1611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770608238.126, "dur": 0.969,
    "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": 870, "tid": 870,
    "ts": 7429770608241.719, "dur": 1.147,
    "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": 870, "tid": 870,
    "ts": 7429770608244.659, "dur": 1.714,
    "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": 870, "tid": 870,
    "ts": 7429770608245.534, "dur": 0.371,
    "args": {
      "External id": 1616,"Record function id": 0, "finished": true, "Ev Idx": 1615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770608254.322, "dur": 30.474,
    "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": 870, "tid": 870,
    "ts": 7429770608254.650, "dur": 1.867,
    "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": 870, "tid": 870,
    "ts": 7429770608255.373, "dur": 0.746,
    "args": {
      "External id": 1619,"Record function id": 0, "finished": true, "Ev Idx": 1618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770608255.749, "dur": 0.277,
    "args": {
      "External id": 1620,"Record function id": 0, "finished": true, "Ev Idx": 1619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770608256.787, "dur": 27.914,
    "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": 870, "tid": 870,
    "ts": 7429770608257.167, "dur": 1.300,
    "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": 870, "tid": 870,
    "ts": 7429770608257.489, "dur": 0.873,
    "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": 870, "tid": 870,
    "ts": 7429770608258.752, "dur": 24.356,
    "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": 870, "tid": 870,
    "ts": 7429770608268.385, "dur": 1.074,
    "args": {
      "External id": 1625,"Record function id": 0, "finished": true, "Ev Idx": 1624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770608283.648, "dur": 0.816,
    "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": 870, "tid": 870,
    "ts": 7429770608287.166, "dur": 1.048,
    "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": 870, "tid": 870,
    "ts": 7429770608290.017, "dur": 1.540,
    "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": 870, "tid": 870,
    "ts": 7429770608290.792, "dur": 0.320,
    "args": {
      "External id": 1629,"Record function id": 0, "finished": true, "Ev Idx": 1628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770608294.484, "dur": 2.123,
    "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": 870, "tid": 870,
    "ts": 7429770608295.788, "dur": 0.300,
    "args": {
      "External id": 1631,"Record function id": 0, "finished": true, "Ev Idx": 1630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770608297.676, "dur": 1.042,
    "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": 870, "tid": 870,
    "ts": 7429770608298.100, "dur": 0.322,
    "args": {
      "External id": 1633,"Record function id": 0, "finished": true, "Ev Idx": 1632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770608299.526, "dur": 9.411,
    "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": 870, "tid": 870,
    "ts": 7429770608314.100, "dur": 4.462,
    "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": 870, "tid": 870,
    "ts": 7429770608316.514, "dur": 0.624,
    "args": {
      "External id": 1636,"Record function id": 0, "finished": true, "Ev Idx": 1635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770608320.806, "dur": 2.297,
    "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": 870, "tid": 870,
    "ts": 7429770608322.188, "dur": 0.326,
    "args": {
      "External id": 1638,"Record function id": 0, "finished": true, "Ev Idx": 1637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770608324.887, "dur": 9.902,
    "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": 870, "tid": 870,
    "ts": 7429770608339.112, "dur": 20.079,
    "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": 870, "tid": 870,
    "ts": 7429770608362.406, "dur": 7.744,
    "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": 870, "tid": 870,
    "ts": 7429770608371.863, "dur": 8.546,
    "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": 870, "tid": 870,
    "ts": 7429770608382.574, "dur": 8.686,
    "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": 870, "tid": 870,
    "ts": 7429770608393.666, "dur": 2.055,
    "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": 870, "tid": 870,
    "ts": 7429770608394.665, "dur": 0.415,
    "args": {
      "External id": 1645,"Record function id": 0, "finished": true, "Ev Idx": 1644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770608397.392, "dur": 1.633,
    "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": 870, "tid": 870,
    "ts": 7429770608398.169, "dur": 0.302,
    "args": {
      "External id": 1647,"Record function id": 0, "finished": true, "Ev Idx": 1646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770608399.942, "dur": 8.520,
    "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": 870, "tid": 870,
    "ts": 7429770608410.255, "dur": 9.441,
    "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": 870, "tid": 870,
    "ts": 7429770608421.753, "dur": 8.576,
    "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": 870, "tid": 870,
    "ts": 7429770608432.181, "dur": 6.368,
    "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": 870, "tid": 870,
    "ts": 7429770608463.276, "dur": 2.506,
    "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": 870, "tid": 870,
    "ts": 7429770608464.512, "dur": 0.422,
    "args": {
      "External id": 1653,"Record function id": 0, "finished": true, "Ev Idx": 1652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770608466.475, "dur": 2.654,
    "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": 870, "tid": 870,
    "ts": 7429770608468.295, "dur": 0.323,
    "args": {
      "External id": 1655,"Record function id": 0, "finished": true, "Ev Idx": 1654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770608469.804, "dur": 2.478,
    "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": 870, "tid": 870,
    "ts": 7429770608471.074, "dur": 0.646,
    "args": {
      "External id": 1657,"Record function id": 0, "finished": true, "Ev Idx": 1656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770608472.929, "dur": 2.856,
    "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": 870, "tid": 870,
    "ts": 7429770608474.725, "dur": 0.533,
    "args": {
      "External id": 1659,"Record function id": 0, "finished": true, "Ev Idx": 1658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770608476.359, "dur": 1.487,
    "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": 870, "tid": 870,
    "ts": 7429770608477.049, "dur": 0.314,
    "args": {
      "External id": 1661,"Record function id": 0, "finished": true, "Ev Idx": 1660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770608480.403, "dur": 3.084,
    "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": 870, "tid": 870,
    "ts": 7429770608482.439, "dur": 0.396,
    "args": {
      "External id": 1663,"Record function id": 0, "finished": true, "Ev Idx": 1662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770608485.973, "dur": 19.953,
    "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": 870, "tid": 870,
    "ts": 7429770608488.579, "dur": 14.337,
    "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": 870, "tid": 870,
    "ts": 7429770608490.271, "dur": 2.580,
    "args": {
      "External id": 1666,"Record function id": 0, "finished": true, "Ev Idx": 1665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770608491.123, "dur": 1.482,
    "args": {
      "External id": 1667,"Record function id": 0, "finished": true, "Ev Idx": 1666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770608493.596, "dur": 8.743,
    "args": {
      "External id": 1668,"Record function id": 0, "finished": true, "Ev Idx": 1667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770608503.531, "dur": 2.147,
    "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": 870, "tid": 870,
    "ts": 7429770608509.386, "dur": 2.738,
    "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": 870, "tid": 870,
    "ts": 7429770608511.121, "dur": 0.408,
    "args": {
      "External id": 1671,"Record function id": 0, "finished": true, "Ev Idx": 1670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770608512.885, "dur": 1.527,
    "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": 870, "tid": 870,
    "ts": 7429770608513.665, "dur": 0.273,
    "args": {
      "External id": 1673,"Record function id": 0, "finished": true, "Ev Idx": 1672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770608514.932, "dur": 1.584,
    "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": 870, "tid": 870,
    "ts": 7429770608515.621, "dur": 0.502,
    "args": {
      "External id": 1675,"Record function id": 0, "finished": true, "Ev Idx": 1674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770608517.063, "dur": 1.598,
    "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": 870, "tid": 870,
    "ts": 7429770608517.753, "dur": 0.438,
    "args": {
      "External id": 1677,"Record function id": 0, "finished": true, "Ev Idx": 1676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770608519.248, "dur": 1.582,
    "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": 870, "tid": 870,
    "ts": 7429770608519.910, "dur": 0.461,
    "args": {
      "External id": 1679,"Record function id": 0, "finished": true, "Ev Idx": 1678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770608522.348, "dur": 2.600,
    "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": 870, "tid": 870,
    "ts": 7429770608524.068, "dur": 0.456,
    "args": {
      "External id": 1681,"Record function id": 0, "finished": true, "Ev Idx": 1680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770608526.361, "dur": 11.747,
    "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": 870, "tid": 870,
    "ts": 7429770608527.315, "dur": 9.361,
    "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": 870, "tid": 870,
    "ts": 7429770608528.299, "dur": 1.713,
    "args": {
      "External id": 1684,"Record function id": 0, "finished": true, "Ev Idx": 1683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770608528.630, "dur": 1.271,
    "args": {
      "External id": 1685,"Record function id": 0, "finished": true, "Ev Idx": 1684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770608530.311, "dur": 6.062,
    "args": {
      "External id": 1686,"Record function id": 0, "finished": true, "Ev Idx": 1685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770608537.040, "dur": 0.889,
    "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": 870, "tid": 870,
    "ts": 7429770608540.015, "dur": 1.958,
    "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": 870, "tid": 870,
    "ts": 7429770608541.086, "dur": 0.428,
    "args": {
      "External id": 1689,"Record function id": 0, "finished": true, "Ev Idx": 1688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770608543.499, "dur": 46.300,
    "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": 870, "tid": 870,
    "ts": 7429770608544.134, "dur": 1.564,
    "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": 870, "tid": 870,
    "ts": 7429770608544.909, "dur": 0.330,
    "args": {
      "External id": 1692,"Record function id": 0, "finished": true, "Ev Idx": 1691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770608546.055, "dur": 4.668,
    "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": 870, "tid": 870,
    "ts": 7429770608547.853, "dur": 2.723,
    "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": 870, "tid": 870,
    "ts": 7429770608551.169, "dur": 2.088,
    "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": 870, "tid": 870,
    "ts": 7429770608552.600, "dur": 0.301,
    "args": {
      "External id": 1696,"Record function id": 0, "finished": true, "Ev Idx": 1695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770608553.552, "dur": 1.908,
    "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": 870, "tid": 870,
    "ts": 7429770608554.568, "dur": 0.803,
    "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": 870, "tid": 870,
    "ts": 7429770608556.036, "dur": 31.903,
    "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": 870, "tid": 870,
    "ts": 7429770608569.135, "dur": 1.143,
    "args": {
      "External id": 1700,"Record function id": 0, "finished": true, "Ev Idx": 1699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770608588.521, "dur": 0.981,
    "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": 870, "tid": 870,
    "ts": 7429770608592.893, "dur": 9.180,
    "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": 870, "tid": 870,
    "ts": 7429770608605.231, "dur": 1.758,
    "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": 870, "tid": 870,
    "ts": 7429770608605.983, "dur": 0.340,
    "args": {
      "External id": 1704,"Record function id": 0, "finished": true, "Ev Idx": 1703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770608607.602, "dur": 1.264,
    "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": 870, "tid": 870,
    "ts": 7429770608608.152, "dur": 0.322,
    "args": {
      "External id": 1706,"Record function id": 0, "finished": true, "Ev Idx": 1705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770608609.531, "dur": 1.326,
    "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": 870, "tid": 870,
    "ts": 7429770608610.290, "dur": 0.224,
    "args": {
      "External id": 1708,"Record function id": 0, "finished": true, "Ev Idx": 1707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770608612.603, "dur": 7.121,
    "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": 870, "tid": 870,
    "ts": 7429770608624.934, "dur": 13.034,
    "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": 870, "tid": 870,
    "ts": 7429770608626.141, "dur": 0.382,
    "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": 870, "tid": 870,
    "ts": 7429770608627.173, "dur": 10.575,
    "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": 870, "tid": 870,
    "ts": 7429770608640.538, "dur": 0.138,
    "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": 870, "tid": 870,
    "ts": 7429770608651.690, "dur": 0.506,
    "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": 870, "tid": 870,
    "ts": 7429770608653.986, "dur": 41.308,
    "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": 870, "tid": 870,
    "ts": 7429770608654.506, "dur": 2.210,
    "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": 870, "tid": 870,
    "ts": 7429770608655.762, "dur": 0.418,
    "args": {
      "External id": 1717,"Record function id": 0, "finished": true, "Ev Idx": 1716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770608657.107, "dur": 3.076,
    "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": 870, "tid": 870,
    "ts": 7429770608657.624, "dur": 2.435,
    "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": 870, "tid": 870,
    "ts": 7429770608660.627, "dur": 1.653,
    "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": 870, "tid": 870,
    "ts": 7429770608661.632, "dur": 0.272,
    "args": {
      "External id": 1721,"Record function id": 0, "finished": true, "Ev Idx": 1720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770608662.570, "dur": 1.217,
    "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": 870, "tid": 870,
    "ts": 7429770608662.821, "dur": 0.883,
    "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": 870, "tid": 870,
    "ts": 7429770608664.147, "dur": 29.291,
    "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": 870, "tid": 870,
    "ts": 7429770608675.537, "dur": 1.264,
    "args": {
      "External id": 1725,"Record function id": 0, "finished": true, "Ev Idx": 1724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770608694.000, "dur": 1.012,
    "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": 870, "tid": 870,
    "ts": 7429770608697.364, "dur": 1.983,
    "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": 870, "tid": 870,
    "ts": 7429770608698.350, "dur": 0.435,
    "args": {
      "External id": 1728,"Record function id": 0, "finished": true, "Ev Idx": 1727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770608701.427, "dur": 12.671,
    "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": 870, "tid": 870,
    "ts": 7429770608702.088, "dur": 11.860,
    "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": 870, "tid": 870,
    "ts": 7429770608703.094, "dur": 2.280,
    "args": {
      "External id": 1731,"Record function id": 0, "finished": true, "Ev Idx": 1730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770608703.661, "dur": 1.560,
    "args": {
      "External id": 1732,"Record function id": 0, "finished": true, "Ev Idx": 1731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770608705.720, "dur": 7.886,
    "args": {
      "External id": 1733,"Record function id": 0, "finished": true, "Ev Idx": 1732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770608719.639, "dur": 1.757,
    "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": 870, "tid": 870,
    "ts": 7429770608720.103, "dur": 1.192,
    "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": 870, "tid": 870,
    "ts": 7429770608735.768, "dur": 39.831,
    "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": 870, "tid": 870,
    "ts": 7429770608736.401, "dur": 3.113,
    "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": 870, "tid": 870,
    "ts": 7429770608737.787, "dur": 1.106,
    "args": {
      "External id": 1738,"Record function id": 0, "finished": true, "Ev Idx": 1737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770608738.201, "dur": 0.577,
    "args": {
      "External id": 1739,"Record function id": 0, "finished": true, "Ev Idx": 1738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770608739.934, "dur": 35.504,
    "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": 870, "tid": 870,
    "ts": 7429770608740.409, "dur": 2.943,
    "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": 870, "tid": 870,
    "ts": 7429770608740.797, "dur": 2.445,
    "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": 870, "tid": 870,
    "ts": 7429770608743.690, "dur": 29.691,
    "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": 870, "tid": 870,
    "ts": 7429770608756.402, "dur": 1.150,
    "args": {
      "External id": 1744,"Record function id": 0, "finished": true, "Ev Idx": 1743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770608773.990, "dur": 1.108,
    "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": 870, "tid": 870,
    "ts": 7429770608780.837, "dur": 8.760,
    "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": 870, "tid": 870,
    "ts": 7429770608799.784, "dur": 0.338,
    "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": 870, "tid": 870,
    "ts": 7429770608801.972, "dur": 10.563,
    "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": 870, "tid": 870,
    "ts": 7429770608803.813, "dur": 0.415,
    "args": {
      "External id": 1749,"Record function id": 0, "finished": true, "Ev Idx": 1748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770608804.583, "dur": 0.407,
    "args": {
      "External id": 1750,"Record function id": 0, "finished": true, "Ev Idx": 1749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770608815.203, "dur": 12.560,
    "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": 870, "tid": 870,
    "ts": 7429770608831.157, "dur": 8.340,
    "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": 870, "tid": 870,
    "ts": 7429770608841.401, "dur": 11.059,
    "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": 870, "tid": 870,
    "ts": 7429770608854.174, "dur": 8.452,
    "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": 870, "tid": 870,
    "ts": 7429770608865.054, "dur": 0.098,
    "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": 870, "tid": 870,
    "ts": 7429770608866.066, "dur": 8.388,
    "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": 870, "tid": 870,
    "ts": 7429770608889.771, "dur": 48.149,
    "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": 870, "tid": 870,
    "ts": 7429770608890.185, "dur": 2.678,
    "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": 870, "tid": 870,
    "ts": 7429770608891.070, "dur": 1.166,
    "args": {
      "External id": 1759,"Record function id": 0, "finished": true, "Ev Idx": 1758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770608891.667, "dur": 0.428,
    "args": {
      "External id": 1760,"Record function id": 0, "finished": true, "Ev Idx": 1759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770608893.237, "dur": 44.545,
    "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": 870, "tid": 870,
    "ts": 7429770608893.680, "dur": 2.439,
    "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": 870, "tid": 870,
    "ts": 7429770608894.135, "dur": 1.865,
    "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": 870, "tid": 870,
    "ts": 7429770608896.452, "dur": 39.397,
    "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": 870, "tid": 870,
    "ts": 7429770608906.734, "dur": 14.565,
    "args": {
      "External id": 1765,"Record function id": 0, "finished": true, "Ev Idx": 1764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770608936.424, "dur": 1.076,
    "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": 870, "tid": 870,
    "ts": 7429770608946.109, "dur": 9.184,
    "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": 870, "tid": 870,
    "ts": 7429770608963.549, "dur": 33.292,
    "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": 870, "tid": 870,
    "ts": 7429770608963.904, "dur": 2.332,
    "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": 870, "tid": 870,
    "ts": 7429770608964.812, "dur": 0.931,
    "args": {
      "External id": 1770,"Record function id": 0, "finished": true, "Ev Idx": 1769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770608965.214, "dur": 0.386,
    "args": {
      "External id": 1771,"Record function id": 0, "finished": true, "Ev Idx": 1770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770608966.594, "dur": 30.129,
    "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": 870, "tid": 870,
    "ts": 7429770608967.160, "dur": 1.389,
    "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": 870, "tid": 870,
    "ts": 7429770608967.419, "dur": 1.017,
    "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": 870, "tid": 870,
    "ts": 7429770608968.879, "dur": 26.133,
    "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": 870, "tid": 870,
    "ts": 7429770608979.199, "dur": 1.397,
    "args": {
      "External id": 1776,"Record function id": 0, "finished": true, "Ev Idx": 1775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770608995.646, "dur": 0.828,
    "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": 870, "tid": 870,
    "ts": 7429770608998.471, "dur": 8.688,
    "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": 870, "tid": 870,
    "ts": 7429770609021.244, "dur": 40.887,
    "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": 870, "tid": 870,
    "ts": 7429770609021.750, "dur": 2.430,
    "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": 870, "tid": 870,
    "ts": 7429770609022.732, "dur": 0.864,
    "args": {
      "External id": 1781,"Record function id": 0, "finished": true, "Ev Idx": 1780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770609023.148, "dur": 0.314,
    "args": {
      "External id": 1782,"Record function id": 0, "finished": true, "Ev Idx": 1781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770609024.474, "dur": 37.537,
    "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": 870, "tid": 870,
    "ts": 7429770609024.886, "dur": 1.520,
    "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": 870, "tid": 870,
    "ts": 7429770609025.149, "dur": 1.144,
    "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": 870, "tid": 870,
    "ts": 7429770609026.705, "dur": 33.579,
    "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": 870, "tid": 870,
    "ts": 7429770609033.537, "dur": 1.108,
    "args": {
      "External id": 1787,"Record function id": 0, "finished": true, "Ev Idx": 1786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770609060.840, "dur": 0.895,
    "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": 870, "tid": 870,
    "ts": 7429770609064.799, "dur": 9.517,
    "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": 870, "tid": 870,
    "ts": 7429770609093.450, "dur": 0.290,
    "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": 870, "tid": 870,
    "ts": 7429770609095.381, "dur": 9.095,
    "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": 870, "tid": 870,
    "ts": 7429770609096.758, "dur": 0.218,
    "args": {
      "External id": 1792,"Record function id": 0, "finished": true, "Ev Idx": 1791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770609097.251, "dur": 0.098,
    "args": {
      "External id": 1793,"Record function id": 0, "finished": true, "Ev Idx": 1792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770609106.781, "dur": 9.756,
    "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": 870, "tid": 870,
    "ts": 7429770609119.765, "dur": 7.589,
    "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": 870, "tid": 870,
    "ts": 7429770609128.803, "dur": 10.944,
    "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": 870, "tid": 870,
    "ts": 7429770609141.078, "dur": 7.505,
    "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": 870, "tid": 870,
    "ts": 7429770609150.895, "dur": 0.099,
    "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": 870, "tid": 870,
    "ts": 7429770609151.803, "dur": 8.990,
    "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": 870, "tid": 870,
    "ts": 7429770609178.203, "dur": 37.858,
    "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": 870, "tid": 870,
    "ts": 7429770609178.802, "dur": 3.505,
    "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": 870, "tid": 870,
    "ts": 7429770609180.153, "dur": 1.251,
    "args": {
      "External id": 1802,"Record function id": 0, "finished": true, "Ev Idx": 1801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770609180.778, "dur": 0.474,
    "args": {
      "External id": 1803,"Record function id": 0, "finished": true, "Ev Idx": 1802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770609182.982, "dur": 32.957,
    "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": 870, "tid": 870,
    "ts": 7429770609183.637, "dur": 1.991,
    "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": 870, "tid": 870,
    "ts": 7429770609183.929, "dur": 1.583,
    "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": 870, "tid": 870,
    "ts": 7429770609186.090, "dur": 27.807,
    "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": 870, "tid": 870,
    "ts": 7429770609197.084, "dur": 1.203,
    "args": {
      "External id": 1808,"Record function id": 0, "finished": true, "Ev Idx": 1807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770609214.542, "dur": 1.105,
    "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": 870, "tid": 870,
    "ts": 7429770609219.012, "dur": 1.385,
    "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": 870, "tid": 870,
    "ts": 7429770609222.139, "dur": 2.043,
    "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": 870, "tid": 870,
    "ts": 7429770609223.083, "dur": 0.442,
    "args": {
      "External id": 1812,"Record function id": 0, "finished": true, "Ev Idx": 1811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770609231.508, "dur": 31.750,
    "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": 870, "tid": 870,
    "ts": 7429770609231.967, "dur": 2.022,
    "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": 870, "tid": 870,
    "ts": 7429770609232.737, "dur": 0.834,
    "args": {
      "External id": 1815,"Record function id": 0, "finished": true, "Ev Idx": 1814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770609233.157, "dur": 0.286,
    "args": {
      "External id": 1816,"Record function id": 0, "finished": true, "Ev Idx": 1815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770609234.298, "dur": 28.882,
    "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": 870, "tid": 870,
    "ts": 7429770609234.686, "dur": 1.165,
    "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": 870, "tid": 870,
    "ts": 7429770609234.947, "dur": 0.798,
    "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": 870, "tid": 870,
    "ts": 7429770609236.164, "dur": 25.252,
    "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": 870, "tid": 870,
    "ts": 7429770609245.919, "dur": 1.265,
    "args": {
      "External id": 1821,"Record function id": 0, "finished": true, "Ev Idx": 1820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770609261.948, "dur": 0.991,
    "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": 870, "tid": 870,
    "ts": 7429770609265.302, "dur": 1.064,
    "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": 870, "tid": 870,
    "ts": 7429770609267.631, "dur": 1.663,
    "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": 870, "tid": 870,
    "ts": 7429770609268.326, "dur": 0.506,
    "args": {
      "External id": 1825,"Record function id": 0, "finished": true, "Ev Idx": 1824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770609276.976, "dur": 30.428,
    "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": 870, "tid": 870,
    "ts": 7429770609277.258, "dur": 1.856,
    "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": 870, "tid": 870,
    "ts": 7429770609277.977, "dur": 0.769,
    "args": {
      "External id": 1828,"Record function id": 0, "finished": true, "Ev Idx": 1827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770609278.315, "dur": 0.283,
    "args": {
      "External id": 1829,"Record function id": 0, "finished": true, "Ev Idx": 1828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770609279.542, "dur": 27.764,
    "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": 870, "tid": 870,
    "ts": 7429770609279.908, "dur": 1.366,
    "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": 870, "tid": 870,
    "ts": 7429770609280.302, "dur": 0.849,
    "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": 870, "tid": 870,
    "ts": 7429770609281.695, "dur": 23.957,
    "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": 870, "tid": 870,
    "ts": 7429770609291.011, "dur": 1.147,
    "args": {
      "External id": 1834,"Record function id": 0, "finished": true, "Ev Idx": 1833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770609306.195, "dur": 0.884,
    "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": 870, "tid": 870,
    "ts": 7429770609309.590, "dur": 1.232,
    "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": 870, "tid": 870,
    "ts": 7429770609312.173, "dur": 1.775,
    "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": 870, "tid": 870,
    "ts": 7429770609312.974, "dur": 0.475,
    "args": {
      "External id": 1838,"Record function id": 0, "finished": true, "Ev Idx": 1837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770609316.556, "dur": 1.864,
    "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": 870, "tid": 870,
    "ts": 7429770609317.593, "dur": 0.312,
    "args": {
      "External id": 1840,"Record function id": 0, "finished": true, "Ev Idx": 1839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770609319.506, "dur": 0.979,
    "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": 870, "tid": 870,
    "ts": 7429770609319.901, "dur": 0.276,
    "args": {
      "External id": 1842,"Record function id": 0, "finished": true, "Ev Idx": 1841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770609321.494, "dur": 9.791,
    "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": 870, "tid": 870,
    "ts": 7429770609335.901, "dur": 4.671,
    "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": 870, "tid": 870,
    "ts": 7429770609337.710, "dur": 1.065,
    "args": {
      "External id": 1845,"Record function id": 0, "finished": true, "Ev Idx": 1844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770609343.012, "dur": 2.340,
    "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": 870, "tid": 870,
    "ts": 7429770609344.410, "dur": 0.327,
    "args": {
      "External id": 1847,"Record function id": 0, "finished": true, "Ev Idx": 1846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770609347.111, "dur": 8.473,
    "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": 870, "tid": 870,
    "ts": 7429770609359.420, "dur": 16.512,
    "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": 870, "tid": 870,
    "ts": 7429770609378.776, "dur": 8.210,
    "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": 870, "tid": 870,
    "ts": 7429770609388.619, "dur": 8.994,
    "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": 870, "tid": 870,
    "ts": 7429770609399.628, "dur": 7.906,
    "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": 870, "tid": 870,
    "ts": 7429770609410.010, "dur": 2.164,
    "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": 870, "tid": 870,
    "ts": 7429770609410.968, "dur": 0.525,
    "args": {
      "External id": 1854,"Record function id": 0, "finished": true, "Ev Idx": 1853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770609413.784, "dur": 2.618,
    "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": 870, "tid": 870,
    "ts": 7429770609415.386, "dur": 0.446,
    "args": {
      "External id": 1856,"Record function id": 0, "finished": true, "Ev Idx": 1855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770609417.277, "dur": 8.321,
    "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": 870, "tid": 870,
    "ts": 7429770609427.402, "dur": 11.047,
    "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": 870, "tid": 870,
    "ts": 7429770609440.649, "dur": 8.129,
    "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": 870, "tid": 870,
    "ts": 7429770609450.110, "dur": 7.009,
    "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": 870, "tid": 870,
    "ts": 7429770609479.225, "dur": 3.791,
    "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": 870, "tid": 870,
    "ts": 7429770609481.714, "dur": 0.564,
    "args": {
      "External id": 1862,"Record function id": 0, "finished": true, "Ev Idx": 1861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770609483.736, "dur": 2.037,
    "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": 870, "tid": 870,
    "ts": 7429770609484.880, "dur": 0.329,
    "args": {
      "External id": 1864,"Record function id": 0, "finished": true, "Ev Idx": 1863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770609486.499, "dur": 2.548,
    "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": 870, "tid": 870,
    "ts": 7429770609487.770, "dur": 0.625,
    "args": {
      "External id": 1866,"Record function id": 0, "finished": true, "Ev Idx": 1865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770609489.686, "dur": 1.837,
    "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": 870, "tid": 870,
    "ts": 7429770609490.465, "dur": 0.563,
    "args": {
      "External id": 1868,"Record function id": 0, "finished": true, "Ev Idx": 1867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770609492.164, "dur": 1.499,
    "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": 870, "tid": 870,
    "ts": 7429770609492.896, "dur": 0.310,
    "args": {
      "External id": 1870,"Record function id": 0, "finished": true, "Ev Idx": 1869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770609496.205, "dur": 2.617,
    "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": 870, "tid": 870,
    "ts": 7429770609497.765, "dur": 0.412,
    "args": {
      "External id": 1872,"Record function id": 0, "finished": true, "Ev Idx": 1871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770609501.060, "dur": 20.400,
    "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": 870, "tid": 870,
    "ts": 7429770609503.115, "dur": 15.789,
    "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": 870, "tid": 870,
    "ts": 7429770609504.585, "dur": 2.484,
    "args": {
      "External id": 1875,"Record function id": 0, "finished": true, "Ev Idx": 1874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770609505.278, "dur": 1.606,
    "args": {
      "External id": 1876,"Record function id": 0, "finished": true, "Ev Idx": 1875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770609507.847, "dur": 10.524,
    "args": {
      "External id": 1877,"Record function id": 0, "finished": true, "Ev Idx": 1876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770609519.386, "dur": 1.861,
    "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": 870, "tid": 870,
    "ts": 7429770609524.585, "dur": 1.890,
    "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": 870, "tid": 870,
    "ts": 7429770609525.474, "dur": 0.403,
    "args": {
      "External id": 1880,"Record function id": 0, "finished": true, "Ev Idx": 1879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770609527.092, "dur": 1.588,
    "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": 870, "tid": 870,
    "ts": 7429770609527.929, "dur": 0.273,
    "args": {
      "External id": 1882,"Record function id": 0, "finished": true, "Ev Idx": 1881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770609529.196, "dur": 1.707,
    "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": 870, "tid": 870,
    "ts": 7429770609530.181, "dur": 0.317,
    "args": {
      "External id": 1884,"Record function id": 0, "finished": true, "Ev Idx": 1883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770609531.553, "dur": 1.638,
    "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": 870, "tid": 870,
    "ts": 7429770609532.410, "dur": 0.301,
    "args": {
      "External id": 1886,"Record function id": 0, "finished": true, "Ev Idx": 1885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770609533.792, "dur": 2.878,
    "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": 870, "tid": 870,
    "ts": 7429770609535.796, "dur": 0.418,
    "args": {
      "External id": 1888,"Record function id": 0, "finished": true, "Ev Idx": 1887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770609538.354, "dur": 2.149,
    "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": 870, "tid": 870,
    "ts": 7429770609539.711, "dur": 0.360,
    "args": {
      "External id": 1890,"Record function id": 0, "finished": true, "Ev Idx": 1889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770609541.966, "dur": 11.716,
    "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": 870, "tid": 870,
    "ts": 7429770609542.867, "dur": 9.313,
    "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": 870, "tid": 870,
    "ts": 7429770609543.624, "dur": 1.946,
    "args": {
      "External id": 1893,"Record function id": 0, "finished": true, "Ev Idx": 1892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770609544.145, "dur": 1.323,
    "args": {
      "External id": 1894,"Record function id": 0, "finished": true, "Ev Idx": 1893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770609545.884, "dur": 6.004,
    "args": {
      "External id": 1895,"Record function id": 0, "finished": true, "Ev Idx": 1894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770609552.582, "dur": 0.944,
    "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": 870, "tid": 870,
    "ts": 7429770609555.606, "dur": 1.723,
    "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": 870, "tid": 870,
    "ts": 7429770609556.463, "dur": 0.371,
    "args": {
      "External id": 1898,"Record function id": 0, "finished": true, "Ev Idx": 1897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770609558.788, "dur": 45.029,
    "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": 870, "tid": 870,
    "ts": 7429770609559.406, "dur": 2.422,
    "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": 870, "tid": 870,
    "ts": 7429770609561.066, "dur": 0.322,
    "args": {
      "External id": 1901,"Record function id": 0, "finished": true, "Ev Idx": 1900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770609562.205, "dur": 3.772,
    "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": 870, "tid": 870,
    "ts": 7429770609563.739, "dur": 2.120,
    "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": 870, "tid": 870,
    "ts": 7429770609566.396, "dur": 1.321,
    "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": 870, "tid": 870,
    "ts": 7429770609567.066, "dur": 0.291,
    "args": {
      "External id": 1905,"Record function id": 0, "finished": true, "Ev Idx": 1904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770609567.985, "dur": 1.748,
    "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": 870, "tid": 870,
    "ts": 7429770609568.810, "dur": 0.838,
    "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": 870, "tid": 870,
    "ts": 7429770609570.359, "dur": 31.539,
    "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": 870, "tid": 870,
    "ts": 7429770609583.722, "dur": 1.118,
    "args": {
      "External id": 1909,"Record function id": 0, "finished": true, "Ev Idx": 1908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770609602.473, "dur": 1.007,
    "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": 870, "tid": 870,
    "ts": 7429770609606.666, "dur": 9.148,
    "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": 870, "tid": 870,
    "ts": 7429770609618.679, "dur": 1.988,
    "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": 870, "tid": 870,
    "ts": 7429770609619.597, "dur": 0.422,
    "args": {
      "External id": 1913,"Record function id": 0, "finished": true, "Ev Idx": 1912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770609621.424, "dur": 1.265,
    "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": 870, "tid": 870,
    "ts": 7429770609621.967, "dur": 0.318,
    "args": {
      "External id": 1915,"Record function id": 0, "finished": true, "Ev Idx": 1914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770609623.339, "dur": 1.070,
    "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": 870, "tid": 870,
    "ts": 7429770609623.808, "dur": 0.237,
    "args": {
      "External id": 1917,"Record function id": 0, "finished": true, "Ev Idx": 1916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770609626.043, "dur": 7.696,
    "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": 870, "tid": 870,
    "ts": 7429770609638.198, "dur": 12.114,
    "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": 870, "tid": 870,
    "ts": 7429770609639.188, "dur": 0.415,
    "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": 870, "tid": 870,
    "ts": 7429770609640.203, "dur": 9.863,
    "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": 870, "tid": 870,
    "ts": 7429770609652.796, "dur": 0.180,
    "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": 870, "tid": 870,
    "ts": 7429770609662.861, "dur": 0.417,
    "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": 870, "tid": 870,
    "ts": 7429770609664.886, "dur": 39.122,
    "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": 870, "tid": 870,
    "ts": 7429770609665.353, "dur": 1.947,
    "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": 870, "tid": 870,
    "ts": 7429770609666.441, "dur": 0.323,
    "args": {
      "External id": 1926,"Record function id": 0, "finished": true, "Ev Idx": 1925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770609667.680, "dur": 3.110,
    "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": 870, "tid": 870,
    "ts": 7429770609667.987, "dur": 2.680,
    "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": 870, "tid": 870,
    "ts": 7429770609671.203, "dur": 2.074,
    "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": 870, "tid": 870,
    "ts": 7429770609672.632, "dur": 0.248,
    "args": {
      "External id": 1930,"Record function id": 0, "finished": true, "Ev Idx": 1929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770609673.559, "dur": 1.358,
    "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": 870, "tid": 870,
    "ts": 7429770609673.830, "dur": 1.016,
    "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": 870, "tid": 870,
    "ts": 7429770609675.272, "dur": 26.845,
    "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": 870, "tid": 870,
    "ts": 7429770609685.941, "dur": 1.258,
    "args": {
      "External id": 1934,"Record function id": 0, "finished": true, "Ev Idx": 1933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770609702.686, "dur": 1.029,
    "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": 870, "tid": 870,
    "ts": 7429770609705.818, "dur": 1.941,
    "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": 870, "tid": 870,
    "ts": 7429770609706.834, "dur": 0.366,
    "args": {
      "External id": 1937,"Record function id": 0, "finished": true, "Ev Idx": 1936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770609709.825, "dur": 12.573,
    "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": 870, "tid": 870,
    "ts": 7429770609710.509, "dur": 11.742,
    "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": 870, "tid": 870,
    "ts": 7429770609711.359, "dur": 2.659,
    "args": {
      "External id": 1940,"Record function id": 0, "finished": true, "Ev Idx": 1939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770609711.932, "dur": 1.943,
    "args": {
      "External id": 1941,"Record function id": 0, "finished": true, "Ev Idx": 1940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770609714.389, "dur": 7.497,
    "args": {
      "External id": 1942,"Record function id": 0, "finished": true, "Ev Idx": 1941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770609727.257, "dur": 1.644,
    "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": 870, "tid": 870,
    "ts": 7429770609727.580, "dur": 1.216,
    "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": 870, "tid": 870,
    "ts": 7429770609740.933, "dur": 36.708,
    "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": 870, "tid": 870,
    "ts": 7429770609741.452, "dur": 3.615,
    "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": 870, "tid": 870,
    "ts": 7429770609743.649, "dur": 0.931,
    "args": {
      "External id": 1947,"Record function id": 0, "finished": true, "Ev Idx": 1946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770609744.106, "dur": 0.298,
    "args": {
      "External id": 1948,"Record function id": 0, "finished": true, "Ev Idx": 1947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770609745.448, "dur": 32.053,
    "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": 870, "tid": 870,
    "ts": 7429770609745.857, "dur": 1.705,
    "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": 870, "tid": 870,
    "ts": 7429770609746.097, "dur": 1.375,
    "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": 870, "tid": 870,
    "ts": 7429770609748.110, "dur": 27.342,
    "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": 870, "tid": 870,
    "ts": 7429770609758.150, "dur": 1.156,
    "args": {
      "External id": 1953,"Record function id": 0, "finished": true, "Ev Idx": 1952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770609776.208, "dur": 0.970,
    "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": 870, "tid": 870,
    "ts": 7429770609782.098, "dur": 7.785,
    "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": 870, "tid": 870,
    "ts": 7429770609798.728, "dur": 0.180,
    "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": 870, "tid": 870,
    "ts": 7429770609800.742, "dur": 9.787,
    "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": 870, "tid": 870,
    "ts": 7429770609802.276, "dur": 0.536,
    "args": {
      "External id": 1958,"Record function id": 0, "finished": true, "Ev Idx": 1957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770609803.112, "dur": 0.076,
    "args": {
      "External id": 1959,"Record function id": 0, "finished": true, "Ev Idx": 1958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770609813.366, "dur": 11.421,
    "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": 870, "tid": 870,
    "ts": 7429770609828.100, "dur": 8.178,
    "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": 870, "tid": 870,
    "ts": 7429770609838.018, "dur": 9.572,
    "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": 870, "tid": 870,
    "ts": 7429770609849.269, "dur": 9.081,
    "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": 870, "tid": 870,
    "ts": 7429770609860.667, "dur": 0.469,
    "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": 870, "tid": 870,
    "ts": 7429770609862.173, "dur": 8.654,
    "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": 870, "tid": 870,
    "ts": 7429770609885.118, "dur": 35.674,
    "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": 870, "tid": 870,
    "ts": 7429770609885.558, "dur": 3.379,
    "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": 870, "tid": 870,
    "ts": 7429770609886.952, "dur": 1.360,
    "args": {
      "External id": 1968,"Record function id": 0, "finished": true, "Ev Idx": 1967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770609887.722, "dur": 0.434,
    "args": {
      "External id": 1969,"Record function id": 0, "finished": true, "Ev Idx": 1968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770609889.306, "dur": 31.356,
    "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": 870, "tid": 870,
    "ts": 7429770609889.767, "dur": 2.217,
    "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": 870, "tid": 870,
    "ts": 7429770609890.059, "dur": 1.821,
    "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": 870, "tid": 870,
    "ts": 7429770609892.328, "dur": 26.405,
    "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": 870, "tid": 870,
    "ts": 7429770609902.319, "dur": 1.367,
    "args": {
      "External id": 1974,"Record function id": 0, "finished": true, "Ev Idx": 1973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770609919.407, "dur": 0.970,
    "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": 870, "tid": 870,
    "ts": 7429770609928.041, "dur": 9.558,
    "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": 870, "tid": 870,
    "ts": 7429770609944.651, "dur": 33.872,
    "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": 870, "tid": 870,
    "ts": 7429770609944.967, "dur": 2.471,
    "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": 870, "tid": 870,
    "ts": 7429770609945.836, "dur": 1.141,
    "args": {
      "External id": 1979,"Record function id": 0, "finished": true, "Ev Idx": 1978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770609946.498, "dur": 0.327,
    "args": {
      "External id": 1980,"Record function id": 0, "finished": true, "Ev Idx": 1979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770609947.749, "dur": 30.669,
    "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": 870, "tid": 870,
    "ts": 7429770609948.153, "dur": 1.362,
    "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": 870, "tid": 870,
    "ts": 7429770609948.442, "dur": 0.973,
    "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": 870, "tid": 870,
    "ts": 7429770609949.817, "dur": 26.657,
    "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": 870, "tid": 870,
    "ts": 7429770609960.538, "dur": 1.205,
    "args": {
      "External id": 1985,"Record function id": 0, "finished": true, "Ev Idx": 1984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770609977.123, "dur": 1.037,
    "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": 870, "tid": 870,
    "ts": 7429770609980.130, "dur": 7.929,
    "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": 870, "tid": 870,
    "ts": 7429770609993.685, "dur": 160.445,
    "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": 870, "tid": 870,
    "ts": 7429770609994.189, "dur": 2.258,
    "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": 870, "tid": 870,
    "ts": 7429770609995.133, "dur": 0.819,
    "args": {
      "External id": 1990,"Record function id": 0, "finished": true, "Ev Idx": 1989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770609995.527, "dur": 0.299,
    "args": {
      "External id": 1991,"Record function id": 0, "finished": true, "Ev Idx": 1990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770609996.879, "dur": 157.049,
    "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": 870, "tid": 870,
    "ts": 7429770609997.288, "dur": 1.333,
    "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": 870, "tid": 870,
    "ts": 7429770609997.556, "dur": 0.955,
    "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": 870, "tid": 870,
    "ts": 7429770609999.057, "dur": 150.612,
    "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": 870, "tid": 870,
    "ts": 7429770610013.246, "dur": 97.375,
    "args": {
      "External id": 1996,"Record function id": 0, "finished": true, "Ev Idx": 1995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770610151.336, "dur": 2.075,
    "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": 870, "tid": 870,
    "ts": 7429770610160.530, "dur": 11.150,
    "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": 870, "tid": 870,
    "ts": 7429770610196.553, "dur": 0.176,
    "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": 870, "tid": 870,
    "ts": 7429770610198.875, "dur": 13.448,
    "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": 870, "tid": 870,
    "ts": 7429770610200.919, "dur": 0.729,
    "args": {
      "External id": 2001,"Record function id": 0, "finished": true, "Ev Idx": 2000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770610201.992, "dur": 0.126,
    "args": {
      "External id": 2002,"Record function id": 0, "finished": true, "Ev Idx": 2001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770610214.879, "dur": 15.420,
    "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": 870, "tid": 870,
    "ts": 7429770610234.446, "dur": 9.501,
    "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": 870, "tid": 870,
    "ts": 7429770610245.871, "dur": 9.239,
    "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": 870, "tid": 870,
    "ts": 7429770610256.865, "dur": 9.267,
    "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": 870, "tid": 870,
    "ts": 7429770610268.668, "dur": 0.110,
    "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": 870, "tid": 870,
    "ts": 7429770610269.705, "dur": 8.531,
    "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": 870, "tid": 870,
    "ts": 7429770610298.333, "dur": 43.902,
    "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": 870, "tid": 870,
    "ts": 7429770610298.942, "dur": 5.590,
    "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": 870, "tid": 870,
    "ts": 7429770610301.298, "dur": 2.173,
    "args": {
      "External id": 2011,"Record function id": 0, "finished": true, "Ev Idx": 2010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770610302.322, "dur": 0.909,
    "args": {
      "External id": 2012,"Record function id": 0, "finished": true, "Ev Idx": 2011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770610305.013, "dur": 37.117,
    "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": 870, "tid": 870,
    "ts": 7429770610305.773, "dur": 2.107,
    "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": 870, "tid": 870,
    "ts": 7429770610306.477, "dur": 1.254,
    "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": 870, "tid": 870,
    "ts": 7429770610308.338, "dur": 31.799,
    "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": 870, "tid": 870,
    "ts": 7429770610322.026, "dur": 1.788,
    "args": {
      "External id": 2017,"Record function id": 0, "finished": true, "Ev Idx": 2016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770610340.761, "dur": 1.076,
    "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": 870, "tid": 870,
    "ts": 7429770610345.577, "dur": 1.416,
    "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": 870, "tid": 870,
    "ts": 7429770610349.054, "dur": 2.847,
    "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": 870, "tid": 870,
    "ts": 7429770610350.325, "dur": 0.699,
    "args": {
      "External id": 2021,"Record function id": 0, "finished": true, "Ev Idx": 2020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770610553.295, "dur": 46.633,
    "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": 870, "tid": 870,
    "ts": 7429770610554.119, "dur": 3.923,
    "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": 870, "tid": 870,
    "ts": 7429770610555.833, "dur": 1.515,
    "args": {
      "External id": 2024,"Record function id": 0, "finished": true, "Ev Idx": 2023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770610556.686, "dur": 0.459,
    "args": {
      "External id": 2025,"Record function id": 0, "finished": true, "Ev Idx": 2024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770610558.465, "dur": 41.318,
    "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": 870, "tid": 870,
    "ts": 7429770610559.165, "dur": 2.154,
    "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": 870, "tid": 870,
    "ts": 7429770610559.539, "dur": 1.662,
    "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": 870, "tid": 870,
    "ts": 7429770610561.758, "dur": 35.711,
    "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": 870, "tid": 870,
    "ts": 7429770610575.281, "dur": 1.388,
    "args": {
      "External id": 2030,"Record function id": 0, "finished": true, "Ev Idx": 2029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770610598.100, "dur": 1.265,
    "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": 870, "tid": 870,
    "ts": 7429770610604.002, "dur": 1.383,
    "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": 870, "tid": 870,
    "ts": 7429770610607.289, "dur": 2.832,
    "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": 870, "tid": 870,
    "ts": 7429770610608.933, "dur": 0.505,
    "args": {
      "External id": 2034,"Record function id": 0, "finished": true, "Ev Idx": 2033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770610620.864, "dur": 36.160,
    "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": 870, "tid": 870,
    "ts": 7429770610621.400, "dur": 3.000,
    "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": 870, "tid": 870,
    "ts": 7429770610622.756, "dur": 1.226,
    "args": {
      "External id": 2037,"Record function id": 0, "finished": true, "Ev Idx": 2036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770610623.421, "dur": 0.446,
    "args": {
      "External id": 2038,"Record function id": 0, "finished": true, "Ev Idx": 2037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770610624.849, "dur": 32.077,
    "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": 870, "tid": 870,
    "ts": 7429770610625.320, "dur": 2.228,
    "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": 870, "tid": 870,
    "ts": 7429770610625.637, "dur": 1.837,
    "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": 870, "tid": 870,
    "ts": 7429770610627.873, "dur": 27.142,
    "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": 870, "tid": 870,
    "ts": 7429770610639.973, "dur": 1.124,
    "args": {
      "External id": 2043,"Record function id": 0, "finished": true, "Ev Idx": 2042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770610655.563, "dur": 1.108,
    "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": 870, "tid": 870,
    "ts": 7429770610659.884, "dur": 1.087,
    "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": 870, "tid": 870,
    "ts": 7429770610662.420, "dur": 1.814,
    "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": 870, "tid": 870,
    "ts": 7429770610663.417, "dur": 0.346,
    "args": {
      "External id": 2047,"Record function id": 0, "finished": true, "Ev Idx": 2046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770610668.178, "dur": 35.479,
    "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": 870, "tid": 870,
    "ts": 7429770610702.768, "dur": 0.368,
    "args": {
      "External id": 2049,"Record function id": 0, "finished": true, "Ev Idx": 2048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770610704.977, "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": 870, "tid": 870,
    "ts": 7429770610705.433, "dur": 0.320,
    "args": {
      "External id": 2051,"Record function id": 0, "finished": true, "Ev Idx": 2050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770610707.209, "dur": 13.674,
    "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": 870, "tid": 870,
    "ts": 7429770610728.144, "dur": 5.512,
    "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": 870, "tid": 870,
    "ts": 7429770610731.377, "dur": 0.637,
    "args": {
      "External id": 2054,"Record function id": 0, "finished": true, "Ev Idx": 2053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770610736.113, "dur": 1.795,
    "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": 870, "tid": 870,
    "ts": 7429770610736.982, "dur": 0.312,
    "args": {
      "External id": 2056,"Record function id": 0, "finished": true, "Ev Idx": 2055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770610740.091, "dur": 11.967,
    "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": 870, "tid": 870,
    "ts": 7429770610756.729, "dur": 18.604,
    "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": 870, "tid": 870,
    "ts": 7429770610779.011, "dur": 8.424,
    "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": 870, "tid": 870,
    "ts": 7429770610789.458, "dur": 9.465,
    "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": 870, "tid": 870,
    "ts": 7429770610801.079, "dur": 7.362,
    "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": 870, "tid": 870,
    "ts": 7429770610810.918, "dur": 2.492,
    "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": 870, "tid": 870,
    "ts": 7429770610812.334, "dur": 0.434,
    "args": {
      "External id": 2063,"Record function id": 0, "finished": true, "Ev Idx": 2062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770610815.038, "dur": 2.446,
    "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": 870, "tid": 870,
    "ts": 7429770610816.646, "dur": 0.305,
    "args": {
      "External id": 2065,"Record function id": 0, "finished": true, "Ev Idx": 2064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770610818.414, "dur": 8.808,
    "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": 870, "tid": 870,
    "ts": 7429770610828.903, "dur": 12.010,
    "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": 870, "tid": 870,
    "ts": 7429770610843.293, "dur": 7.699,
    "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": 870, "tid": 870,
    "ts": 7429770610852.471, "dur": 7.069,
    "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": 870, "tid": 870,
    "ts": 7429770610886.843, "dur": 2.422,
    "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": 870, "tid": 870,
    "ts": 7429770610888.030, "dur": 0.414,
    "args": {
      "External id": 2071,"Record function id": 0, "finished": true, "Ev Idx": 2070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770610889.914, "dur": 2.281,
    "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": 870, "tid": 870,
    "ts": 7429770610891.392, "dur": 0.295,
    "args": {
      "External id": 2073,"Record function id": 0, "finished": true, "Ev Idx": 2072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770610892.911, "dur": 3.114,
    "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": 870, "tid": 870,
    "ts": 7429770610894.431, "dur": 0.799,
    "args": {
      "External id": 2075,"Record function id": 0, "finished": true, "Ev Idx": 2074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770610896.640, "dur": 1.732,
    "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": 870, "tid": 870,
    "ts": 7429770610897.420, "dur": 0.427,
    "args": {
      "External id": 2077,"Record function id": 0, "finished": true, "Ev Idx": 2076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770610898.942, "dur": 1.461,
    "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": 870, "tid": 870,
    "ts": 7429770610899.617, "dur": 0.306,
    "args": {
      "External id": 2079,"Record function id": 0, "finished": true, "Ev Idx": 2078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770610903.602, "dur": 4.068,
    "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": 870, "tid": 870,
    "ts": 7429770610906.341, "dur": 0.520,
    "args": {
      "External id": 2081,"Record function id": 0, "finished": true, "Ev Idx": 2080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770610910.220, "dur": 23.712,
    "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": 870, "tid": 870,
    "ts": 7429770610913.144, "dur": 18.154,
    "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": 870, "tid": 870,
    "ts": 7429770610915.967, "dur": 2.775,
    "args": {
      "External id": 2084,"Record function id": 0, "finished": true, "Ev Idx": 2083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770610916.920, "dur": 1.581,
    "args": {
      "External id": 2085,"Record function id": 0, "finished": true, "Ev Idx": 2084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770610919.631, "dur": 11.025,
    "args": {
      "External id": 2086,"Record function id": 0, "finished": true, "Ev Idx": 2085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770610931.876, "dur": 1.746,
    "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": 870, "tid": 870,
    "ts": 7429770610937.179, "dur": 2.585,
    "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": 870, "tid": 870,
    "ts": 7429770610938.743, "dur": 0.414,
    "args": {
      "External id": 2089,"Record function id": 0, "finished": true, "Ev Idx": 2088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770610940.388, "dur": 1.424,
    "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": 870, "tid": 870,
    "ts": 7429770610941.070, "dur": 0.271,
    "args": {
      "External id": 2091,"Record function id": 0, "finished": true, "Ev Idx": 2090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770610942.365, "dur": 2.460,
    "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": 870, "tid": 870,
    "ts": 7429770610944.012, "dur": 0.379,
    "args": {
      "External id": 2093,"Record function id": 0, "finished": true, "Ev Idx": 2092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770610945.363, "dur": 2.329,
    "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": 870, "tid": 870,
    "ts": 7429770610946.874, "dur": 0.341,
    "args": {
      "External id": 2095,"Record function id": 0, "finished": true, "Ev Idx": 2094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770610948.248, "dur": 1.370,
    "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": 870, "tid": 870,
    "ts": 7429770610948.885, "dur": 0.293,
    "args": {
      "External id": 2097,"Record function id": 0, "finished": true, "Ev Idx": 2096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770610951.125, "dur": 1.606,
    "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": 870, "tid": 870,
    "ts": 7429770610951.952, "dur": 0.346,
    "args": {
      "External id": 2099,"Record function id": 0, "finished": true, "Ev Idx": 2098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770610954.241, "dur": 12.687,
    "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": 870, "tid": 870,
    "ts": 7429770610955.164, "dur": 8.993,
    "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": 870, "tid": 870,
    "ts": 7429770610955.807, "dur": 1.730,
    "args": {
      "External id": 2102,"Record function id": 0, "finished": true, "Ev Idx": 2101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770610956.172, "dur": 1.229,
    "args": {
      "External id": 2103,"Record function id": 0, "finished": true, "Ev Idx": 2102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770610957.852, "dur": 5.974,
    "args": {
      "External id": 2104,"Record function id": 0, "finished": true, "Ev Idx": 2103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770610964.555, "dur": 2.203,
    "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": 870, "tid": 870,
    "ts": 7429770610969.158, "dur": 1.988,
    "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": 870, "tid": 870,
    "ts": 7429770610970.296, "dur": 0.334,
    "args": {
      "External id": 2107,"Record function id": 0, "finished": true, "Ev Idx": 2106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770610972.952, "dur": 48.870,
    "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": 870, "tid": 870,
    "ts": 7429770610973.771, "dur": 1.611,
    "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": 870, "tid": 870,
    "ts": 7429770610974.618, "dur": 0.324,
    "args": {
      "External id": 2110,"Record function id": 0, "finished": true, "Ev Idx": 2109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770610975.778, "dur": 3.975,
    "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": 870, "tid": 870,
    "ts": 7429770610977.566, "dur": 2.043,
    "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": 870, "tid": 870,
    "ts": 7429770610980.258, "dur": 1.528,
    "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": 870, "tid": 870,
    "ts": 7429770610981.127, "dur": 0.293,
    "args": {
      "External id": 2114,"Record function id": 0, "finished": true, "Ev Idx": 2113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770610982.082, "dur": 1.717,
    "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": 870, "tid": 870,
    "ts": 7429770610982.900, "dur": 0.810,
    "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": 870, "tid": 870,
    "ts": 7429770610984.525, "dur": 35.214,
    "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": 870, "tid": 870,
    "ts": 7429770610999.171, "dur": 1.140,
    "args": {
      "External id": 2118,"Record function id": 0, "finished": true, "Ev Idx": 2117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770611020.420, "dur": 1.025,
    "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": 870, "tid": 870,
    "ts": 7429770611025.685, "dur": 10.410,
    "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": 870, "tid": 870,
    "ts": 7429770611039.379, "dur": 1.812,
    "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": 870, "tid": 870,
    "ts": 7429770611040.146, "dur": 0.381,
    "args": {
      "External id": 2122,"Record function id": 0, "finished": true, "Ev Idx": 2121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770611041.783, "dur": 1.569,
    "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": 870, "tid": 870,
    "ts": 7429770611042.521, "dur": 0.458,
    "args": {
      "External id": 2124,"Record function id": 0, "finished": true, "Ev Idx": 2123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770611043.969, "dur": 1.304,
    "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": 870, "tid": 870,
    "ts": 7429770611044.536, "dur": 0.384,
    "args": {
      "External id": 2126,"Record function id": 0, "finished": true, "Ev Idx": 2125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770611046.876, "dur": 8.119,
    "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": 870, "tid": 870,
    "ts": 7429770611061.371, "dur": 15.504,
    "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": 870, "tid": 870,
    "ts": 7429770611062.640, "dur": 0.418,
    "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": 870, "tid": 870,
    "ts": 7429770611063.782, "dur": 12.858,
    "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": 870, "tid": 870,
    "ts": 7429770611080.152, "dur": 0.337,
    "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": 870, "tid": 870,
    "ts": 7429770611093.033, "dur": 0.469,
    "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": 870, "tid": 870,
    "ts": 7429770611095.249, "dur": 42.074,
    "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": 870, "tid": 870,
    "ts": 7429770611095.887, "dur": 2.196,
    "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": 870, "tid": 870,
    "ts": 7429770611097.178, "dur": 0.398,
    "args": {
      "External id": 2135,"Record function id": 0, "finished": true, "Ev Idx": 2134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770611098.626, "dur": 2.766,
    "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": 870, "tid": 870,
    "ts": 7429770611099.034, "dur": 2.237,
    "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": 870, "tid": 870,
    "ts": 7429770611101.797, "dur": 1.272,
    "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": 870, "tid": 870,
    "ts": 7429770611102.437, "dur": 0.256,
    "args": {
      "External id": 2139,"Record function id": 0, "finished": true, "Ev Idx": 2138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770611103.373, "dur": 1.925,
    "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": 870, "tid": 870,
    "ts": 7429770611103.599, "dur": 1.625,
    "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": 870, "tid": 870,
    "ts": 7429770611105.762, "dur": 29.503,
    "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": 870, "tid": 870,
    "ts": 7429770611118.040, "dur": 1.239,
    "args": {
      "External id": 2143,"Record function id": 0, "finished": true, "Ev Idx": 2142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770611135.935, "dur": 1.101,
    "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": 870, "tid": 870,
    "ts": 7429770611139.370, "dur": 1.886,
    "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": 870, "tid": 870,
    "ts": 7429770611140.355, "dur": 0.360,
    "args": {
      "External id": 2146,"Record function id": 0, "finished": true, "Ev Idx": 2145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770611144.050, "dur": 11.964,
    "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": 870, "tid": 870,
    "ts": 7429770611144.614, "dur": 11.250,
    "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": 870, "tid": 870,
    "ts": 7429770611145.573, "dur": 2.172,
    "args": {
      "External id": 2149,"Record function id": 0, "finished": true, "Ev Idx": 2148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770611145.976, "dur": 1.608,
    "args": {
      "External id": 2150,"Record function id": 0, "finished": true, "Ev Idx": 2149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770611148.124, "dur": 7.327,
    "args": {
      "External id": 2151,"Record function id": 0, "finished": true, "Ev Idx": 2150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770611161.708, "dur": 1.612,
    "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": 870, "tid": 870,
    "ts": 7429770611162.035, "dur": 1.196,
    "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": 870, "tid": 870,
    "ts": 7429770611178.042, "dur": 38.077,
    "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": 870, "tid": 870,
    "ts": 7429770611178.612, "dur": 3.955,
    "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": 870, "tid": 870,
    "ts": 7429770611181.001, "dur": 0.877,
    "args": {
      "External id": 2156,"Record function id": 0, "finished": true, "Ev Idx": 2155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770611181.426, "dur": 0.358,
    "args": {
      "External id": 2157,"Record function id": 0, "finished": true, "Ev Idx": 2156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770611182.942, "dur": 33.013,
    "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": 870, "tid": 870,
    "ts": 7429770611183.365, "dur": 1.778,
    "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": 870, "tid": 870,
    "ts": 7429770611183.630, "dur": 1.428,
    "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": 870, "tid": 870,
    "ts": 7429770611185.669, "dur": 28.174,
    "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": 870, "tid": 870,
    "ts": 7429770611196.158, "dur": 1.172,
    "args": {
      "External id": 2162,"Record function id": 0, "finished": true, "Ev Idx": 2161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770611214.520, "dur": 1.083,
    "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": 870, "tid": 870,
    "ts": 7429770611222.012, "dur": 9.296,
    "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": 870, "tid": 870,
    "ts": 7429770611249.384, "dur": 0.212,
    "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": 870, "tid": 870,
    "ts": 7429770611252.004, "dur": 14.416,
    "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": 870, "tid": 870,
    "ts": 7429770611253.808, "dur": 0.409,
    "args": {
      "External id": 2167,"Record function id": 0, "finished": true, "Ev Idx": 2166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770611254.626, "dur": 0.257,
    "args": {
      "External id": 2168,"Record function id": 0, "finished": true, "Ev Idx": 2167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770611269.758, "dur": 13.893,
    "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": 870, "tid": 870,
    "ts": 7429770611287.474, "dur": 9.651,
    "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": 870, "tid": 870,
    "ts": 7429770611299.533, "dur": 9.656,
    "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": 870, "tid": 870,
    "ts": 7429770611310.920, "dur": 9.019,
    "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": 870, "tid": 870,
    "ts": 7429770611322.476, "dur": 0.137,
    "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": 870, "tid": 870,
    "ts": 7429770611323.539, "dur": 8.062,
    "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": 870, "tid": 870,
    "ts": 7429770611348.860, "dur": 38.880,
    "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": 870, "tid": 870,
    "ts": 7429770611349.302, "dur": 2.919,
    "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": 870, "tid": 870,
    "ts": 7429770611350.493, "dur": 1.128,
    "args": {
      "External id": 2177,"Record function id": 0, "finished": true, "Ev Idx": 2176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770611350.983, "dur": 0.460,
    "args": {
      "External id": 2178,"Record function id": 0, "finished": true, "Ev Idx": 2177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770611352.561, "dur": 35.026,
    "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": 870, "tid": 870,
    "ts": 7429770611353.074, "dur": 1.829,
    "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": 870, "tid": 870,
    "ts": 7429770611353.562, "dur": 1.221,
    "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": 870, "tid": 870,
    "ts": 7429770611355.242, "dur": 30.338,
    "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": 870, "tid": 870,
    "ts": 7429770611368.990, "dur": 1.212,
    "args": {
      "External id": 2183,"Record function id": 0, "finished": true, "Ev Idx": 2182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770611386.183, "dur": 1.103,
    "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": 870, "tid": 870,
    "ts": 7429770611396.609, "dur": 10.882,
    "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": 870, "tid": 870,
    "ts": 7429770611415.631, "dur": 34.525,
    "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": 870, "tid": 870,
    "ts": 7429770611415.994, "dur": 2.437,
    "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": 870, "tid": 870,
    "ts": 7429770611416.771, "dur": 1.176,
    "args": {
      "External id": 2188,"Record function id": 0, "finished": true, "Ev Idx": 2187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770611417.429, "dur": 0.363,
    "args": {
      "External id": 2189,"Record function id": 0, "finished": true, "Ev Idx": 2188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770611418.776, "dur": 31.277,
    "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": 870, "tid": 870,
    "ts": 7429770611419.294, "dur": 2.933,
    "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": 870, "tid": 870,
    "ts": 7429770611419.561, "dur": 2.545,
    "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": 870, "tid": 870,
    "ts": 7429770611422.570, "dur": 25.715,
    "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": 870, "tid": 870,
    "ts": 7429770611432.109, "dur": 1.367,
    "args": {
      "External id": 2194,"Record function id": 0, "finished": true, "Ev Idx": 2193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770611448.881, "dur": 0.925,
    "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": 870, "tid": 870,
    "ts": 7429770611451.894, "dur": 8.641,
    "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": 870, "tid": 870,
    "ts": 7429770611467.279, "dur": 47.705,
    "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": 870, "tid": 870,
    "ts": 7429770611467.611, "dur": 2.424,
    "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": 870, "tid": 870,
    "ts": 7429770611468.766, "dur": 0.805,
    "args": {
      "External id": 2199,"Record function id": 0, "finished": true, "Ev Idx": 2198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770611469.159, "dur": 0.296,
    "args": {
      "External id": 2200,"Record function id": 0, "finished": true, "Ev Idx": 2199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770611470.482, "dur": 44.396,
    "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": 870, "tid": 870,
    "ts": 7429770611471.031, "dur": 1.324,
    "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": 870, "tid": 870,
    "ts": 7429770611471.296, "dur": 0.937,
    "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": 870, "tid": 870,
    "ts": 7429770611472.653, "dur": 40.550,
    "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": 870, "tid": 870,
    "ts": 7429770611483.727, "dur": 1.104,
    "args": {
      "External id": 2205,"Record function id": 0, "finished": true, "Ev Idx": 2204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770611513.722, "dur": 0.890,
    "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": 870, "tid": 870,
    "ts": 7429770611517.826, "dur": 7.959,
    "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": 870, "tid": 870,
    "ts": 7429770611545.859, "dur": 0.188,
    "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": 870, "tid": 870,
    "ts": 7429770611547.964, "dur": 9.397,
    "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": 870, "tid": 870,
    "ts": 7429770611549.191, "dur": 0.291,
    "args": {
      "External id": 2210,"Record function id": 0, "finished": true, "Ev Idx": 2209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770611550.018, "dur": 0.076,
    "args": {
      "External id": 2211,"Record function id": 0, "finished": true, "Ev Idx": 2210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770611559.873, "dur": 11.654,
    "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": 870, "tid": 870,
    "ts": 7429770611574.715, "dur": 8.395,
    "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": 870, "tid": 870,
    "ts": 7429770611584.779, "dur": 10.475,
    "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": 870, "tid": 870,
    "ts": 7429770611596.776, "dur": 7.389,
    "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": 870, "tid": 870,
    "ts": 7429770611606.502, "dur": 0.116,
    "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": 870, "tid": 870,
    "ts": 7429770611607.343, "dur": 8.323,
    "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": 870, "tid": 870,
    "ts": 7429770611633.539, "dur": 38.015,
    "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": 870, "tid": 870,
    "ts": 7429770611634.172, "dur": 4.378,
    "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": 870, "tid": 870,
    "ts": 7429770611635.916, "dur": 1.401,
    "args": {
      "External id": 2220,"Record function id": 0, "finished": true, "Ev Idx": 2219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770611636.699, "dur": 0.479,
    "args": {
      "External id": 2221,"Record function id": 0, "finished": true, "Ev Idx": 2220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770611639.017, "dur": 32.435,
    "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": 870, "tid": 870,
    "ts": 7429770611639.629, "dur": 1.706,
    "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": 870, "tid": 870,
    "ts": 7429770611640.108, "dur": 1.115,
    "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": 870, "tid": 870,
    "ts": 7429770611641.808, "dur": 27.648,
    "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": 870, "tid": 870,
    "ts": 7429770611653.118, "dur": 1.229,
    "args": {
      "External id": 2226,"Record function id": 0, "finished": true, "Ev Idx": 2225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770611670.105, "dur": 1.055,
    "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": 870, "tid": 870,
    "ts": 7429770611674.536, "dur": 1.301,
    "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": 870, "tid": 870,
    "ts": 7429770611677.582, "dur": 2.308,
    "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": 870, "tid": 870,
    "ts": 7429770611678.538, "dur": 0.600,
    "args": {
      "External id": 2230,"Record function id": 0, "finished": true, "Ev Idx": 2229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770611686.903, "dur": 32.933,
    "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": 870, "tid": 870,
    "ts": 7429770611687.192, "dur": 2.678,
    "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": 870, "tid": 870,
    "ts": 7429770611688.392, "dur": 0.861,
    "args": {
      "External id": 2233,"Record function id": 0, "finished": true, "Ev Idx": 2232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770611688.874, "dur": 0.277,
    "args": {
      "External id": 2234,"Record function id": 0, "finished": true, "Ev Idx": 2233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770611690.202, "dur": 29.510,
    "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": 870, "tid": 870,
    "ts": 7429770611690.587, "dur": 1.223,
    "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": 870, "tid": 870,
    "ts": 7429770611690.845, "dur": 0.869,
    "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": 870, "tid": 870,
    "ts": 7429770611692.122, "dur": 25.916,
    "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": 870, "tid": 870,
    "ts": 7429770611702.722, "dur": 1.106,
    "args": {
      "External id": 2239,"Record function id": 0, "finished": true, "Ev Idx": 2238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770611718.607, "dur": 0.874,
    "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": 870, "tid": 870,
    "ts": 7429770611722.028, "dur": 1.131,
    "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": 870, "tid": 870,
    "ts": 7429770611724.609, "dur": 1.828,
    "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": 870, "tid": 870,
    "ts": 7429770611725.454, "dur": 0.470,
    "args": {
      "External id": 2243,"Record function id": 0, "finished": true, "Ev Idx": 2242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770611732.875, "dur": 32.819,
    "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": 870, "tid": 870,
    "ts": 7429770611733.318, "dur": 2.047,
    "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": 870, "tid": 870,
    "ts": 7429770611734.215, "dur": 0.774,
    "args": {
      "External id": 2246,"Record function id": 0, "finished": true, "Ev Idx": 2245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770611734.645, "dur": 0.239,
    "args": {
      "External id": 2247,"Record function id": 0, "finished": true, "Ev Idx": 2246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770611735.795, "dur": 29.802,
    "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": 870, "tid": 870,
    "ts": 7429770611736.367, "dur": 1.535,
    "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": 870, "tid": 870,
    "ts": 7429770611736.782, "dur": 1.019,
    "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": 870, "tid": 870,
    "ts": 7429770611738.349, "dur": 25.617,
    "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": 870, "tid": 870,
    "ts": 7429770611748.150, "dur": 1.110,
    "args": {
      "External id": 2252,"Record function id": 0, "finished": true, "Ev Idx": 2251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770611764.484, "dur": 0.875,
    "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": 870, "tid": 870,
    "ts": 7429770611768.014, "dur": 1.099,
    "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": 870, "tid": 870,
    "ts": 7429770611770.677, "dur": 1.571,
    "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": 870, "tid": 870,
    "ts": 7429770611771.414, "dur": 0.354,
    "args": {
      "External id": 2256,"Record function id": 0, "finished": true, "Ev Idx": 2255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770611775.319, "dur": 1.762,
    "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": 870, "tid": 870,
    "ts": 7429770611776.209, "dur": 0.308,
    "args": {
      "External id": 2258,"Record function id": 0, "finished": true, "Ev Idx": 2257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770611778.107, "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": 870, "tid": 870,
    "ts": 7429770611778.547, "dur": 0.265,
    "args": {
      "External id": 2260,"Record function id": 0, "finished": true, "Ev Idx": 2259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770611780.031, "dur": 10.401,
    "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": 870, "tid": 870,
    "ts": 7429770611794.953, "dur": 3.565,
    "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": 870, "tid": 870,
    "ts": 7429770611796.742, "dur": 0.513,
    "args": {
      "External id": 2263,"Record function id": 0, "finished": true, "Ev Idx": 2262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770611800.565, "dur": 3.298,
    "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": 870, "tid": 870,
    "ts": 7429770611801.969, "dur": 0.604,
    "args": {
      "External id": 2265,"Record function id": 0, "finished": true, "Ev Idx": 2264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770611805.597, "dur": 10.949,
    "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": 870, "tid": 870,
    "ts": 7429770611819.993, "dur": 17.021,
    "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": 870, "tid": 870,
    "ts": 7429770611840.079, "dur": 7.517,
    "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": 870, "tid": 870,
    "ts": 7429770611849.193, "dur": 9.823,
    "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": 870, "tid": 870,
    "ts": 7429770611861.081, "dur": 8.637,
    "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": 870, "tid": 870,
    "ts": 7429770611872.070, "dur": 2.203,
    "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": 870, "tid": 870,
    "ts": 7429770611873.239, "dur": 0.366,
    "args": {
      "External id": 2272,"Record function id": 0, "finished": true, "Ev Idx": 2271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770611875.813, "dur": 1.788,
    "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": 870, "tid": 870,
    "ts": 7429770611876.758, "dur": 0.314,
    "args": {
      "External id": 2274,"Record function id": 0, "finished": true, "Ev Idx": 2273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770611878.571, "dur": 10.633,
    "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": 870, "tid": 870,
    "ts": 7429770611890.744, "dur": 9.476,
    "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": 870, "tid": 870,
    "ts": 7429770611902.305, "dur": 9.479,
    "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": 870, "tid": 870,
    "ts": 7429770611913.364, "dur": 7.050,
    "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": 870, "tid": 870,
    "ts": 7429770611942.789, "dur": 2.802,
    "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": 870, "tid": 870,
    "ts": 7429770611944.259, "dur": 0.538,
    "args": {
      "External id": 2280,"Record function id": 0, "finished": true, "Ev Idx": 2279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770611946.377, "dur": 2.933,
    "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": 870, "tid": 870,
    "ts": 7429770611948.267, "dur": 0.509,
    "args": {
      "External id": 2282,"Record function id": 0, "finished": true, "Ev Idx": 2281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770611950.129, "dur": 2.439,
    "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": 870, "tid": 870,
    "ts": 7429770611951.252, "dur": 0.651,
    "args": {
      "External id": 2284,"Record function id": 0, "finished": true, "Ev Idx": 2283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770611953.207, "dur": 2.299,
    "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": 870, "tid": 870,
    "ts": 7429770611954.452, "dur": 0.523,
    "args": {
      "External id": 2286,"Record function id": 0, "finished": true, "Ev Idx": 2285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770611956.129, "dur": 1.640,
    "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": 870, "tid": 870,
    "ts": 7429770611956.818, "dur": 0.469,
    "args": {
      "External id": 2288,"Record function id": 0, "finished": true, "Ev Idx": 2287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770611960.449, "dur": 2.639,
    "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": 870, "tid": 870,
    "ts": 7429770611962.017, "dur": 0.389,
    "args": {
      "External id": 2290,"Record function id": 0, "finished": true, "Ev Idx": 2289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770611965.119, "dur": 18.579,
    "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": 870, "tid": 870,
    "ts": 7429770611967.178, "dur": 13.592,
    "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": 870, "tid": 870,
    "ts": 7429770611968.373, "dur": 2.469,
    "args": {
      "External id": 2293,"Record function id": 0, "finished": true, "Ev Idx": 2292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770611969.061, "dur": 1.594,
    "args": {
      "External id": 2294,"Record function id": 0, "finished": true, "Ev Idx": 2293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770611971.531, "dur": 8.771,
    "args": {
      "External id": 2295,"Record function id": 0, "finished": true, "Ev Idx": 2294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770611981.243, "dur": 2.167,
    "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": 870, "tid": 870,
    "ts": 7429770611986.734, "dur": 2.349,
    "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": 870, "tid": 870,
    "ts": 7429770611988.097, "dur": 0.387,
    "args": {
      "External id": 2298,"Record function id": 0, "finished": true, "Ev Idx": 2297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770611989.708, "dur": 1.733,
    "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": 870, "tid": 870,
    "ts": 7429770611990.679, "dur": 0.296,
    "args": {
      "External id": 2300,"Record function id": 0, "finished": true, "Ev Idx": 2299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770611991.967, "dur": 1.802,
    "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": 870, "tid": 870,
    "ts": 7429770611993.069, "dur": 0.315,
    "args": {
      "External id": 2302,"Record function id": 0, "finished": true, "Ev Idx": 2301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770611994.471, "dur": 1.921,
    "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": 870, "tid": 870,
    "ts": 7429770611995.593, "dur": 0.311,
    "args": {
      "External id": 2304,"Record function id": 0, "finished": true, "Ev Idx": 2303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770611997.116, "dur": 1.573,
    "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": 870, "tid": 870,
    "ts": 7429770611997.939, "dur": 0.302,
    "args": {
      "External id": 2306,"Record function id": 0, "finished": true, "Ev Idx": 2305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770612000.223, "dur": 2.338,
    "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": 870, "tid": 870,
    "ts": 7429770612001.656, "dur": 0.457,
    "args": {
      "External id": 2308,"Record function id": 0, "finished": true, "Ev Idx": 2307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770612003.951, "dur": 11.529,
    "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": 870, "tid": 870,
    "ts": 7429770612004.879, "dur": 9.107,
    "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": 870, "tid": 870,
    "ts": 7429770612006.100, "dur": 1.707,
    "args": {
      "External id": 2311,"Record function id": 0, "finished": true, "Ev Idx": 2310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770612006.452, "dur": 1.248,
    "args": {
      "External id": 2312,"Record function id": 0, "finished": true, "Ev Idx": 2311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770612008.068, "dur": 5.599,
    "args": {
      "External id": 2313,"Record function id": 0, "finished": true, "Ev Idx": 2312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770612014.364, "dur": 0.936,
    "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": 870, "tid": 870,
    "ts": 7429770612018.801, "dur": 1.862,
    "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": 870, "tid": 870,
    "ts": 7429770612019.785, "dur": 0.411,
    "args": {
      "External id": 2316,"Record function id": 0, "finished": true, "Ev Idx": 2315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770612022.080, "dur": 44.378,
    "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": 870, "tid": 870,
    "ts": 7429770612022.620, "dur": 1.900,
    "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": 870, "tid": 870,
    "ts": 7429770612023.720, "dur": 0.327,
    "args": {
      "External id": 2319,"Record function id": 0, "finished": true, "Ev Idx": 2318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770612024.861, "dur": 3.188,
    "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": 870, "tid": 870,
    "ts": 7429770612026.489, "dur": 1.425,
    "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": 870, "tid": 870,
    "ts": 7429770612028.486, "dur": 2.521,
    "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": 870, "tid": 870,
    "ts": 7429770612030.367, "dur": 0.294,
    "args": {
      "External id": 2323,"Record function id": 0, "finished": true, "Ev Idx": 2322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770612031.300, "dur": 2.486,
    "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": 870, "tid": 870,
    "ts": 7429770612032.081, "dur": 1.614,
    "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": 870, "tid": 870,
    "ts": 7429770612034.299, "dur": 30.217,
    "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": 870, "tid": 870,
    "ts": 7429770612045.547, "dur": 1.113,
    "args": {
      "External id": 2327,"Record function id": 0, "finished": true, "Ev Idx": 2326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770612065.088, "dur": 1.074,
    "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": 870, "tid": 870,
    "ts": 7429770612069.348, "dur": 8.963,
    "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": 870, "tid": 870,
    "ts": 7429770612081.501, "dur": 1.851,
    "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": 870, "tid": 870,
    "ts": 7429770612082.238, "dur": 0.439,
    "args": {
      "External id": 2331,"Record function id": 0, "finished": true, "Ev Idx": 2330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770612083.957, "dur": 1.247,
    "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": 870, "tid": 870,
    "ts": 7429770612084.517, "dur": 0.323,
    "args": {
      "External id": 2333,"Record function id": 0, "finished": true, "Ev Idx": 2332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770612085.796, "dur": 1.033,
    "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": 870, "tid": 870,
    "ts": 7429770612086.247, "dur": 0.234,
    "args": {
      "External id": 2335,"Record function id": 0, "finished": true, "Ev Idx": 2334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770612088.312, "dur": 7.526,
    "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": 870, "tid": 870,
    "ts": 7429770612100.733, "dur": 13.430,
    "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": 870, "tid": 870,
    "ts": 7429770612101.712, "dur": 0.335,
    "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": 870, "tid": 870,
    "ts": 7429770612102.577, "dur": 11.314,
    "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": 870, "tid": 870,
    "ts": 7429770612116.912, "dur": 0.353,
    "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": 870, "tid": 870,
    "ts": 7429770612127.686, "dur": 0.365,
    "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": 870, "tid": 870,
    "ts": 7429770612129.732, "dur": 39.328,
    "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": 870, "tid": 870,
    "ts": 7429770612130.171, "dur": 1.875,
    "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": 870, "tid": 870,
    "ts": 7429770612131.069, "dur": 0.351,
    "args": {
      "External id": 2344,"Record function id": 0, "finished": true, "Ev Idx": 2343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770612132.408, "dur": 2.888,
    "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": 870, "tid": 870,
    "ts": 7429770612132.759, "dur": 2.418,
    "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": 870, "tid": 870,
    "ts": 7429770612135.819, "dur": 2.031,
    "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": 870, "tid": 870,
    "ts": 7429770612137.241, "dur": 0.232,
    "args": {
      "External id": 2348,"Record function id": 0, "finished": true, "Ev Idx": 2347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770612138.285, "dur": 1.866,
    "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": 870, "tid": 870,
    "ts": 7429770612138.696, "dur": 1.382,
    "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": 870, "tid": 870,
    "ts": 7429770612140.669, "dur": 26.398,
    "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": 870, "tid": 870,
    "ts": 7429770612150.699, "dur": 1.124,
    "args": {
      "External id": 2352,"Record function id": 0, "finished": true, "Ev Idx": 2351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770612167.622, "dur": 1.168,
    "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": 870, "tid": 870,
    "ts": 7429770612171.028, "dur": 1.911,
    "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": 870, "tid": 870,
    "ts": 7429770612172.074, "dur": 0.335,
    "args": {
      "External id": 2355,"Record function id": 0, "finished": true, "Ev Idx": 2354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770612175.533, "dur": 11.664,
    "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": 870, "tid": 870,
    "ts": 7429770612176.241, "dur": 10.811,
    "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": 870, "tid": 870,
    "ts": 7429770612177.541, "dur": 1.840,
    "args": {
      "External id": 2358,"Record function id": 0, "finished": true, "Ev Idx": 2357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770612177.880, "dur": 1.357,
    "args": {
      "External id": 2359,"Record function id": 0, "finished": true, "Ev Idx": 2358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770612179.866, "dur": 6.772,
    "args": {
      "External id": 2360,"Record function id": 0, "finished": true, "Ev Idx": 2359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770612192.390, "dur": 1.913,
    "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": 870, "tid": 870,
    "ts": 7429770612192.889, "dur": 1.322,
    "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": 870, "tid": 870,
    "ts": 7429770612206.066, "dur": 37.210,
    "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": 870, "tid": 870,
    "ts": 7429770612206.676, "dur": 2.590,
    "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": 870, "tid": 870,
    "ts": 7429770612207.877, "dur": 0.873,
    "args": {
      "External id": 2365,"Record function id": 0, "finished": true, "Ev Idx": 2364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770612208.312, "dur": 0.308,
    "args": {
      "External id": 2366,"Record function id": 0, "finished": true, "Ev Idx": 2365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770612209.637, "dur": 33.500,
    "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": 870, "tid": 870,
    "ts": 7429770612210.332, "dur": 1.991,
    "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": 870, "tid": 870,
    "ts": 7429770612210.607, "dur": 1.619,
    "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": 870, "tid": 870,
    "ts": 7429770612212.852, "dur": 28.240,
    "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": 870, "tid": 870,
    "ts": 7429770612224.526, "dur": 1.143,
    "args": {
      "External id": 2371,"Record function id": 0, "finished": true, "Ev Idx": 2370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770612241.786, "dur": 1.065,
    "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": 870, "tid": 870,
    "ts": 7429770612248.003, "dur": 8.052,
    "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": 870, "tid": 870,
    "ts": 7429770612264.138, "dur": 0.179,
    "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": 870, "tid": 870,
    "ts": 7429770612266.077, "dur": 10.429,
    "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": 870, "tid": 870,
    "ts": 7429770612267.543, "dur": 0.527,
    "args": {
      "External id": 2376,"Record function id": 0, "finished": true, "Ev Idx": 2375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770612268.550, "dur": 0.105,
    "args": {
      "External id": 2377,"Record function id": 0, "finished": true, "Ev Idx": 2376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770612279.094, "dur": 12.123,
    "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": 870, "tid": 870,
    "ts": 7429770612294.673, "dur": 17.437,
    "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": 870, "tid": 870,
    "ts": 7429770612315.583, "dur": 10.990,
    "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": 870, "tid": 870,
    "ts": 7429770612328.479, "dur": 11.219,
    "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": 870, "tid": 870,
    "ts": 7429770612342.800, "dur": 0.128,
    "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": 870, "tid": 870,
    "ts": 7429770612343.902, "dur": 9.086,
    "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": 870, "tid": 870,
    "ts": 7429770612368.754, "dur": 43.742,
    "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": 870, "tid": 870,
    "ts": 7429770612369.551, "dur": 3.507,
    "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": 870, "tid": 870,
    "ts": 7429770612370.969, "dur": 1.409,
    "args": {
      "External id": 2386,"Record function id": 0, "finished": true, "Ev Idx": 2385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770612371.567, "dur": 0.604,
    "args": {
      "External id": 2387,"Record function id": 0, "finished": true, "Ev Idx": 2386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770612373.570, "dur": 38.793,
    "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": 870, "tid": 870,
    "ts": 7429770612374.460, "dur": 2.071,
    "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": 870, "tid": 870,
    "ts": 7429770612375.125, "dur": 1.256,
    "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": 870, "tid": 870,
    "ts": 7429770612376.909, "dur": 33.140,
    "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": 870, "tid": 870,
    "ts": 7429770612391.354, "dur": 1.289,
    "args": {
      "External id": 2392,"Record function id": 0, "finished": true, "Ev Idx": 2391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770612410.768, "dur": 1.200,
    "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": 870, "tid": 870,
    "ts": 7429770612420.392, "dur": 9.366,
    "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": 870, "tid": 870,
    "ts": 7429770612437.238, "dur": 33.594,
    "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": 870, "tid": 870,
    "ts": 7429770612437.627, "dur": 2.588,
    "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": 870, "tid": 870,
    "ts": 7429770612438.815, "dur": 0.898,
    "args": {
      "External id": 2397,"Record function id": 0, "finished": true, "Ev Idx": 2396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770612439.183, "dur": 0.378,
    "args": {
      "External id": 2398,"Record function id": 0, "finished": true, "Ev Idx": 2397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770612440.563, "dur": 30.163,
    "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": 870, "tid": 870,
    "ts": 7429770612441.026, "dur": 1.701,
    "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": 870, "tid": 870,
    "ts": 7429770612441.432, "dur": 1.191,
    "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": 870, "tid": 870,
    "ts": 7429770612443.052, "dur": 25.755,
    "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": 870, "tid": 870,
    "ts": 7429770612453.088, "dur": 1.181,
    "args": {
      "External id": 2403,"Record function id": 0, "finished": true, "Ev Idx": 2402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770612469.415, "dur": 1.072,
    "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": 870, "tid": 870,
    "ts": 7429770612472.309, "dur": 8.066,
    "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": 870, "tid": 870,
    "ts": 7429770612486.184, "dur": 47.491,
    "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": 870, "tid": 870,
    "ts": 7429770612486.734, "dur": 2.120,
    "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": 870, "tid": 870,
    "ts": 7429770612487.559, "dur": 0.864,
    "args": {
      "External id": 2408,"Record function id": 0, "finished": true, "Ev Idx": 2407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770612487.946, "dur": 0.358,
    "args": {
      "External id": 2409,"Record function id": 0, "finished": true, "Ev Idx": 2408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770612489.321, "dur": 44.249,
    "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": 870, "tid": 870,
    "ts": 7429770612490.033, "dur": 1.529,
    "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": 870, "tid": 870,
    "ts": 7429770612490.462, "dur": 0.988,
    "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": 870, "tid": 870,
    "ts": 7429770612492.013, "dur": 39.821,
    "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": 870, "tid": 870,
    "ts": 7429770612501.911, "dur": 1.055,
    "args": {
      "External id": 2414,"Record function id": 0, "finished": true, "Ev Idx": 2413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770612532.433, "dur": 0.882,
    "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": 870, "tid": 870,
    "ts": 7429770612536.228, "dur": 7.553,
    "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": 870, "tid": 870,
    "ts": 7429770612562.633, "dur": 0.136,
    "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": 870, "tid": 870,
    "ts": 7429770612564.764, "dur": 9.690,
    "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": 870, "tid": 870,
    "ts": 7429770612566.288, "dur": 0.352,
    "args": {
      "External id": 2419,"Record function id": 0, "finished": true, "Ev Idx": 2418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770612567.150, "dur": 0.256,
    "args": {
      "External id": 2420,"Record function id": 0, "finished": true, "Ev Idx": 2419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770612577.093, "dur": 11.925,
    "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": 870, "tid": 870,
    "ts": 7429770612592.890, "dur": 9.443,
    "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": 870, "tid": 870,
    "ts": 7429770612603.822, "dur": 10.425,
    "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": 870, "tid": 870,
    "ts": 7429770612615.781, "dur": 8.803,
    "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": 870, "tid": 870,
    "ts": 7429770612627.072, "dur": 0.102,
    "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": 870, "tid": 870,
    "ts": 7429770612627.887, "dur": 8.901,
    "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": 870, "tid": 870,
    "ts": 7429770612654.451, "dur": 35.812,
    "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": 870, "tid": 870,
    "ts": 7429770612654.848, "dur": 2.792,
    "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": 870, "tid": 870,
    "ts": 7429770612655.886, "dur": 1.155,
    "args": {
      "External id": 2429,"Record function id": 0, "finished": true, "Ev Idx": 2428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770612656.355, "dur": 0.509,
    "args": {
      "External id": 2430,"Record function id": 0, "finished": true, "Ev Idx": 2429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770612657.990, "dur": 32.171,
    "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": 870, "tid": 870,
    "ts": 7429770612658.497, "dur": 1.527,
    "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": 870, "tid": 870,
    "ts": 7429770612658.778, "dur": 1.119,
    "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": 870, "tid": 870,
    "ts": 7429770612660.419, "dur": 27.681,
    "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": 870, "tid": 870,
    "ts": 7429770612671.709, "dur": 1.128,
    "args": {
      "External id": 2435,"Record function id": 0, "finished": true, "Ev Idx": 2434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770612688.717, "dur": 1.182,
    "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": 870, "tid": 870,
    "ts": 7429770612693.291, "dur": 1.457,
    "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": 870, "tid": 870,
    "ts": 7429770612696.645, "dur": 2.520,
    "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": 870, "tid": 870,
    "ts": 7429770612697.902, "dur": 0.551,
    "args": {
      "External id": 2439,"Record function id": 0, "finished": true, "Ev Idx": 2438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770612706.235, "dur": 29.610,
    "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": 870, "tid": 870,
    "ts": 7429770612706.560, "dur": 2.343,
    "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": 870, "tid": 870,
    "ts": 7429770612707.512, "dur": 0.963,
    "args": {
      "External id": 2442,"Record function id": 0, "finished": true, "Ev Idx": 2441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770612707.909, "dur": 0.446,
    "args": {
      "External id": 2443,"Record function id": 0, "finished": true, "Ev Idx": 2442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770612709.218, "dur": 26.507,
    "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": 870, "tid": 870,
    "ts": 7429770612709.650, "dur": 1.369,
    "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": 870, "tid": 870,
    "ts": 7429770612710.054, "dur": 0.872,
    "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": 870, "tid": 870,
    "ts": 7429770612711.336, "dur": 22.595,
    "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": 870, "tid": 870,
    "ts": 7429770612719.181, "dur": 1.207,
    "args": {
      "External id": 2448,"Record function id": 0, "finished": true, "Ev Idx": 2447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770612734.517, "dur": 0.963,
    "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": 870, "tid": 870,
    "ts": 7429770612738.012, "dur": 1.100,
    "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": 870, "tid": 870,
    "ts": 7429770612740.474, "dur": 1.820,
    "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": 870, "tid": 870,
    "ts": 7429770612741.260, "dur": 0.508,
    "args": {
      "External id": 2452,"Record function id": 0, "finished": true, "Ev Idx": 2451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770612749.307, "dur": 30.435,
    "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": 870, "tid": 870,
    "ts": 7429770612749.612, "dur": 1.968,
    "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": 870, "tid": 870,
    "ts": 7429770612750.479, "dur": 0.718,
    "args": {
      "External id": 2455,"Record function id": 0, "finished": true, "Ev Idx": 2454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770612750.819, "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": 870, "tid": 870,
    "ts": 7429770612751.844, "dur": 27.802,
    "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": 870, "tid": 870,
    "ts": 7429770612752.258, "dur": 1.803,
    "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": 870, "tid": 870,
    "ts": 7429770612752.728, "dur": 1.216,
    "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": 870, "tid": 870,
    "ts": 7429770612754.396, "dur": 23.435,
    "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": 870, "tid": 870,
    "ts": 7429770612762.639, "dur": 1.065,
    "args": {
      "External id": 2461,"Record function id": 0, "finished": true, "Ev Idx": 2460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770612778.491, "dur": 0.916,
    "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": 870, "tid": 870,
    "ts": 7429770612782.255, "dur": 1.222,
    "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": 870, "tid": 870,
    "ts": 7429770612786.840, "dur": 1.599,
    "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": 870, "tid": 870,
    "ts": 7429770612787.633, "dur": 0.351,
    "args": {
      "External id": 2465,"Record function id": 0, "finished": true, "Ev Idx": 2464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770612790.972, "dur": 1.919,
    "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": 870, "tid": 870,
    "ts": 7429770612792.150, "dur": 0.288,
    "args": {
      "External id": 2467,"Record function id": 0, "finished": true, "Ev Idx": 2466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770612793.923, "dur": 1.026,
    "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": 870, "tid": 870,
    "ts": 7429770612794.341, "dur": 0.300,
    "args": {
      "External id": 2469,"Record function id": 0, "finished": true, "Ev Idx": 2468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770612795.768, "dur": 11.458,
    "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": 870, "tid": 870,
    "ts": 7429770612812.286, "dur": 3.637,
    "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": 870, "tid": 870,
    "ts": 7429770612814.219, "dur": 0.490,
    "args": {
      "External id": 2472,"Record function id": 0, "finished": true, "Ev Idx": 2471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770612818.265, "dur": 1.830,
    "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": 870, "tid": 870,
    "ts": 7429770612819.156, "dur": 0.342,
    "args": {
      "External id": 2474,"Record function id": 0, "finished": true, "Ev Idx": 2473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770612821.889, "dur": 9.847,
    "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": 870, "tid": 870,
    "ts": 7429770612835.739, "dur": 15.513,
    "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": 870, "tid": 870,
    "ts": 7429770612854.399, "dur": 7.584,
    "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": 870, "tid": 870,
    "ts": 7429770612863.687, "dur": 7.654,
    "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": 870, "tid": 870,
    "ts": 7429770612873.436, "dur": 6.937,
    "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": 870, "tid": 870,
    "ts": 7429770612882.667, "dur": 2.229,
    "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": 870, "tid": 870,
    "ts": 7429770612883.857, "dur": 0.395,
    "args": {
      "External id": 2481,"Record function id": 0, "finished": true, "Ev Idx": 2480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770612886.470, "dur": 2.866,
    "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": 870, "tid": 870,
    "ts": 7429770612888.471, "dur": 0.298,
    "args": {
      "External id": 2483,"Record function id": 0, "finished": true, "Ev Idx": 2482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770612890.308, "dur": 8.481,
    "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": 870, "tid": 870,
    "ts": 7429770612900.475, "dur": 12.019,
    "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": 870, "tid": 870,
    "ts": 7429770612914.902, "dur": 7.775,
    "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": 870, "tid": 870,
    "ts": 7429770612923.999, "dur": 7.023,
    "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": 870, "tid": 870,
    "ts": 7429770612953.027, "dur": 2.494,
    "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": 870, "tid": 870,
    "ts": 7429770612954.270, "dur": 0.427,
    "args": {
      "External id": 2489,"Record function id": 0, "finished": true, "Ev Idx": 2488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770612956.201, "dur": 2.487,
    "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": 870, "tid": 870,
    "ts": 7429770612957.827, "dur": 0.307,
    "args": {
      "External id": 2491,"Record function id": 0, "finished": true, "Ev Idx": 2490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770612959.745, "dur": 3.131,
    "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": 870, "tid": 870,
    "ts": 7429770612961.309, "dur": 0.909,
    "args": {
      "External id": 2493,"Record function id": 0, "finished": true, "Ev Idx": 2492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770612963.668, "dur": 1.808,
    "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": 870, "tid": 870,
    "ts": 7429770612964.583, "dur": 0.380,
    "args": {
      "External id": 2495,"Record function id": 0, "finished": true, "Ev Idx": 2494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770612966.093, "dur": 1.774,
    "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": 870, "tid": 870,
    "ts": 7429770612966.831, "dur": 0.559,
    "args": {
      "External id": 2497,"Record function id": 0, "finished": true, "Ev Idx": 2496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770612970.531, "dur": 3.569,
    "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": 870, "tid": 870,
    "ts": 7429770612972.783, "dur": 0.682,
    "args": {
      "External id": 2499,"Record function id": 0, "finished": true, "Ev Idx": 2498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770612976.502, "dur": 19.331,
    "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": 870, "tid": 870,
    "ts": 7429770612979.213, "dur": 14.370,
    "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": 870, "tid": 870,
    "ts": 7429770612981.049, "dur": 2.728,
    "args": {
      "External id": 2502,"Record function id": 0, "finished": true, "Ev Idx": 2501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770612981.994, "dur": 1.532,
    "args": {
      "External id": 2503,"Record function id": 0, "finished": true, "Ev Idx": 2502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770612984.477, "dur": 8.658,
    "args": {
      "External id": 2504,"Record function id": 0, "finished": true, "Ev Idx": 2503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770612994.062, "dur": 1.383,
    "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": 870, "tid": 870,
    "ts": 7429770612998.971, "dur": 1.858,
    "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": 870, "tid": 870,
    "ts": 7429770612999.827, "dur": 0.401,
    "args": {
      "External id": 2507,"Record function id": 0, "finished": true, "Ev Idx": 2506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770613001.432, "dur": 1.845,
    "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": 870, "tid": 870,
    "ts": 7429770613002.345, "dur": 0.454,
    "args": {
      "External id": 2509,"Record function id": 0, "finished": true, "Ev Idx": 2508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770613003.814, "dur": 2.712,
    "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": 870, "tid": 870,
    "ts": 7429770613005.790, "dur": 0.318,
    "args": {
      "External id": 2511,"Record function id": 0, "finished": true, "Ev Idx": 2510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770613007.068, "dur": 2.678,
    "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": 870, "tid": 870,
    "ts": 7429770613008.804, "dur": 0.443,
    "args": {
      "External id": 2513,"Record function id": 0, "finished": true, "Ev Idx": 2512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770613010.468, "dur": 1.949,
    "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": 870, "tid": 870,
    "ts": 7429770613011.509, "dur": 0.418,
    "args": {
      "External id": 2515,"Record function id": 0, "finished": true, "Ev Idx": 2514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770613013.914, "dur": 2.100,
    "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": 870, "tid": 870,
    "ts": 7429770613015.170, "dur": 0.378,
    "args": {
      "External id": 2517,"Record function id": 0, "finished": true, "Ev Idx": 2516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770613017.407, "dur": 13.046,
    "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": 870, "tid": 870,
    "ts": 7429770613018.516, "dur": 8.990,
    "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": 870, "tid": 870,
    "ts": 7429770613019.394, "dur": 1.641,
    "args": {
      "External id": 2520,"Record function id": 0, "finished": true, "Ev Idx": 2519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770613019.770, "dur": 1.128,
    "args": {
      "External id": 2521,"Record function id": 0, "finished": true, "Ev Idx": 2520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770613021.370, "dur": 5.837,
    "args": {
      "External id": 2522,"Record function id": 0, "finished": true, "Ev Idx": 2521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770613027.880, "dur": 2.431,
    "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": 870, "tid": 870,
    "ts": 7429770613032.442, "dur": 1.945,
    "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": 870, "tid": 870,
    "ts": 7429770613033.523, "dur": 0.355,
    "args": {
      "External id": 2525,"Record function id": 0, "finished": true, "Ev Idx": 2524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770613035.966, "dur": 47.005,
    "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": 870, "tid": 870,
    "ts": 7429770613036.543, "dur": 1.624,
    "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": 870, "tid": 870,
    "ts": 7429770613037.359, "dur": 0.341,
    "args": {
      "External id": 2528,"Record function id": 0, "finished": true, "Ev Idx": 2527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770613038.525, "dur": 3.511,
    "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": 870, "tid": 870,
    "ts": 7429770613040.189, "dur": 1.694,
    "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": 870, "tid": 870,
    "ts": 7429770613042.499, "dur": 1.330,
    "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": 870, "tid": 870,
    "ts": 7429770613043.184, "dur": 0.287,
    "args": {
      "External id": 2532,"Record function id": 0, "finished": true, "Ev Idx": 2531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770613044.121, "dur": 1.777,
    "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": 870, "tid": 870,
    "ts": 7429770613044.960, "dur": 0.836,
    "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": 870, "tid": 870,
    "ts": 7429770613046.484, "dur": 34.451,
    "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": 870, "tid": 870,
    "ts": 7429770613062.344, "dur": 1.120,
    "args": {
      "External id": 2536,"Record function id": 0, "finished": true, "Ev Idx": 2535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770613081.527, "dur": 1.047,
    "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": 870, "tid": 870,
    "ts": 7429770613085.956, "dur": 10.072,
    "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": 870, "tid": 870,
    "ts": 7429770613099.362, "dur": 2.158,
    "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": 870, "tid": 870,
    "ts": 7429770613100.299, "dur": 0.543,
    "args": {
      "External id": 2540,"Record function id": 0, "finished": true, "Ev Idx": 2539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770613102.131, "dur": 1.186,
    "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": 870, "tid": 870,
    "ts": 7429770613102.631, "dur": 0.299,
    "args": {
      "External id": 2542,"Record function id": 0, "finished": true, "Ev Idx": 2541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770613103.965, "dur": 1.195,
    "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": 870, "tid": 870,
    "ts": 7429770613104.569, "dur": 0.242,
    "args": {
      "External id": 2544,"Record function id": 0, "finished": true, "Ev Idx": 2543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770613106.821, "dur": 7.922,
    "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": 870, "tid": 870,
    "ts": 7429770613119.753, "dur": 13.759,
    "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": 870, "tid": 870,
    "ts": 7429770613120.872, "dur": 0.469,
    "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": 870, "tid": 870,
    "ts": 7429770613122.015, "dur": 11.230,
    "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": 870, "tid": 870,
    "ts": 7429770613136.025, "dur": 0.197,
    "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": 870, "tid": 870,
    "ts": 7429770613146.913, "dur": 0.488,
    "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": 870, "tid": 870,
    "ts": 7429770613149.149, "dur": 159.156,
    "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": 870, "tid": 870,
    "ts": 7429770613149.690, "dur": 2.504,
    "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": 870, "tid": 870,
    "ts": 7429770613150.915, "dur": 0.673,
    "args": {
      "External id": 2553,"Record function id": 0, "finished": true, "Ev Idx": 2552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770613152.623, "dur": 2.785,
    "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": 870, "tid": 870,
    "ts": 7429770613153.154, "dur": 2.130,
    "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": 870, "tid": 870,
    "ts": 7429770613155.848, "dur": 2.404,
    "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": 870, "tid": 870,
    "ts": 7429770613157.583, "dur": 0.265,
    "args": {
      "External id": 2557,"Record function id": 0, "finished": true, "Ev Idx": 2556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770613158.545, "dur": 2.573,
    "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": 870, "tid": 870,
    "ts": 7429770613158.798, "dur": 2.252,
    "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": 870, "tid": 870,
    "ts": 7429770613161.489, "dur": 144.033,
    "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": 870, "tid": 870,
    "ts": 7429770613206.419, "dur": 79.044,
    "args": {
      "External id": 2561,"Record function id": 0, "finished": true, "Ev Idx": 2560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770613306.468, "dur": 1.480,
    "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": 870, "tid": 870,
    "ts": 7429770613311.017, "dur": 2.504,
    "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": 870, "tid": 870,
    "ts": 7429770613312.233, "dur": 0.420,
    "args": {
      "External id": 2564,"Record function id": 0, "finished": true, "Ev Idx": 2563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770613316.501, "dur": 13.902,
    "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": 870, "tid": 870,
    "ts": 7429770613317.128, "dur": 13.135,
    "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": 870, "tid": 870,
    "ts": 7429770613318.272, "dur": 2.157,
    "args": {
      "External id": 2567,"Record function id": 0, "finished": true, "Ev Idx": 2566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770613318.762, "dur": 1.485,
    "args": {
      "External id": 2568,"Record function id": 0, "finished": true, "Ev Idx": 2567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770613320.905, "dur": 8.862,
    "args": {
      "External id": 2569,"Record function id": 0, "finished": true, "Ev Idx": 2568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770613337.439, "dur": 1.818,
    "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": 870, "tid": 870,
    "ts": 7429770613337.781, "dur": 1.378,
    "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": 870, "tid": 870,
    "ts": 7429770613355.994, "dur": 44.206,
    "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": 870, "tid": 870,
    "ts": 7429770613356.888, "dur": 3.458,
    "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": 870, "tid": 870,
    "ts": 7429770613358.551, "dur": 1.104,
    "args": {
      "External id": 2574,"Record function id": 0, "finished": true, "Ev Idx": 2573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770613359.178, "dur": 0.344,
    "args": {
      "External id": 2575,"Record function id": 0, "finished": true, "Ev Idx": 2574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770613360.837, "dur": 39.219,
    "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": 870, "tid": 870,
    "ts": 7429770613361.515, "dur": 2.412,
    "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": 870, "tid": 870,
    "ts": 7429770613361.766, "dur": 2.067,
    "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": 870, "tid": 870,
    "ts": 7429770613364.421, "dur": 33.577,
    "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": 870, "tid": 870,
    "ts": 7429770613378.914, "dur": 1.328,
    "args": {
      "External id": 2580,"Record function id": 0, "finished": true, "Ev Idx": 2579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770613398.613, "dur": 1.074,
    "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": 870, "tid": 870,
    "ts": 7429770613407.020, "dur": 9.558,
    "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": 870, "tid": 870,
    "ts": 7429770613426.688, "dur": 0.201,
    "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": 870, "tid": 870,
    "ts": 7429770613429.351, "dur": 13.924,
    "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": 870, "tid": 870,
    "ts": 7429770613431.413, "dur": 0.482,
    "args": {
      "External id": 2585,"Record function id": 0, "finished": true, "Ev Idx": 2584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770613432.287, "dur": 0.089,
    "args": {
      "External id": 2586,"Record function id": 0, "finished": true, "Ev Idx": 2585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770613446.402, "dur": 14.261,
    "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": 870, "tid": 870,
    "ts": 7429770613464.584, "dur": 8.527,
    "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": 870, "tid": 870,
    "ts": 7429770613475.278, "dur": 12.650,
    "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": 870, "tid": 870,
    "ts": 7429770613490.016, "dur": 10.772,
    "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": 870, "tid": 870,
    "ts": 7429770613503.616, "dur": 0.149,
    "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": 870, "tid": 870,
    "ts": 7429770613504.600, "dur": 9.191,
    "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": 870, "tid": 870,
    "ts": 7429770613529.837, "dur": 40.160,
    "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": 870, "tid": 870,
    "ts": 7429770613530.233, "dur": 3.607,
    "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": 870, "tid": 870,
    "ts": 7429770613531.624, "dur": 1.596,
    "args": {
      "External id": 2595,"Record function id": 0, "finished": true, "Ev Idx": 2594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770613532.576, "dur": 0.480,
    "args": {
      "External id": 2596,"Record function id": 0, "finished": true, "Ev Idx": 2595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770613534.295, "dur": 35.572,
    "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": 870, "tid": 870,
    "ts": 7429770613535.150, "dur": 2.549,
    "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": 870, "tid": 870,
    "ts": 7429770613535.412, "dur": 2.156,
    "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": 870, "tid": 870,
    "ts": 7429770613538.094, "dur": 29.660,
    "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": 870, "tid": 870,
    "ts": 7429770613550.207, "dur": 1.252,
    "args": {
      "External id": 2601,"Record function id": 0, "finished": true, "Ev Idx": 2600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770613568.447, "dur": 1.111,
    "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": 870, "tid": 870,
    "ts": 7429770613578.985, "dur": 11.015,
    "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": 870, "tid": 870,
    "ts": 7429770613597.483, "dur": 34.662,
    "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": 870, "tid": 870,
    "ts": 7429770613597.833, "dur": 2.218,
    "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": 870, "tid": 870,
    "ts": 7429770613598.611, "dur": 0.931,
    "args": {
      "External id": 2606,"Record function id": 0, "finished": true, "Ev Idx": 2605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770613599.004, "dur": 0.387,
    "args": {
      "External id": 2607,"Record function id": 0, "finished": true, "Ev Idx": 2606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770613600.403, "dur": 31.641,
    "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": 870, "tid": 870,
    "ts": 7429770613600.798, "dur": 2.373,
    "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": 870, "tid": 870,
    "ts": 7429770613601.241, "dur": 1.792,
    "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": 870, "tid": 870,
    "ts": 7429770613603.613, "dur": 26.599,
    "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": 870, "tid": 870,
    "ts": 7429770613614.804, "dur": 1.233,
    "args": {
      "External id": 2612,"Record function id": 0, "finished": true, "Ev Idx": 2611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770613630.909, "dur": 0.897,
    "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": 870, "tid": 870,
    "ts": 7429770613633.781, "dur": 8.285,
    "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": 870, "tid": 870,
    "ts": 7429770613648.344, "dur": 48.178,
    "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": 870, "tid": 870,
    "ts": 7429770613648.807, "dur": 2.146,
    "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": 870, "tid": 870,
    "ts": 7429770613649.651, "dur": 0.811,
    "args": {
      "External id": 2617,"Record function id": 0, "finished": true, "Ev Idx": 2616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770613650.048, "dur": 0.295,
    "args": {
      "External id": 2618,"Record function id": 0, "finished": true, "Ev Idx": 2617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770613651.255, "dur": 45.167,
    "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": 870, "tid": 870,
    "ts": 7429770613651.665, "dur": 1.566,
    "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": 870, "tid": 870,
    "ts": 7429770613652.036, "dur": 1.073,
    "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": 870, "tid": 870,
    "ts": 7429770613653.522, "dur": 40.941,
    "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": 870, "tid": 870,
    "ts": 7429770613666.195, "dur": 1.098,
    "args": {
      "External id": 2623,"Record function id": 0, "finished": true, "Ev Idx": 2622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770613695.161, "dur": 1.012,
    "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": 870, "tid": 870,
    "ts": 7429770613699.096, "dur": 8.499,
    "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": 870, "tid": 870,
    "ts": 7429770613726.532, "dur": 0.137,
    "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": 870, "tid": 870,
    "ts": 7429770613728.368, "dur": 8.628,
    "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": 870, "tid": 870,
    "ts": 7429770613729.673, "dur": 0.238,
    "args": {
      "External id": 2628,"Record function id": 0, "finished": true, "Ev Idx": 2627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770613730.199, "dur": 0.066,
    "args": {
      "External id": 2629,"Record function id": 0, "finished": true, "Ev Idx": 2628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770613739.252, "dur": 11.092,
    "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": 870, "tid": 870,
    "ts": 7429770613753.346, "dur": 7.823,
    "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": 870, "tid": 870,
    "ts": 7429770613762.597, "dur": 10.851,
    "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": 870, "tid": 870,
    "ts": 7429770613774.824, "dur": 7.123,
    "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": 870, "tid": 870,
    "ts": 7429770613784.169, "dur": 0.097,
    "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": 870, "tid": 870,
    "ts": 7429770613784.948, "dur": 8.887,
    "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": 870, "tid": 870,
    "ts": 7429770613812.012, "dur": 37.702,
    "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": 870, "tid": 870,
    "ts": 7429770613812.601, "dur": 3.027,
    "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": 870, "tid": 870,
    "ts": 7429770613813.459, "dur": 1.221,
    "args": {
      "External id": 2638,"Record function id": 0, "finished": true, "Ev Idx": 2637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770613813.987, "dur": 0.507,
    "args": {
      "External id": 2639,"Record function id": 0, "finished": true, "Ev Idx": 2638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770613815.926, "dur": 33.690,
    "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": 870, "tid": 870,
    "ts": 7429770613816.500, "dur": 3.495,
    "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": 870, "tid": 870,
    "ts": 7429770613816.784, "dur": 3.091,
    "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": 870, "tid": 870,
    "ts": 7429770613820.341, "dur": 27.399,
    "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": 870, "tid": 870,
    "ts": 7429770613830.262, "dur": 1.187,
    "args": {
      "External id": 2644,"Record function id": 0, "finished": true, "Ev Idx": 2643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770613848.402, "dur": 0.915,
    "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": 870, "tid": 870,
    "ts": 7429770613852.999, "dur": 1.492,
    "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": 870, "tid": 870,
    "ts": 7429770613856.315, "dur": 2.348,
    "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": 870, "tid": 870,
    "ts": 7429770613857.257, "dur": 0.530,
    "args": {
      "External id": 2648,"Record function id": 0, "finished": true, "Ev Idx": 2647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770613866.230, "dur": 32.136,
    "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": 870, "tid": 870,
    "ts": 7429770613866.534, "dur": 2.033,
    "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": 870, "tid": 870,
    "ts": 7429770613867.273, "dur": 0.858,
    "args": {
      "External id": 2651,"Record function id": 0, "finished": true, "Ev Idx": 2650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770613867.746, "dur": 0.298,
    "args": {
      "External id": 2652,"Record function id": 0, "finished": true, "Ev Idx": 2651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770613868.923, "dur": 29.350,
    "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": 870, "tid": 870,
    "ts": 7429770613869.322, "dur": 1.491,
    "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": 870, "tid": 870,
    "ts": 7429770613869.731, "dur": 0.971,
    "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": 870, "tid": 870,
    "ts": 7429770613871.132, "dur": 25.580,
    "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": 870, "tid": 870,
    "ts": 7429770613880.540, "dur": 1.150,
    "args": {
      "External id": 2657,"Record function id": 0, "finished": true, "Ev Idx": 2656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770613897.224, "dur": 0.822,
    "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": 870, "tid": 870,
    "ts": 7429770613900.740, "dur": 1.292,
    "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": 870, "tid": 870,
    "ts": 7429770613903.350, "dur": 1.802,
    "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": 870, "tid": 870,
    "ts": 7429770613904.256, "dur": 0.369,
    "args": {
      "External id": 2661,"Record function id": 0, "finished": true, "Ev Idx": 2660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770613912.012, "dur": 31.126,
    "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": 870, "tid": 870,
    "ts": 7429770613912.505, "dur": 2.440,
    "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": 870, "tid": 870,
    "ts": 7429770613913.422, "dur": 1.096,
    "args": {
      "External id": 2664,"Record function id": 0, "finished": true, "Ev Idx": 2663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770613913.997, "dur": 0.406,
    "args": {
      "External id": 2665,"Record function id": 0, "finished": true, "Ev Idx": 2664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770613915.248, "dur": 27.796,
    "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": 870, "tid": 870,
    "ts": 7429770613915.620, "dur": 1.306,
    "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": 870, "tid": 870,
    "ts": 7429770613915.943, "dur": 0.885,
    "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": 870, "tid": 870,
    "ts": 7429770613917.216, "dur": 24.190,
    "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": 870, "tid": 870,
    "ts": 7429770613926.228, "dur": 1.039,
    "args": {
      "External id": 2670,"Record function id": 0, "finished": true, "Ev Idx": 2669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770613941.980, "dur": 0.827,
    "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": 870, "tid": 870,
    "ts": 7429770613945.381, "dur": 1.125,
    "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": 870, "tid": 870,
    "ts": 7429770613947.995, "dur": 1.719,
    "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": 870, "tid": 870,
    "ts": 7429770613948.914, "dur": 0.339,
    "args": {
      "External id": 2674,"Record function id": 0, "finished": true, "Ev Idx": 2673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770613952.733, "dur": 2.262,
    "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": 870, "tid": 870,
    "ts": 7429770613954.118, "dur": 0.326,
    "args": {
      "External id": 2676,"Record function id": 0, "finished": true, "Ev Idx": 2675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770613956.100, "dur": 1.007,
    "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": 870, "tid": 870,
    "ts": 7429770613956.516, "dur": 0.271,
    "args": {
      "External id": 2678,"Record function id": 0, "finished": true, "Ev Idx": 2677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770613958.218, "dur": 10.475,
    "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": 870, "tid": 870,
    "ts": 7429770613973.582, "dur": 3.968,
    "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": 870, "tid": 870,
    "ts": 7429770613975.615, "dur": 0.480,
    "args": {
      "External id": 2681,"Record function id": 0, "finished": true, "Ev Idx": 2680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770613979.635, "dur": 3.408,
    "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": 870, "tid": 870,
    "ts": 7429770613981.087, "dur": 0.959,
    "args": {
      "External id": 2683,"Record function id": 0, "finished": true, "Ev Idx": 2682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770613984.663, "dur": 11.183,
    "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": 870, "tid": 870,
    "ts": 7429770613999.804, "dur": 17.632,
    "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": 870, "tid": 870,
    "ts": 7429770614020.146, "dur": 8.453,
    "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": 870, "tid": 870,
    "ts": 7429770614030.212, "dur": 7.239,
    "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": 870, "tid": 870,
    "ts": 7429770614039.587, "dur": 8.257,
    "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": 870, "tid": 870,
    "ts": 7429770614050.190, "dur": 2.081,
    "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": 870, "tid": 870,
    "ts": 7429770614051.199, "dur": 0.413,
    "args": {
      "External id": 2690,"Record function id": 0, "finished": true, "Ev Idx": 2689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770614053.857, "dur": 1.651,
    "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": 870, "tid": 870,
    "ts": 7429770614054.680, "dur": 0.304,
    "args": {
      "External id": 2692,"Record function id": 0, "finished": true, "Ev Idx": 2691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770614056.351, "dur": 9.835,
    "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": 870, "tid": 870,
    "ts": 7429770614068.032, "dur": 31.733,
    "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": 870, "tid": 870,
    "ts": 7429770614103.952, "dur": 10.365,
    "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": 870, "tid": 870,
    "ts": 7429770614115.806, "dur": 8.490,
    "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": 870, "tid": 870,
    "ts": 7429770614149.405, "dur": 2.837,
    "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": 870, "tid": 870,
    "ts": 7429770614150.772, "dur": 0.561,
    "args": {
      "External id": 2698,"Record function id": 0, "finished": true, "Ev Idx": 2697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770614152.926, "dur": 1.854,
    "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": 870, "tid": 870,
    "ts": 7429770614153.759, "dur": 0.445,
    "args": {
      "External id": 2700,"Record function id": 0, "finished": true, "Ev Idx": 2699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770614155.512, "dur": 2.775,
    "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": 870, "tid": 870,
    "ts": 7429770614156.858, "dur": 0.815,
    "args": {
      "External id": 2702,"Record function id": 0, "finished": true, "Ev Idx": 2701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770614158.965, "dur": 2.772,
    "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": 870, "tid": 870,
    "ts": 7429770614160.888, "dur": 0.332,
    "args": {
      "External id": 2704,"Record function id": 0, "finished": true, "Ev Idx": 2703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770614162.482, "dur": 2.513,
    "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": 870, "tid": 870,
    "ts": 7429770614164.214, "dur": 0.310,
    "args": {
      "External id": 2706,"Record function id": 0, "finished": true, "Ev Idx": 2705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770614167.856, "dur": 3.283,
    "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": 870, "tid": 870,
    "ts": 7429770614169.922, "dur": 0.448,
    "args": {
      "External id": 2708,"Record function id": 0, "finished": true, "Ev Idx": 2707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770614173.606, "dur": 19.399,
    "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": 870, "tid": 870,
    "ts": 7429770614176.308, "dur": 14.189,
    "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": 870, "tid": 870,
    "ts": 7429770614177.676, "dur": 2.869,
    "args": {
      "External id": 2711,"Record function id": 0, "finished": true, "Ev Idx": 2710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770614178.502, "dur": 1.788,
    "args": {
      "External id": 2712,"Record function id": 0, "finished": true, "Ev Idx": 2711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770614181.219, "dur": 8.811,
    "args": {
      "External id": 2713,"Record function id": 0, "finished": true, "Ev Idx": 2712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770614190.987, "dur": 1.771,
    "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": 870, "tid": 870,
    "ts": 7429770614196.198, "dur": 2.625,
    "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": 870, "tid": 870,
    "ts": 7429770614197.949, "dur": 0.320,
    "args": {
      "External id": 2716,"Record function id": 0, "finished": true, "Ev Idx": 2715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770614199.418, "dur": 2.077,
    "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": 870, "tid": 870,
    "ts": 7429770614200.748, "dur": 0.292,
    "args": {
      "External id": 2718,"Record function id": 0, "finished": true, "Ev Idx": 2717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770614201.987, "dur": 1.623,
    "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": 870, "tid": 870,
    "ts": 7429770614202.877, "dur": 0.341,
    "args": {
      "External id": 2720,"Record function id": 0, "finished": true, "Ev Idx": 2719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770614204.306, "dur": 1.509,
    "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": 870, "tid": 870,
    "ts": 7429770614204.995, "dur": 0.316,
    "args": {
      "External id": 2722,"Record function id": 0, "finished": true, "Ev Idx": 2721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770614206.365, "dur": 1.844,
    "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": 870, "tid": 870,
    "ts": 7429770614207.339, "dur": 0.410,
    "args": {
      "External id": 2724,"Record function id": 0, "finished": true, "Ev Idx": 2723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770614209.692, "dur": 2.462,
    "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": 870, "tid": 870,
    "ts": 7429770614211.094, "dur": 0.635,
    "args": {
      "External id": 2726,"Record function id": 0, "finished": true, "Ev Idx": 2725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770614213.672, "dur": 12.900,
    "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": 870, "tid": 870,
    "ts": 7429770614214.497, "dur": 10.410,
    "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": 870, "tid": 870,
    "ts": 7429770614215.866, "dur": 2.089,
    "args": {
      "External id": 2729,"Record function id": 0, "finished": true, "Ev Idx": 2728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770614216.603, "dur": 1.244,
    "args": {
      "External id": 2730,"Record function id": 0, "finished": true, "Ev Idx": 2729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770614218.516, "dur": 6.076,
    "args": {
      "External id": 2731,"Record function id": 0, "finished": true, "Ev Idx": 2730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770614225.424, "dur": 0.994,
    "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": 870, "tid": 870,
    "ts": 7429770614228.727, "dur": 1.924,
    "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": 870, "tid": 870,
    "ts": 7429770614229.612, "dur": 0.517,
    "args": {
      "External id": 2734,"Record function id": 0, "finished": true, "Ev Idx": 2733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770614232.136, "dur": 49.623,
    "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": 870, "tid": 870,
    "ts": 7429770614232.945, "dur": 1.508,
    "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": 870, "tid": 870,
    "ts": 7429770614233.701, "dur": 0.329,
    "args": {
      "External id": 2737,"Record function id": 0, "finished": true, "Ev Idx": 2736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770614234.797, "dur": 4.628,
    "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": 870, "tid": 870,
    "ts": 7429770614236.710, "dur": 2.563,
    "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": 870, "tid": 870,
    "ts": 7429770614239.870, "dur": 1.970,
    "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": 870, "tid": 870,
    "ts": 7429770614241.157, "dur": 0.303,
    "args": {
      "External id": 2741,"Record function id": 0, "finished": true, "Ev Idx": 2740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770614242.150, "dur": 2.173,
    "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": 870, "tid": 870,
    "ts": 7429770614242.971, "dur": 1.228,
    "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": 870, "tid": 870,
    "ts": 7429770614245.023, "dur": 34.890,
    "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": 870, "tid": 870,
    "ts": 7429770614258.481, "dur": 1.146,
    "args": {
      "External id": 2745,"Record function id": 0, "finished": true, "Ev Idx": 2744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770614280.460, "dur": 1.001,
    "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": 870, "tid": 870,
    "ts": 7429770614285.273, "dur": 9.585,
    "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": 870, "tid": 870,
    "ts": 7429770614298.077, "dur": 1.753,
    "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": 870, "tid": 870,
    "ts": 7429770614298.776, "dur": 0.408,
    "args": {
      "External id": 2749,"Record function id": 0, "finished": true, "Ev Idx": 2748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770614300.439, "dur": 1.246,
    "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": 870, "tid": 870,
    "ts": 7429770614301.014, "dur": 0.294,
    "args": {
      "External id": 2751,"Record function id": 0, "finished": true, "Ev Idx": 2750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770614302.316, "dur": 1.253,
    "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": 870, "tid": 870,
    "ts": 7429770614302.949, "dur": 0.263,
    "args": {
      "External id": 2753,"Record function id": 0, "finished": true, "Ev Idx": 2752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770614305.381, "dur": 8.049,
    "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": 870, "tid": 870,
    "ts": 7429770614318.813, "dur": 13.973,
    "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": 870, "tid": 870,
    "ts": 7429770614320.089, "dur": 0.390,
    "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": 870, "tid": 870,
    "ts": 7429770614321.290, "dur": 11.219,
    "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": 870, "tid": 870,
    "ts": 7429770614335.473, "dur": 0.335,
    "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": 870, "tid": 870,
    "ts": 7429770614348.744, "dur": 0.653,
    "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": 870, "tid": 870,
    "ts": 7429770614351.167, "dur": 39.497,
    "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": 870, "tid": 870,
    "ts": 7429770614352.030, "dur": 2.216,
    "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": 870, "tid": 870,
    "ts": 7429770614353.394, "dur": 0.395,
    "args": {
      "External id": 2762,"Record function id": 0, "finished": true, "Ev Idx": 2761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770614354.738, "dur": 2.523,
    "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": 870, "tid": 870,
    "ts": 7429770614355.106, "dur": 2.036,
    "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": 870, "tid": 870,
    "ts": 7429770614357.692, "dur": 1.997,
    "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": 870, "tid": 870,
    "ts": 7429770614358.937, "dur": 0.402,
    "args": {
      "External id": 2766,"Record function id": 0, "finished": true, "Ev Idx": 2765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770614359.977, "dur": 2.097,
    "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": 870, "tid": 870,
    "ts": 7429770614360.388, "dur": 1.615,
    "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": 870, "tid": 870,
    "ts": 7429770614362.455, "dur": 26.326,
    "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": 870, "tid": 870,
    "ts": 7429770614372.110, "dur": 1.358,
    "args": {
      "External id": 2770,"Record function id": 0, "finished": true, "Ev Idx": 2769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770614389.327, "dur": 1.027,
    "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": 870, "tid": 870,
    "ts": 7429770614392.674, "dur": 1.772,
    "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": 870, "tid": 870,
    "ts": 7429770614393.637, "dur": 0.340,
    "args": {
      "External id": 2773,"Record function id": 0, "finished": true, "Ev Idx": 2772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770614396.919, "dur": 11.872,
    "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": 870, "tid": 870,
    "ts": 7429770614397.416, "dur": 11.233,
    "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": 870, "tid": 870,
    "ts": 7429770614398.393, "dur": 1.944,
    "args": {
      "External id": 2776,"Record function id": 0, "finished": true, "Ev Idx": 2775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770614398.786, "dur": 1.428,
    "args": {
      "External id": 2777,"Record function id": 0, "finished": true, "Ev Idx": 2776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770614400.679, "dur": 7.641,
    "args": {
      "External id": 2778,"Record function id": 0, "finished": true, "Ev Idx": 2777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770614414.089, "dur": 1.614,
    "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": 870, "tid": 870,
    "ts": 7429770614414.398, "dur": 1.199,
    "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": 870, "tid": 870,
    "ts": 7429770614429.970, "dur": 39.346,
    "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": 870, "tid": 870,
    "ts": 7429770614430.675, "dur": 3.258,
    "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": 870, "tid": 870,
    "ts": 7429770614432.176, "dur": 1.165,
    "args": {
      "External id": 2783,"Record function id": 0, "finished": true, "Ev Idx": 2782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770614432.671, "dur": 0.454,
    "args": {
      "External id": 2784,"Record function id": 0, "finished": true, "Ev Idx": 2783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770614434.330, "dur": 34.848,
    "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": 870, "tid": 870,
    "ts": 7429770614434.954, "dur": 2.307,
    "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": 870, "tid": 870,
    "ts": 7429770614435.540, "dur": 1.624,
    "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": 870, "tid": 870,
    "ts": 7429770614437.937, "dur": 29.285,
    "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": 870, "tid": 870,
    "ts": 7429770614449.992, "dur": 1.180,
    "args": {
      "External id": 2789,"Record function id": 0, "finished": true, "Ev Idx": 2788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770614467.805, "dur": 1.084,
    "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": 870, "tid": 870,
    "ts": 7429770614474.176, "dur": 9.606,
    "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": 870, "tid": 870,
    "ts": 7429770614492.476, "dur": 0.269,
    "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": 870, "tid": 870,
    "ts": 7429770614494.776, "dur": 10.639,
    "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": 870, "tid": 870,
    "ts": 7429770614496.531, "dur": 0.325,
    "args": {
      "External id": 2794,"Record function id": 0, "finished": true, "Ev Idx": 2793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770614497.239, "dur": 0.256,
    "args": {
      "External id": 2795,"Record function id": 0, "finished": true, "Ev Idx": 2794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770614508.060, "dur": 13.009,
    "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": 870, "tid": 870,
    "ts": 7429770614524.532, "dur": 9.003,
    "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": 870, "tid": 870,
    "ts": 7429770614535.622, "dur": 10.007,
    "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": 870, "tid": 870,
    "ts": 7429770614547.114, "dur": 10.045,
    "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": 870, "tid": 870,
    "ts": 7429770614559.600, "dur": 0.275,
    "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": 870, "tid": 870,
    "ts": 7429770614560.781, "dur": 8.831,
    "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": 870, "tid": 870,
    "ts": 7429770614584.164, "dur": 36.476,
    "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": 870, "tid": 870,
    "ts": 7429770614585.062, "dur": 2.900,
    "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": 870, "tid": 870,
    "ts": 7429770614586.137, "dur": 1.259,
    "args": {
      "External id": 2804,"Record function id": 0, "finished": true, "Ev Idx": 2803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770614586.854, "dur": 0.408,
    "args": {
      "External id": 2805,"Record function id": 0, "finished": true, "Ev Idx": 2804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770614588.321, "dur": 32.182,
    "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": 870, "tid": 870,
    "ts": 7429770614588.936, "dur": 2.408,
    "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": 870, "tid": 870,
    "ts": 7429770614589.212, "dur": 2.015,
    "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": 870, "tid": 870,
    "ts": 7429770614591.686, "dur": 27.080,
    "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": 870, "tid": 870,
    "ts": 7429770614602.594, "dur": 1.335,
    "args": {
      "External id": 2810,"Record function id": 0, "finished": true, "Ev Idx": 2809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770614619.335, "dur": 0.878,
    "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": 870, "tid": 870,
    "ts": 7429770614628.865, "dur": 10.106,
    "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": 870, "tid": 870,
    "ts": 7429770614647.416, "dur": 34.107,
    "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": 870, "tid": 870,
    "ts": 7429770614648.128, "dur": 2.480,
    "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": 870, "tid": 870,
    "ts": 7429770614649.059, "dur": 1.111,
    "args": {
      "External id": 2815,"Record function id": 0, "finished": true, "Ev Idx": 2814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770614649.651, "dur": 0.407,
    "args": {
      "External id": 2816,"Record function id": 0, "finished": true, "Ev Idx": 2815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770614650.973, "dur": 30.451,
    "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": 870, "tid": 870,
    "ts": 7429770614651.422, "dur": 1.611,
    "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": 870, "tid": 870,
    "ts": 7429770614651.830, "dur": 1.099,
    "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": 870, "tid": 870,
    "ts": 7429770614653.348, "dur": 26.154,
    "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": 870, "tid": 870,
    "ts": 7429770614664.069, "dur": 1.111,
    "args": {
      "External id": 2821,"Record function id": 0, "finished": true, "Ev Idx": 2820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770614680.092, "dur": 1.066,
    "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": 870, "tid": 870,
    "ts": 7429770614683.022, "dur": 7.938,
    "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": 870, "tid": 870,
    "ts": 7429770614697.841, "dur": 47.692,
    "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": 870, "tid": 870,
    "ts": 7429770614698.476, "dur": 2.294,
    "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": 870, "tid": 870,
    "ts": 7429770614699.444, "dur": 0.874,
    "args": {
      "External id": 2826,"Record function id": 0, "finished": true, "Ev Idx": 2825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770614699.897, "dur": 0.307,
    "args": {
      "External id": 2827,"Record function id": 0, "finished": true, "Ev Idx": 2826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770614701.214, "dur": 44.218,
    "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": 870, "tid": 870,
    "ts": 7429770614701.770, "dur": 1.308,
    "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": 870, "tid": 870,
    "ts": 7429770614702.024, "dur": 0.939,
    "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": 870, "tid": 870,
    "ts": 7429770614703.367, "dur": 40.350,
    "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": 870, "tid": 870,
    "ts": 7429770614716.087, "dur": 1.094,
    "args": {
      "External id": 2832,"Record function id": 0, "finished": true, "Ev Idx": 2831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770614744.284, "dur": 0.900,
    "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": 870, "tid": 870,
    "ts": 7429770614748.107, "dur": 7.828,
    "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": 870, "tid": 870,
    "ts": 7429770614774.696, "dur": 0.159,
    "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": 870, "tid": 870,
    "ts": 7429770614776.592, "dur": 8.942,
    "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": 870, "tid": 870,
    "ts": 7429770614777.874, "dur": 0.230,
    "args": {
      "External id": 2837,"Record function id": 0, "finished": true, "Ev Idx": 2836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770614778.679, "dur": 0.080,
    "args": {
      "External id": 2838,"Record function id": 0, "finished": true, "Ev Idx": 2837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770614788.068, "dur": 10.634,
    "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": 870, "tid": 870,
    "ts": 7429770614801.964, "dur": 9.408,
    "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": 870, "tid": 870,
    "ts": 7429770614812.998, "dur": 10.309,
    "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": 870, "tid": 870,
    "ts": 7429770614824.811, "dur": 7.444,
    "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": 870, "tid": 870,
    "ts": 7429770614834.768, "dur": 0.100,
    "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": 870, "tid": 870,
    "ts": 7429770614835.613, "dur": 8.421,
    "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": 870, "tid": 870,
    "ts": 7429770614861.403, "dur": 38.468,
    "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": 870, "tid": 870,
    "ts": 7429770614861.801, "dur": 4.219,
    "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": 870, "tid": 870,
    "ts": 7429770614863.291, "dur": 1.530,
    "args": {
      "External id": 2847,"Record function id": 0, "finished": true, "Ev Idx": 2846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770614864.098, "dur": 0.567,
    "args": {
      "External id": 2848,"Record function id": 0, "finished": true, "Ev Idx": 2847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770614866.402, "dur": 33.354,
    "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": 870, "tid": 870,
    "ts": 7429770614866.915, "dur": 2.637,
    "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": 870, "tid": 870,
    "ts": 7429770614867.552, "dur": 1.871,
    "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": 870, "tid": 870,
    "ts": 7429770614869.932, "dur": 27.594,
    "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": 870, "tid": 870,
    "ts": 7429770614880.612, "dur": 1.304,
    "args": {
      "External id": 2853,"Record function id": 0, "finished": true, "Ev Idx": 2852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770614898.327, "dur": 1.157,
    "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": 870, "tid": 870,
    "ts": 7429770614902.776, "dur": 1.472,
    "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": 870, "tid": 870,
    "ts": 7429770614905.914, "dur": 2.267,
    "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": 870, "tid": 870,
    "ts": 7429770614907.086, "dur": 0.472,
    "args": {
      "External id": 2857,"Record function id": 0, "finished": true, "Ev Idx": 2856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770614915.291, "dur": 31.883,
    "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": 870, "tid": 870,
    "ts": 7429770614915.946, "dur": 2.281,
    "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": 870, "tid": 870,
    "ts": 7429770614916.864, "dur": 0.940,
    "args": {
      "External id": 2860,"Record function id": 0, "finished": true, "Ev Idx": 2859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770614917.411, "dur": 0.281,
    "args": {
      "External id": 2861,"Record function id": 0, "finished": true, "Ev Idx": 2860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770614918.651, "dur": 28.408,
    "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": 870, "tid": 870,
    "ts": 7429770614919.046, "dur": 1.260,
    "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": 870, "tid": 870,
    "ts": 7429770614919.305, "dur": 0.899,
    "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": 870, "tid": 870,
    "ts": 7429770614920.630, "dur": 24.719,
    "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": 870, "tid": 870,
    "ts": 7429770614929.699, "dur": 1.314,
    "args": {
      "External id": 2866,"Record function id": 0, "finished": true, "Ev Idx": 2865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770614945.902, "dur": 0.926,
    "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": 870, "tid": 870,
    "ts": 7429770614949.099, "dur": 1.138,
    "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": 870, "tid": 870,
    "ts": 7429770614951.576, "dur": 1.988,
    "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": 870, "tid": 870,
    "ts": 7429770614952.511, "dur": 0.523,
    "args": {
      "External id": 2870,"Record function id": 0, "finished": true, "Ev Idx": 2869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770614960.732, "dur": 29.331,
    "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": 870, "tid": 870,
    "ts": 7429770614961.043, "dur": 1.951,
    "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": 870, "tid": 870,
    "ts": 7429770614961.886, "dur": 0.743,
    "args": {
      "External id": 2873,"Record function id": 0, "finished": true, "Ev Idx": 2872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770614962.284, "dur": 0.241,
    "args": {
      "External id": 2874,"Record function id": 0, "finished": true, "Ev Idx": 2873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770614963.271, "dur": 26.671,
    "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": 870, "tid": 870,
    "ts": 7429770614963.687, "dur": 1.295,
    "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": 870, "tid": 870,
    "ts": 7429770614964.012, "dur": 0.858,
    "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": 870, "tid": 870,
    "ts": 7429770614965.301, "dur": 22.888,
    "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": 870, "tid": 870,
    "ts": 7429770614974.017, "dur": 1.108,
    "args": {
      "External id": 2879,"Record function id": 0, "finished": true, "Ev Idx": 2878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770614988.704, "dur": 1.016,
    "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": 870, "tid": 870,
    "ts": 7429770614992.179, "dur": 1.410,
    "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": 870, "tid": 870,
    "ts": 7429770614995.022, "dur": 1.959,
    "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": 870, "tid": 870,
    "ts": 7429770614995.854, "dur": 0.548,
    "args": {
      "External id": 2883,"Record function id": 0, "finished": true, "Ev Idx": 2882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770614999.317, "dur": 1.640,
    "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": 870, "tid": 870,
    "ts": 7429770615000.166, "dur": 0.280,
    "args": {
      "External id": 2885,"Record function id": 0, "finished": true, "Ev Idx": 2884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770615001.975, "dur": 1.137,
    "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": 870, "tid": 870,
    "ts": 7429770615002.385, "dur": 0.309,
    "args": {
      "External id": 2887,"Record function id": 0, "finished": true, "Ev Idx": 2886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770615004.010, "dur": 9.332,
    "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": 870, "tid": 870,
    "ts": 7429770615017.915, "dur": 3.581,
    "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": 870, "tid": 870,
    "ts": 7429770615019.876, "dur": 0.434,
    "args": {
      "External id": 2890,"Record function id": 0, "finished": true, "Ev Idx": 2889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770615023.857, "dur": 1.857,
    "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": 870, "tid": 870,
    "ts": 7429770615024.844, "dur": 0.293,
    "args": {
      "External id": 2892,"Record function id": 0, "finished": true, "Ev Idx": 2891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770615027.526, "dur": 11.699,
    "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": 870, "tid": 870,
    "ts": 7429770615043.390, "dur": 14.540,
    "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": 870, "tid": 870,
    "ts": 7429770615060.867, "dur": 9.303,
    "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": 870, "tid": 870,
    "ts": 7429770615071.894, "dur": 9.139,
    "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": 870, "tid": 870,
    "ts": 7429770615083.012, "dur": 8.686,
    "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": 870, "tid": 870,
    "ts": 7429770615094.113, "dur": 2.095,
    "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": 870, "tid": 870,
    "ts": 7429770615095.076, "dur": 0.363,
    "args": {
      "External id": 2899,"Record function id": 0, "finished": true, "Ev Idx": 2898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770615097.755, "dur": 1.684,
    "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": 870, "tid": 870,
    "ts": 7429770615098.612, "dur": 0.295,
    "args": {
      "External id": 2901,"Record function id": 0, "finished": true, "Ev Idx": 2900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770615100.419, "dur": 9.147,
    "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": 870, "tid": 870,
    "ts": 7429770615111.575, "dur": 11.025,
    "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": 870, "tid": 870,
    "ts": 7429770615124.829, "dur": 7.403,
    "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": 870, "tid": 870,
    "ts": 7429770615133.677, "dur": 6.088,
    "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": 870, "tid": 870,
    "ts": 7429770615161.180, "dur": 2.447,
    "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": 870, "tid": 870,
    "ts": 7429770615162.291, "dur": 0.580,
    "args": {
      "External id": 2907,"Record function id": 0, "finished": true, "Ev Idx": 2906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770615164.241, "dur": 1.636,
    "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": 870, "tid": 870,
    "ts": 7429770615165.027, "dur": 0.325,
    "args": {
      "External id": 2909,"Record function id": 0, "finished": true, "Ev Idx": 2908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770615166.644, "dur": 2.649,
    "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": 870, "tid": 870,
    "ts": 7429770615167.544, "dur": 1.050,
    "args": {
      "External id": 2911,"Record function id": 0, "finished": true, "Ev Idx": 2910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770615169.919, "dur": 3.266,
    "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": 870, "tid": 870,
    "ts": 7429770615172.094, "dur": 0.560,
    "args": {
      "External id": 2913,"Record function id": 0, "finished": true, "Ev Idx": 2912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770615173.772, "dur": 2.435,
    "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": 870, "tid": 870,
    "ts": 7429770615175.295, "dur": 0.430,
    "args": {
      "External id": 2915,"Record function id": 0, "finished": true, "Ev Idx": 2914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770615178.856, "dur": 3.047,
    "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": 870, "tid": 870,
    "ts": 7429770615180.811, "dur": 0.373,
    "args": {
      "External id": 2917,"Record function id": 0, "finished": true, "Ev Idx": 2916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770615184.319, "dur": 17.419,
    "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": 870, "tid": 870,
    "ts": 7429770615186.341, "dur": 13.169,
    "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": 870, "tid": 870,
    "ts": 7429770615187.992, "dur": 2.627,
    "args": {
      "External id": 2920,"Record function id": 0, "finished": true, "Ev Idx": 2919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770615188.888, "dur": 1.519,
    "args": {
      "External id": 2921,"Record function id": 0, "finished": true, "Ev Idx": 2920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770615191.247, "dur": 7.707,
    "args": {
      "External id": 2922,"Record function id": 0, "finished": true, "Ev Idx": 2921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770615200.073, "dur": 1.406,
    "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": 870, "tid": 870,
    "ts": 7429770615204.845, "dur": 2.470,
    "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": 870, "tid": 870,
    "ts": 7429770615206.336, "dur": 0.387,
    "args": {
      "External id": 2925,"Record function id": 0, "finished": true, "Ev Idx": 2924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770615207.963, "dur": 1.863,
    "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": 870, "tid": 870,
    "ts": 7429770615209.046, "dur": 0.288,
    "args": {
      "External id": 2927,"Record function id": 0, "finished": true, "Ev Idx": 2926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770615210.353, "dur": 1.391,
    "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": 870, "tid": 870,
    "ts": 7429770615211.033, "dur": 0.306,
    "args": {
      "External id": 2929,"Record function id": 0, "finished": true, "Ev Idx": 2928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770615212.312, "dur": 1.482,
    "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": 870, "tid": 870,
    "ts": 7429770615212.978, "dur": 0.321,
    "args": {
      "External id": 2931,"Record function id": 0, "finished": true, "Ev Idx": 2930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770615214.369, "dur": 1.456,
    "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": 870, "tid": 870,
    "ts": 7429770615215.082, "dur": 0.283,
    "args": {
      "External id": 2933,"Record function id": 0, "finished": true, "Ev Idx": 2932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770615217.341, "dur": 2.007,
    "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": 870, "tid": 870,
    "ts": 7429770615218.552, "dur": 0.323,
    "args": {
      "External id": 2935,"Record function id": 0, "finished": true, "Ev Idx": 2934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770615220.791, "dur": 11.550,
    "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": 870, "tid": 870,
    "ts": 7429770615221.779, "dur": 9.107,
    "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": 870, "tid": 870,
    "ts": 7429770615223.343, "dur": 1.637,
    "args": {
      "External id": 2938,"Record function id": 0, "finished": true, "Ev Idx": 2937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770615223.674, "dur": 1.195,
    "args": {
      "External id": 2939,"Record function id": 0, "finished": true, "Ev Idx": 2938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770615225.331, "dur": 5.241,
    "args": {
      "External id": 2940,"Record function id": 0, "finished": true, "Ev Idx": 2939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770615231.275, "dur": 0.898,
    "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": 870, "tid": 870,
    "ts": 7429770615234.476, "dur": 1.716,
    "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": 870, "tid": 870,
    "ts": 7429770615235.311, "dur": 0.360,
    "args": {
      "External id": 2943,"Record function id": 0, "finished": true, "Ev Idx": 2942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770615237.809, "dur": 44.348,
    "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": 870, "tid": 870,
    "ts": 7429770615238.355, "dur": 1.602,
    "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": 870, "tid": 870,
    "ts": 7429770615239.151, "dur": 0.345,
    "args": {
      "External id": 2946,"Record function id": 0, "finished": true, "Ev Idx": 2945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770615240.305, "dur": 3.619,
    "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": 870, "tid": 870,
    "ts": 7429770615241.759, "dur": 2.008,
    "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": 870, "tid": 870,
    "ts": 7429770615244.505, "dur": 2.260,
    "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": 870, "tid": 870,
    "ts": 7429770615245.960, "dur": 0.439,
    "args": {
      "External id": 2950,"Record function id": 0, "finished": true, "Ev Idx": 2949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770615247.034, "dur": 2.499,
    "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": 870, "tid": 870,
    "ts": 7429770615248.031, "dur": 1.405,
    "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": 870, "tid": 870,
    "ts": 7429770615250.144, "dur": 29.810,
    "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": 870, "tid": 870,
    "ts": 7429770615261.648, "dur": 1.241,
    "args": {
      "External id": 2954,"Record function id": 0, "finished": true, "Ev Idx": 2953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770615280.580, "dur": 1.222,
    "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": 870, "tid": 870,
    "ts": 7429770615285.143, "dur": 9.235,
    "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": 870, "tid": 870,
    "ts": 7429770615297.556, "dur": 1.769,
    "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": 870, "tid": 870,
    "ts": 7429770615298.334, "dur": 0.361,
    "args": {
      "External id": 2958,"Record function id": 0, "finished": true, "Ev Idx": 2957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770615299.933, "dur": 1.259,
    "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": 870, "tid": 870,
    "ts": 7429770615300.505, "dur": 0.286,
    "args": {
      "External id": 2960,"Record function id": 0, "finished": true, "Ev Idx": 2959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770615301.806, "dur": 0.984,
    "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": 870, "tid": 870,
    "ts": 7429770615302.219, "dur": 0.237,
    "args": {
      "External id": 2962,"Record function id": 0, "finished": true, "Ev Idx": 2961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770615304.265, "dur": 7.232,
    "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": 870, "tid": 870,
    "ts": 7429770615316.394, "dur": 12.476,
    "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": 870, "tid": 870,
    "ts": 7429770615317.406, "dur": 0.329,
    "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": 870, "tid": 870,
    "ts": 7429770615318.646, "dur": 9.995,
    "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": 870, "tid": 870,
    "ts": 7429770615331.470, "dur": 0.174,
    "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": 870, "tid": 870,
    "ts": 7429770615341.496, "dur": 0.356,
    "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": 870, "tid": 870,
    "ts": 7429770615343.808, "dur": 39.691,
    "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": 870, "tid": 870,
    "ts": 7429770615344.462, "dur": 2.505,
    "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": 870, "tid": 870,
    "ts": 7429770615346.014, "dur": 0.450,
    "args": {
      "External id": 2971,"Record function id": 0, "finished": true, "Ev Idx": 2970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770615347.325, "dur": 3.386,
    "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": 870, "tid": 870,
    "ts": 7429770615347.625, "dur": 2.954,
    "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": 870, "tid": 870,
    "ts": 7429770615351.427, "dur": 2.186,
    "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": 870, "tid": 870,
    "ts": 7429770615352.993, "dur": 0.262,
    "args": {
      "External id": 2975,"Record function id": 0, "finished": true, "Ev Idx": 2974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770615353.904, "dur": 1.633,
    "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": 870, "tid": 870,
    "ts": 7429770615354.436, "dur": 1.027,
    "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": 870, "tid": 870,
    "ts": 7429770615355.903, "dur": 25.614,
    "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": 870, "tid": 870,
    "ts": 7429770615366.147, "dur": 1.102,
    "args": {
      "External id": 2979,"Record function id": 0, "finished": true, "Ev Idx": 2978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770615382.067, "dur": 1.142,
    "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": 870, "tid": 870,
    "ts": 7429770615385.476, "dur": 1.913,
    "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": 870, "tid": 870,
    "ts": 7429770615386.537, "dur": 0.310,
    "args": {
      "External id": 2982,"Record function id": 0, "finished": true, "Ev Idx": 2981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770615389.377, "dur": 11.783,
    "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": 870, "tid": 870,
    "ts": 7429770615389.874, "dur": 11.141,
    "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": 870, "tid": 870,
    "ts": 7429770615390.875, "dur": 1.905,
    "args": {
      "External id": 2985,"Record function id": 0, "finished": true, "Ev Idx": 2984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770615391.233, "dur": 1.435,
    "args": {
      "External id": 2986,"Record function id": 0, "finished": true, "Ev Idx": 2985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770615393.116, "dur": 7.520,
    "args": {
      "External id": 2987,"Record function id": 0, "finished": true, "Ev Idx": 2986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770615406.541, "dur": 1.601,
    "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": 870, "tid": 870,
    "ts": 7429770615406.860, "dur": 1.177,
    "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": 870, "tid": 870,
    "ts": 7429770615421.780, "dur": 35.551,
    "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": 870, "tid": 870,
    "ts": 7429770615422.243, "dur": 2.734,
    "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": 870, "tid": 870,
    "ts": 7429770615423.483, "dur": 0.977,
    "args": {
      "External id": 2992,"Record function id": 0, "finished": true, "Ev Idx": 2991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770615423.925, "dur": 0.372,
    "args": {
      "External id": 2993,"Record function id": 0, "finished": true, "Ev Idx": 2992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770615425.343, "dur": 31.862,
    "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": 870, "tid": 870,
    "ts": 7429770615425.793, "dur": 1.408,
    "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": 870, "tid": 870,
    "ts": 7429770615426.088, "dur": 1.013,
    "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": 870, "tid": 870,
    "ts": 7429770615427.556, "dur": 27.720,
    "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": 870, "tid": 870,
    "ts": 7429770615439.250, "dur": 1.217,
    "args": {
      "External id": 2998,"Record function id": 0, "finished": true, "Ev Idx": 2997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770615455.886, "dur": 1.033,
    "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": 870, "tid": 870,
    "ts": 7429770615462.017, "dur": 8.441,
    "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": 870, "tid": 870,
    "ts": 7429770615480.301, "dur": 0.153,
    "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": 870, "tid": 870,
    "ts": 7429770615482.265, "dur": 11.492,
    "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": 870, "tid": 870,
    "ts": 7429770615483.911, "dur": 0.329,
    "args": {
      "External id": 3003,"Record function id": 0, "finished": true, "Ev Idx": 3002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770615484.908, "dur": 0.088,
    "args": {
      "External id": 3004,"Record function id": 0, "finished": true, "Ev Idx": 3003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770615496.071, "dur": 11.790,
    "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": 870, "tid": 870,
    "ts": 7429770615511.019, "dur": 8.071,
    "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": 870, "tid": 870,
    "ts": 7429770615520.878, "dur": 8.248,
    "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": 870, "tid": 870,
    "ts": 7429770615530.745, "dur": 9.914,
    "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": 870, "tid": 870,
    "ts": 7429770615543.061, "dur": 0.233,
    "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": 870, "tid": 870,
    "ts": 7429770615544.365, "dur": 10.409,
    "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": 870, "tid": 870,
    "ts": 7429770615569.188, "dur": 38.581,
    "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": 870, "tid": 870,
    "ts": 7429770615569.728, "dur": 3.115,
    "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": 870, "tid": 870,
    "ts": 7429770615570.653, "dur": 1.546,
    "args": {
      "External id": 3013,"Record function id": 0, "finished": true, "Ev Idx": 3012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770615571.507, "dur": 0.532,
    "args": {
      "External id": 3014,"Record function id": 0, "finished": true, "Ev Idx": 3013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770615573.218, "dur": 34.431,
    "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": 870, "tid": 870,
    "ts": 7429770615573.851, "dur": 1.806,
    "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": 870, "tid": 870,
    "ts": 7429770615574.327, "dur": 1.200,
    "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": 870, "tid": 870,
    "ts": 7429770615576.007, "dur": 29.712,
    "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": 870, "tid": 870,
    "ts": 7429770615589.648, "dur": 1.266,
    "args": {
      "External id": 3019,"Record function id": 0, "finished": true, "Ev Idx": 3018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770615606.335, "dur": 0.989,
    "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": 870, "tid": 870,
    "ts": 7429770615615.569, "dur": 9.790,
    "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": 870, "tid": 870,
    "ts": 7429770615633.898, "dur": 114.757,
    "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": 870, "tid": 870,
    "ts": 7429770615634.279, "dur": 2.445,
    "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": 870, "tid": 870,
    "ts": 7429770615635.263, "dur": 0.979,
    "args": {
      "External id": 3024,"Record function id": 0, "finished": true, "Ev Idx": 3023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770615635.679, "dur": 0.436,
    "args": {
      "External id": 3025,"Record function id": 0, "finished": true, "Ev Idx": 3024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770615637.135, "dur": 111.373,
    "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": 870, "tid": 870,
    "ts": 7429770615637.620, "dur": 2.394,
    "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": 870, "tid": 870,
    "ts": 7429770615638.098, "dur": 1.811,
    "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": 870, "tid": 870,
    "ts": 7429770615640.474, "dur": 105.336,
    "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": 870, "tid": 870,
    "ts": 7429770615649.912, "dur": 77.151,
    "args": {
      "External id": 3030,"Record function id": 0, "finished": true, "Ev Idx": 3029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770615746.789, "dur": 1.301,
    "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": 870, "tid": 870,
    "ts": 7429770615751.002, "dur": 9.407,
    "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": 870, "tid": 870,
    "ts": 7429770615769.448, "dur": 49.324,
    "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": 870, "tid": 870,
    "ts": 7429770615770.145, "dur": 3.254,
    "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": 870, "tid": 870,
    "ts": 7429770615771.337, "dur": 1.315,
    "args": {
      "External id": 3035,"Record function id": 0, "finished": true, "Ev Idx": 3034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770615771.959, "dur": 0.540,
    "args": {
      "External id": 3036,"Record function id": 0, "finished": true, "Ev Idx": 3035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770615773.761, "dur": 44.889,
    "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": 870, "tid": 870,
    "ts": 7429770615774.209, "dur": 1.710,
    "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": 870, "tid": 870,
    "ts": 7429770615774.559, "dur": 1.228,
    "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": 870, "tid": 870,
    "ts": 7429770615776.279, "dur": 40.423,
    "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": 870, "tid": 870,
    "ts": 7429770615788.229, "dur": 1.279,
    "args": {
      "External id": 3041,"Record function id": 0, "finished": true, "Ev Idx": 3040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770615817.368, "dur": 0.998,
    "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": 870, "tid": 870,
    "ts": 7429770615821.593, "dur": 9.433,
    "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": 870, "tid": 870,
    "ts": 7429770615851.648, "dur": 0.177,
    "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": 870, "tid": 870,
    "ts": 7429770615853.650, "dur": 9.844,
    "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": 870, "tid": 870,
    "ts": 7429770615855.119, "dur": 0.420,
    "args": {
      "External id": 3046,"Record function id": 0, "finished": true, "Ev Idx": 3045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770615855.857, "dur": 0.249,
    "args": {
      "External id": 3047,"Record function id": 0, "finished": true, "Ev Idx": 3046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770615865.922, "dur": 11.170,
    "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": 870, "tid": 870,
    "ts": 7429770615880.998, "dur": 8.434,
    "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": 870, "tid": 870,
    "ts": 7429770615891.121, "dur": 11.113,
    "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": 870, "tid": 870,
    "ts": 7429770615904.077, "dur": 7.855,
    "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": 870, "tid": 870,
    "ts": 7429770615914.645, "dur": 0.104,
    "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": 870, "tid": 870,
    "ts": 7429770615915.527, "dur": 8.956,
    "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": 870, "tid": 870,
    "ts": 7429770615943.082, "dur": 38.556,
    "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": 870, "tid": 870,
    "ts": 7429770615943.612, "dur": 4.122,
    "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": 870, "tid": 870,
    "ts": 7429770615944.903, "dur": 1.592,
    "args": {
      "External id": 3056,"Record function id": 0, "finished": true, "Ev Idx": 3055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770615945.822, "dur": 0.527,
    "args": {
      "External id": 3057,"Record function id": 0, "finished": true, "Ev Idx": 3056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770615948.101, "dur": 33.417,
    "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": 870, "tid": 870,
    "ts": 7429770615948.774, "dur": 2.580,
    "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": 870, "tid": 870,
    "ts": 7429770615949.154, "dur": 2.071,
    "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": 870, "tid": 870,
    "ts": 7429770615951.698, "dur": 28.039,
    "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": 870, "tid": 870,
    "ts": 7429770615962.804, "dur": 1.451,
    "args": {
      "External id": 3062,"Record function id": 0, "finished": true, "Ev Idx": 3061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770615980.276, "dur": 0.971,
    "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": 870, "tid": 870,
    "ts": 7429770615984.733, "dur": 1.287,
    "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": 870, "tid": 870,
    "ts": 7429770615987.809, "dur": 2.612,
    "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": 870, "tid": 870,
    "ts": 7429770615988.952, "dur": 0.801,
    "args": {
      "External id": 3066,"Record function id": 0, "finished": true, "Ev Idx": 3065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770615997.491, "dur": 61.545,
    "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": 870, "tid": 870,
    "ts": 7429770615997.967, "dur": 2.695,
    "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": 870, "tid": 870,
    "ts": 7429770615999.114, "dur": 1.129,
    "args": {
      "External id": 3069,"Record function id": 0, "finished": true, "Ev Idx": 3068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770615999.667, "dur": 0.455,
    "args": {
      "External id": 3070,"Record function id": 0, "finished": true, "Ev Idx": 3069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770616001.115, "dur": 57.785,
    "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": 870, "tid": 870,
    "ts": 7429770616001.689, "dur": 29.093,
    "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": 870, "tid": 870,
    "ts": 7429770616029.662, "dur": 0.988,
    "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": 870, "tid": 870,
    "ts": 7429770616031.157, "dur": 26.022,
    "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": 870, "tid": 870,
    "ts": 7429770616041.076, "dur": 1.113,
    "args": {
      "External id": 3075,"Record function id": 0, "finished": true, "Ev Idx": 3074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770616057.682, "dur": 0.920,
    "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": 870, "tid": 870,
    "ts": 7429770616061.255, "dur": 1.158,
    "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": 870, "tid": 870,
    "ts": 7429770616063.631, "dur": 1.567,
    "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": 870, "tid": 870,
    "ts": 7429770616064.331, "dur": 0.369,
    "args": {
      "External id": 3079,"Record function id": 0, "finished": true, "Ev Idx": 3078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770616072.232, "dur": 30.156,
    "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": 870, "tid": 870,
    "ts": 7429770616072.526, "dur": 1.917,
    "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": 870, "tid": 870,
    "ts": 7429770616073.316, "dur": 0.743,
    "args": {
      "External id": 3082,"Record function id": 0, "finished": true, "Ev Idx": 3081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770616073.672, "dur": 0.268,
    "args": {
      "External id": 3083,"Record function id": 0, "finished": true, "Ev Idx": 3082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770616074.754, "dur": 27.557,
    "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": 870, "tid": 870,
    "ts": 7429770616075.179, "dur": 1.206,
    "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": 870, "tid": 870,
    "ts": 7429770616075.443, "dur": 0.831,
    "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": 870, "tid": 870,
    "ts": 7429770616076.708, "dur": 23.970,
    "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": 870, "tid": 870,
    "ts": 7429770616085.469, "dur": 1.044,
    "args": {
      "External id": 3088,"Record function id": 0, "finished": true, "Ev Idx": 3087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770616101.183, "dur": 0.903,
    "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": 870, "tid": 870,
    "ts": 7429770616104.643, "dur": 1.120,
    "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": 870, "tid": 870,
    "ts": 7429770616107.394, "dur": 1.552,
    "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": 870, "tid": 870,
    "ts": 7429770616108.135, "dur": 0.332,
    "args": {
      "External id": 3092,"Record function id": 0, "finished": true, "Ev Idx": 3091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770616111.432, "dur": 2.011,
    "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": 870, "tid": 870,
    "ts": 7429770616112.606, "dur": 0.319,
    "args": {
      "External id": 3094,"Record function id": 0, "finished": true, "Ev Idx": 3093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770616114.479, "dur": 0.986,
    "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": 870, "tid": 870,
    "ts": 7429770616114.867, "dur": 0.283,
    "args": {
      "External id": 3096,"Record function id": 0, "finished": true, "Ev Idx": 3095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770616116.234, "dur": 9.380,
    "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": 870, "tid": 870,
    "ts": 7429770616130.851, "dur": 4.105,
    "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": 870, "tid": 870,
    "ts": 7429770616133.028, "dur": 0.607,
    "args": {
      "External id": 3099,"Record function id": 0, "finished": true, "Ev Idx": 3098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770616137.163, "dur": 3.620,
    "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": 870, "tid": 870,
    "ts": 7429770616138.921, "dur": 0.669,
    "args": {
      "External id": 3101,"Record function id": 0, "finished": true, "Ev Idx": 3100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770616142.611, "dur": 9.944,
    "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": 870, "tid": 870,
    "ts": 7429770616156.183, "dur": 17.022,
    "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": 870, "tid": 870,
    "ts": 7429770616175.854, "dur": 7.412,
    "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": 870, "tid": 870,
    "ts": 7429770616184.784, "dur": 9.527,
    "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": 870, "tid": 870,
    "ts": 7429770616196.297, "dur": 7.542,
    "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": 870, "tid": 870,
    "ts": 7429770616206.135, "dur": 2.070,
    "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": 870, "tid": 870,
    "ts": 7429770616207.153, "dur": 0.412,
    "args": {
      "External id": 3108,"Record function id": 0, "finished": true, "Ev Idx": 3107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770616209.741, "dur": 1.623,
    "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": 870, "tid": 870,
    "ts": 7429770616210.511, "dur": 0.339,
    "args": {
      "External id": 3110,"Record function id": 0, "finished": true, "Ev Idx": 3109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770616212.272, "dur": 9.787,
    "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": 870, "tid": 870,
    "ts": 7429770616224.016, "dur": 16.505,
    "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": 870, "tid": 870,
    "ts": 7429770616242.682, "dur": 6.801,
    "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": 870, "tid": 870,
    "ts": 7429770616250.760, "dur": 6.051,
    "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": 870, "tid": 870,
    "ts": 7429770616279.547, "dur": 2.516,
    "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": 870, "tid": 870,
    "ts": 7429770616280.782, "dur": 0.446,
    "args": {
      "External id": 3116,"Record function id": 0, "finished": true, "Ev Idx": 3115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770616282.722, "dur": 1.688,
    "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": 870, "tid": 870,
    "ts": 7429770616283.546, "dur": 0.359,
    "args": {
      "External id": 3118,"Record function id": 0, "finished": true, "Ev Idx": 3117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770616285.017, "dur": 2.095,
    "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": 870, "tid": 870,
    "ts": 7429770616285.926, "dur": 0.619,
    "args": {
      "External id": 3120,"Record function id": 0, "finished": true, "Ev Idx": 3119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770616287.726, "dur": 1.657,
    "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": 870, "tid": 870,
    "ts": 7429770616288.489, "dur": 0.385,
    "args": {
      "External id": 3122,"Record function id": 0, "finished": true, "Ev Idx": 3121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770616289.956, "dur": 1.459,
    "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": 870, "tid": 870,
    "ts": 7429770616290.640, "dur": 0.311,
    "args": {
      "External id": 3124,"Record function id": 0, "finished": true, "Ev Idx": 3123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770616294.455, "dur": 2.970,
    "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": 870, "tid": 870,
    "ts": 7429770616296.332, "dur": 0.392,
    "args": {
      "External id": 3126,"Record function id": 0, "finished": true, "Ev Idx": 3125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770616300.116, "dur": 19.056,
    "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": 870, "tid": 870,
    "ts": 7429770616302.568, "dur": 14.745,
    "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": 870, "tid": 870,
    "ts": 7429770616303.923, "dur": 2.728,
    "args": {
      "External id": 3129,"Record function id": 0, "finished": true, "Ev Idx": 3128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770616304.832, "dur": 1.594,
    "args": {
      "External id": 3130,"Record function id": 0, "finished": true, "Ev Idx": 3129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770616307.320, "dur": 9.402,
    "args": {
      "External id": 3131,"Record function id": 0, "finished": true, "Ev Idx": 3130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770616317.792, "dur": 1.120,
    "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": 870, "tid": 870,
    "ts": 7429770616322.575, "dur": 1.780,
    "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": 870, "tid": 870,
    "ts": 7429770616323.461, "dur": 0.326,
    "args": {
      "External id": 3134,"Record function id": 0, "finished": true, "Ev Idx": 3133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770616325.007, "dur": 1.496,
    "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": 870, "tid": 870,
    "ts": 7429770616325.733, "dur": 0.279,
    "args": {
      "External id": 3136,"Record function id": 0, "finished": true, "Ev Idx": 3135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770616327.043, "dur": 1.448,
    "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": 870, "tid": 870,
    "ts": 7429770616327.775, "dur": 0.299,
    "args": {
      "External id": 3138,"Record function id": 0, "finished": true, "Ev Idx": 3137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770616329.070, "dur": 1.493,
    "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": 870, "tid": 870,
    "ts": 7429770616329.754, "dur": 0.305,
    "args": {
      "External id": 3140,"Record function id": 0, "finished": true, "Ev Idx": 3139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770616331.138, "dur": 1.410,
    "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": 870, "tid": 870,
    "ts": 7429770616331.778, "dur": 0.291,
    "args": {
      "External id": 3142,"Record function id": 0, "finished": true, "Ev Idx": 3141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770616334.046, "dur": 1.572,
    "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": 870, "tid": 870,
    "ts": 7429770616334.849, "dur": 0.329,
    "args": {
      "External id": 3144,"Record function id": 0, "finished": true, "Ev Idx": 3143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770616337.301, "dur": 11.659,
    "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": 870, "tid": 870,
    "ts": 7429770616338.107, "dur": 9.330,
    "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": 870, "tid": 870,
    "ts": 7429770616338.777, "dur": 1.655,
    "args": {
      "External id": 3147,"Record function id": 0, "finished": true, "Ev Idx": 3146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770616339.111, "dur": 1.202,
    "args": {
      "External id": 3148,"Record function id": 0, "finished": true, "Ev Idx": 3147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770616340.801, "dur": 6.306,
    "args": {
      "External id": 3149,"Record function id": 0, "finished": true, "Ev Idx": 3148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770616347.837, "dur": 0.966,
    "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": 870, "tid": 870,
    "ts": 7429770616350.971, "dur": 1.732,
    "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": 870, "tid": 870,
    "ts": 7429770616351.817, "dur": 0.339,
    "args": {
      "External id": 3152,"Record function id": 0, "finished": true, "Ev Idx": 3151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770616354.250, "dur": 108.458,
    "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": 870, "tid": 870,
    "ts": 7429770616354.910, "dur": 1.482,
    "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": 870, "tid": 870,
    "ts": 7429770616355.651, "dur": 0.306,
    "args": {
      "External id": 3155,"Record function id": 0, "finished": true, "Ev Idx": 3154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770616356.708, "dur": 3.855,
    "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": 870, "tid": 870,
    "ts": 7429770616358.538, "dur": 1.855,
    "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": 870, "tid": 870,
    "ts": 7429770616360.992, "dur": 1.327,
    "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": 870, "tid": 870,
    "ts": 7429770616361.670, "dur": 0.288,
    "args": {
      "External id": 3159,"Record function id": 0, "finished": true, "Ev Idx": 3158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770616362.604, "dur": 1.752,
    "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": 870, "tid": 870,
    "ts": 7429770616363.402, "dur": 0.844,
    "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": 870, "tid": 870,
    "ts": 7429770616364.909, "dur": 95.285,
    "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": 870, "tid": 870,
    "ts": 7429770616437.435, "dur": 1.360,
    "args": {
      "External id": 3163,"Record function id": 0, "finished": true, "Ev Idx": 3162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770616460.987, "dur": 1.334,
    "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": 870, "tid": 870,
    "ts": 7429770616466.884, "dur": 10.757,
    "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": 870, "tid": 870,
    "ts": 7429770616481.001, "dur": 2.390,
    "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": 870, "tid": 870,
    "ts": 7429770616481.961, "dur": 0.575,
    "args": {
      "External id": 3167,"Record function id": 0, "finished": true, "Ev Idx": 3166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770616484.078, "dur": 1.603,
    "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": 870, "tid": 870,
    "ts": 7429770616484.878, "dur": 0.330,
    "args": {
      "External id": 3169,"Record function id": 0, "finished": true, "Ev Idx": 3168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770616486.335, "dur": 1.326,
    "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": 870, "tid": 870,
    "ts": 7429770616486.941, "dur": 0.395,
    "args": {
      "External id": 3171,"Record function id": 0, "finished": true, "Ev Idx": 3170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770616489.320, "dur": 7.873,
    "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": 870, "tid": 870,
    "ts": 7429770616503.019, "dur": 14.481,
    "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": 870, "tid": 870,
    "ts": 7429770616504.142, "dur": 0.347,
    "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": 870, "tid": 870,
    "ts": 7429770616505.152, "dur": 12.094,
    "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": 870, "tid": 870,
    "ts": 7429770616520.280, "dur": 0.251,
    "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": 870, "tid": 870,
    "ts": 7429770616531.944, "dur": 0.531,
    "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": 870, "tid": 870,
    "ts": 7429770616534.466, "dur": 41.876,
    "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": 870, "tid": 870,
    "ts": 7429770616535.433, "dur": 2.023,
    "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": 870, "tid": 870,
    "ts": 7429770616536.597, "dur": 0.343,
    "args": {
      "External id": 3180,"Record function id": 0, "finished": true, "Ev Idx": 3179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770616537.832, "dur": 2.870,
    "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": 870, "tid": 870,
    "ts": 7429770616538.382, "dur": 2.185,
    "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": 870, "tid": 870,
    "ts": 7429770616541.146, "dur": 2.305,
    "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": 870, "tid": 870,
    "ts": 7429770616542.830, "dur": 0.238,
    "args": {
      "External id": 3184,"Record function id": 0, "finished": true, "Ev Idx": 3183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770616543.764, "dur": 1.772,
    "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": 870, "tid": 870,
    "ts": 7429770616544.011, "dur": 1.446,
    "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": 870, "tid": 870,
    "ts": 7429770616545.996, "dur": 28.232,
    "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": 870, "tid": 870,
    "ts": 7429770616556.909, "dur": 1.292,
    "args": {
      "External id": 3188,"Record function id": 0, "finished": true, "Ev Idx": 3187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770616574.827, "dur": 1.202,
    "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": 870, "tid": 870,
    "ts": 7429770616578.387, "dur": 1.861,
    "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": 870, "tid": 870,
    "ts": 7429770616579.298, "dur": 0.327,
    "args": {
      "External id": 3191,"Record function id": 0, "finished": true, "Ev Idx": 3190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770616582.819, "dur": 12.813,
    "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": 870, "tid": 870,
    "ts": 7429770616583.383, "dur": 12.103,
    "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": 870, "tid": 870,
    "ts": 7429770616584.515, "dur": 2.134,
    "args": {
      "External id": 3194,"Record function id": 0, "finished": true, "Ev Idx": 3193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770616585.162, "dur": 1.340,
    "args": {
      "External id": 3195,"Record function id": 0, "finished": true, "Ev Idx": 3194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770616587.060, "dur": 7.946,
    "args": {
      "External id": 3196,"Record function id": 0, "finished": true, "Ev Idx": 3195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770616601.590, "dur": 1.796,
    "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": 870, "tid": 870,
    "ts": 7429770616601.999, "dur": 1.268,
    "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": 870, "tid": 870,
    "ts": 7429770616619.014, "dur": 38.961,
    "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": 870, "tid": 870,
    "ts": 7429770616619.919, "dur": 2.917,
    "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": 870, "tid": 870,
    "ts": 7429770616621.378, "dur": 0.929,
    "args": {
      "External id": 3201,"Record function id": 0, "finished": true, "Ev Idx": 3200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770616621.830, "dur": 0.361,
    "args": {
      "External id": 3202,"Record function id": 0, "finished": true, "Ev Idx": 3201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770616623.233, "dur": 34.542,
    "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": 870, "tid": 870,
    "ts": 7429770616623.686, "dur": 1.411,
    "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": 870, "tid": 870,
    "ts": 7429770616623.947, "dur": 1.061,
    "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": 870, "tid": 870,
    "ts": 7429770616625.667, "dur": 30.084,
    "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": 870, "tid": 870,
    "ts": 7429770616638.128, "dur": 1.049,
    "args": {
      "External id": 3207,"Record function id": 0, "finished": true, "Ev Idx": 3206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770616656.396, "dur": 1.062,
    "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": 870, "tid": 870,
    "ts": 7429770616664.033, "dur": 9.389,
    "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": 870, "tid": 870,
    "ts": 7429770616683.609, "dur": 0.214,
    "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": 870, "tid": 870,
    "ts": 7429770616686.207, "dur": 11.800,
    "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": 870, "tid": 870,
    "ts": 7429770616688.203, "dur": 0.419,
    "args": {
      "External id": 3212,"Record function id": 0, "finished": true, "Ev Idx": 3211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770616689.003, "dur": 0.274,
    "args": {
      "External id": 3213,"Record function id": 0, "finished": true, "Ev Idx": 3212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770616700.838, "dur": 14.074,
    "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": 870, "tid": 870,
    "ts": 7429770616718.399, "dur": 8.751,
    "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": 870, "tid": 870,
    "ts": 7429770616728.964, "dur": 10.215,
    "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": 870, "tid": 870,
    "ts": 7429770616740.890, "dur": 10.009,
    "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": 870, "tid": 870,
    "ts": 7429770616753.235, "dur": 0.273,
    "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": 870, "tid": 870,
    "ts": 7429770616754.671, "dur": 10.261,
    "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": 870, "tid": 870,
    "ts": 7429770616781.729, "dur": 37.490,
    "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": 870, "tid": 870,
    "ts": 7429770616782.318, "dur": 3.266,
    "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": 870, "tid": 870,
    "ts": 7429770616783.690, "dur": 1.246,
    "args": {
      "External id": 3222,"Record function id": 0, "finished": true, "Ev Idx": 3221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770616784.347, "dur": 0.422,
    "args": {
      "External id": 3223,"Record function id": 0, "finished": true, "Ev Idx": 3222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770616786.078, "dur": 33.011,
    "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": 870, "tid": 870,
    "ts": 7429770616786.719, "dur": 1.751,
    "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": 870, "tid": 870,
    "ts": 7429770616787.269, "dur": 1.083,
    "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": 870, "tid": 870,
    "ts": 7429770616788.868, "dur": 28.391,
    "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": 870, "tid": 870,
    "ts": 7429770616800.876, "dur": 1.213,
    "args": {
      "External id": 3228,"Record function id": 0, "finished": true, "Ev Idx": 3227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770616817.894, "dur": 0.930,
    "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": 870, "tid": 870,
    "ts": 7429770616827.574, "dur": 12.202,
    "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": 870, "tid": 870,
    "ts": 7429770616848.110, "dur": 33.049,
    "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": 870, "tid": 870,
    "ts": 7429770616848.435, "dur": 3.372,
    "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": 870, "tid": 870,
    "ts": 7429770616849.577, "dur": 1.175,
    "args": {
      "External id": 3233,"Record function id": 0, "finished": true, "Ev Idx": 3232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770616850.106, "dur": 0.528,
    "args": {
      "External id": 3234,"Record function id": 0, "finished": true, "Ev Idx": 3233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770616852.138, "dur": 28.901,
    "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": 870, "tid": 870,
    "ts": 7429770616852.642, "dur": 1.935,
    "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": 870, "tid": 870,
    "ts": 7429770616852.932, "dur": 1.528,
    "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": 870, "tid": 870,
    "ts": 7429770616854.906, "dur": 24.440,
    "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": 870, "tid": 870,
    "ts": 7429770616863.848, "dur": 1.273,
    "args": {
      "External id": 3239,"Record function id": 0, "finished": true, "Ev Idx": 3238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770616879.873, "dur": 0.911,
    "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": 870, "tid": 870,
    "ts": 7429770616883.084, "dur": 8.088,
    "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": 870, "tid": 870,
    "ts": 7429770616898.025, "dur": 47.179,
    "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": 870, "tid": 870,
    "ts": 7429770616898.496, "dur": 2.175,
    "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": 870, "tid": 870,
    "ts": 7429770616899.392, "dur": 0.793,
    "args": {
      "External id": 3244,"Record function id": 0, "finished": true, "Ev Idx": 3243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770616899.768, "dur": 0.309,
    "args": {
      "External id": 3245,"Record function id": 0, "finished": true, "Ev Idx": 3244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770616900.941, "dur": 44.142,
    "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": 870, "tid": 870,
    "ts": 7429770616901.347, "dur": 1.375,
    "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": 870, "tid": 870,
    "ts": 7429770616901.615, "dur": 1.010,
    "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": 870, "tid": 870,
    "ts": 7429770616903.029, "dur": 40.254,
    "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": 870, "tid": 870,
    "ts": 7429770616912.954, "dur": 1.052,
    "args": {
      "External id": 3250,"Record function id": 0, "finished": true, "Ev Idx": 3249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770616944.003, "dur": 0.837,
    "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": 870, "tid": 870,
    "ts": 7429770616948.087, "dur": 7.524,
    "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": 870, "tid": 870,
    "ts": 7429770616973.468, "dur": 0.162,
    "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": 870, "tid": 870,
    "ts": 7429770616975.363, "dur": 9.611,
    "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": 870, "tid": 870,
    "ts": 7429770616976.607, "dur": 0.256,
    "args": {
      "External id": 3255,"Record function id": 0, "finished": true, "Ev Idx": 3254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770616977.186, "dur": 0.098,
    "args": {
      "External id": 3256,"Record function id": 0, "finished": true, "Ev Idx": 3255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770616987.347, "dur": 12.101,
    "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": 870, "tid": 870,
    "ts": 7429770617002.544, "dur": 8.504,
    "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": 870, "tid": 870,
    "ts": 7429770617012.517, "dur": 9.641,
    "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": 870, "tid": 870,
    "ts": 7429770617023.795, "dur": 7.776,
    "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": 870, "tid": 870,
    "ts": 7429770617034.083, "dur": 0.113,
    "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": 870, "tid": 870,
    "ts": 7429770617034.878, "dur": 8.907,
    "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": 870, "tid": 870,
    "ts": 7429770617061.611, "dur": 38.936,
    "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": 870, "tid": 870,
    "ts": 7429770617062.139, "dur": 3.871,
    "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": 870, "tid": 870,
    "ts": 7429770617063.412, "dur": 1.670,
    "args": {
      "External id": 3265,"Record function id": 0, "finished": true, "Ev Idx": 3264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770617064.274, "dur": 0.647,
    "args": {
      "External id": 3266,"Record function id": 0, "finished": true, "Ev Idx": 3265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770617066.372, "dur": 34.078,
    "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": 870, "tid": 870,
    "ts": 7429770617066.966, "dur": 2.698,
    "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": 870, "tid": 870,
    "ts": 7429770617067.359, "dur": 2.193,
    "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": 870, "tid": 870,
    "ts": 7429770617070.033, "dur": 28.421,
    "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": 870, "tid": 870,
    "ts": 7429770617081.356, "dur": 1.143,
    "args": {
      "External id": 3271,"Record function id": 0, "finished": true, "Ev Idx": 3270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770617099.027, "dur": 1.183,
    "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": 870, "tid": 870,
    "ts": 7429770617104.254, "dur": 1.368,
    "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": 870, "tid": 870,
    "ts": 7429770617107.704, "dur": 2.146,
    "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": 870, "tid": 870,
    "ts": 7429770617108.693, "dur": 0.496,
    "args": {
      "External id": 3275,"Record function id": 0, "finished": true, "Ev Idx": 3274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770617116.758, "dur": 32.305,
    "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": 870, "tid": 870,
    "ts": 7429770617117.247, "dur": 2.283,
    "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": 870, "tid": 870,
    "ts": 7429770617118.380, "dur": 0.756,
    "args": {
      "External id": 3278,"Record function id": 0, "finished": true, "Ev Idx": 3277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770617118.753, "dur": 0.290,
    "args": {
      "External id": 3279,"Record function id": 0, "finished": true, "Ev Idx": 3278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770617119.846, "dur": 29.118,
    "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": 870, "tid": 870,
    "ts": 7429770617120.600, "dur": 1.419,
    "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": 870, "tid": 870,
    "ts": 7429770617120.871, "dur": 1.046,
    "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": 870, "tid": 870,
    "ts": 7429770617122.346, "dur": 24.825,
    "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": 870, "tid": 870,
    "ts": 7429770617131.821, "dur": 1.191,
    "args": {
      "External id": 3284,"Record function id": 0, "finished": true, "Ev Idx": 3283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770617147.722, "dur": 1.008,
    "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": 870, "tid": 870,
    "ts": 7429770617151.221, "dur": 1.157,
    "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": 870, "tid": 870,
    "ts": 7429770617154.107, "dur": 1.967,
    "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": 870, "tid": 870,
    "ts": 7429770617155.092, "dur": 0.523,
    "args": {
      "External id": 3288,"Record function id": 0, "finished": true, "Ev Idx": 3287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770617163.364, "dur": 31.113,
    "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": 870, "tid": 870,
    "ts": 7429770617164.058, "dur": 1.934,
    "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": 870, "tid": 870,
    "ts": 7429770617164.803, "dur": 0.819,
    "args": {
      "External id": 3291,"Record function id": 0, "finished": true, "Ev Idx": 3290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770617165.254, "dur": 0.267,
    "args": {
      "External id": 3292,"Record function id": 0, "finished": true, "Ev Idx": 3291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770617166.284, "dur": 28.094,
    "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": 870, "tid": 870,
    "ts": 7429770617166.973, "dur": 1.567,
    "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": 870, "tid": 870,
    "ts": 7429770617167.433, "dur": 0.999,
    "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": 870, "tid": 870,
    "ts": 7429770617168.859, "dur": 23.879,
    "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": 870, "tid": 870,
    "ts": 7429770617177.661, "dur": 1.086,
    "args": {
      "External id": 3297,"Record function id": 0, "finished": true, "Ev Idx": 3296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770617193.268, "dur": 0.885,
    "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": 870, "tid": 870,
    "ts": 7429770617196.842, "dur": 1.122,
    "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": 870, "tid": 870,
    "ts": 7429770617199.550, "dur": 1.496,
    "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": 870, "tid": 870,
    "ts": 7429770617200.268, "dur": 0.335,
    "args": {
      "External id": 3301,"Record function id": 0, "finished": true, "Ev Idx": 3300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770617203.640, "dur": 1.839,
    "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": 870, "tid": 870,
    "ts": 7429770617204.689, "dur": 0.297,
    "args": {
      "External id": 3303,"Record function id": 0, "finished": true, "Ev Idx": 3302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770617206.598, "dur": 1.199,
    "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": 870, "tid": 870,
    "ts": 7429770617207.214, "dur": 0.285,
    "args": {
      "External id": 3305,"Record function id": 0, "finished": true, "Ev Idx": 3304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770617208.770, "dur": 10.905,
    "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": 870, "tid": 870,
    "ts": 7429770617224.457, "dur": 3.352,
    "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": 870, "tid": 870,
    "ts": 7429770617226.201, "dur": 0.504,
    "args": {
      "External id": 3308,"Record function id": 0, "finished": true, "Ev Idx": 3307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770617229.961, "dur": 3.823,
    "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": 870, "tid": 870,
    "ts": 7429770617231.648, "dur": 0.917,
    "args": {
      "External id": 3310,"Record function id": 0, "finished": true, "Ev Idx": 3309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770617235.575, "dur": 11.166,
    "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": 870, "tid": 870,
    "ts": 7429770617250.371, "dur": 17.612,
    "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": 870, "tid": 870,
    "ts": 7429770617271.004, "dur": 8.763,
    "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": 870, "tid": 870,
    "ts": 7429770617281.615, "dur": 7.671,
    "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": 870, "tid": 870,
    "ts": 7429770617291.425, "dur": 9.351,
    "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": 870, "tid": 870,
    "ts": 7429770617303.106, "dur": 2.063,
    "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": 870, "tid": 870,
    "ts": 7429770617304.046, "dur": 0.450,
    "args": {
      "External id": 3317,"Record function id": 0, "finished": true, "Ev Idx": 3316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770617306.662, "dur": 1.652,
    "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": 870, "tid": 870,
    "ts": 7429770617307.463, "dur": 0.329,
    "args": {
      "External id": 3319,"Record function id": 0, "finished": true, "Ev Idx": 3318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770617309.216, "dur": 8.825,
    "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": 870, "tid": 870,
    "ts": 7429770617319.888, "dur": 11.551,
    "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": 870, "tid": 870,
    "ts": 7429770617333.653, "dur": 7.781,
    "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": 870, "tid": 870,
    "ts": 7429770617342.906, "dur": 7.041,
    "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": 870, "tid": 870,
    "ts": 7429770617373.596, "dur": 2.548,
    "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": 870, "tid": 870,
    "ts": 7429770617374.800, "dur": 0.452,
    "args": {
      "External id": 3325,"Record function id": 0, "finished": true, "Ev Idx": 3324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770617376.814, "dur": 1.566,
    "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": 870, "tid": 870,
    "ts": 7429770617377.545, "dur": 0.315,
    "args": {
      "External id": 3327,"Record function id": 0, "finished": true, "Ev Idx": 3326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770617379.066, "dur": 3.269,
    "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": 870, "tid": 870,
    "ts": 7429770617380.897, "dur": 0.800,
    "args": {
      "External id": 3329,"Record function id": 0, "finished": true, "Ev Idx": 3328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770617382.917, "dur": 2.501,
    "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": 870, "tid": 870,
    "ts": 7429770617384.523, "dur": 0.370,
    "args": {
      "External id": 3331,"Record function id": 0, "finished": true, "Ev Idx": 3330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770617385.993, "dur": 2.530,
    "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": 870, "tid": 870,
    "ts": 7429770617387.678, "dur": 0.340,
    "args": {
      "External id": 3333,"Record function id": 0, "finished": true, "Ev Idx": 3332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770617391.293, "dur": 3.025,
    "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": 870, "tid": 870,
    "ts": 7429770617393.093, "dur": 0.548,
    "args": {
      "External id": 3335,"Record function id": 0, "finished": true, "Ev Idx": 3334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770617396.816, "dur": 19.210,
    "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": 870, "tid": 870,
    "ts": 7429770617399.245, "dur": 14.193,
    "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": 870, "tid": 870,
    "ts": 7429770617400.839, "dur": 2.921,
    "args": {
      "External id": 3338,"Record function id": 0, "finished": true, "Ev Idx": 3337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770617401.955, "dur": 1.596,
    "args": {
      "External id": 3339,"Record function id": 0, "finished": true, "Ev Idx": 3338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770617404.351, "dur": 8.534,
    "args": {
      "External id": 3340,"Record function id": 0, "finished": true, "Ev Idx": 3339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770617413.972, "dur": 1.728,
    "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": 870, "tid": 870,
    "ts": 7429770617419.166, "dur": 2.501,
    "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": 870, "tid": 870,
    "ts": 7429770617420.728, "dur": 0.390,
    "args": {
      "External id": 3343,"Record function id": 0, "finished": true, "Ev Idx": 3342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770617422.327, "dur": 1.545,
    "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": 870, "tid": 870,
    "ts": 7429770617423.135, "dur": 0.266,
    "args": {
      "External id": 3345,"Record function id": 0, "finished": true, "Ev Idx": 3344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770617424.385, "dur": 1.690,
    "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": 870, "tid": 870,
    "ts": 7429770617425.324, "dur": 0.293,
    "args": {
      "External id": 3347,"Record function id": 0, "finished": true, "Ev Idx": 3346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770617426.617, "dur": 1.509,
    "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": 870, "tid": 870,
    "ts": 7429770617427.317, "dur": 0.320,
    "args": {
      "External id": 3349,"Record function id": 0, "finished": true, "Ev Idx": 3348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770617428.684, "dur": 1.375,
    "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": 870, "tid": 870,
    "ts": 7429770617429.305, "dur": 0.302,
    "args": {
      "External id": 3351,"Record function id": 0, "finished": true, "Ev Idx": 3350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770617431.636, "dur": 2.276,
    "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": 870, "tid": 870,
    "ts": 7429770617432.999, "dur": 0.465,
    "args": {
      "External id": 3353,"Record function id": 0, "finished": true, "Ev Idx": 3352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770617435.510, "dur": 12.011,
    "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": 870, "tid": 870,
    "ts": 7429770617436.445, "dur": 9.525,
    "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": 870, "tid": 870,
    "ts": 7429770617437.732, "dur": 1.698,
    "args": {
      "External id": 3356,"Record function id": 0, "finished": true, "Ev Idx": 3355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770617438.080, "dur": 1.229,
    "args": {
      "External id": 3357,"Record function id": 0, "finished": true, "Ev Idx": 3356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770617439.899, "dur": 5.746,
    "args": {
      "External id": 3358,"Record function id": 0, "finished": true, "Ev Idx": 3357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770617446.458, "dur": 0.869,
    "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": 870, "tid": 870,
    "ts": 7429770617449.696, "dur": 1.739,
    "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": 870, "tid": 870,
    "ts": 7429770617450.559, "dur": 0.356,
    "args": {
      "External id": 3361,"Record function id": 0, "finished": true, "Ev Idx": 3360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770617453.007, "dur": 112.574,
    "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": 870, "tid": 870,
    "ts": 7429770617453.638, "dur": 1.617,
    "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": 870, "tid": 870,
    "ts": 7429770617454.450, "dur": 0.368,
    "args": {
      "External id": 3364,"Record function id": 0, "finished": true, "Ev Idx": 3363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770617455.604, "dur": 4.268,
    "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": 870, "tid": 870,
    "ts": 7429770617457.231, "dur": 2.468,
    "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": 870, "tid": 870,
    "ts": 7429770617460.327, "dur": 1.958,
    "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": 870, "tid": 870,
    "ts": 7429770617461.602, "dur": 0.298,
    "args": {
      "External id": 3368,"Record function id": 0, "finished": true, "Ev Idx": 3367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770617462.586, "dur": 1.910,
    "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": 870, "tid": 870,
    "ts": 7429770617463.381, "dur": 1.025,
    "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": 870, "tid": 870,
    "ts": 7429770617465.068, "dur": 97.867,
    "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": 870, "tid": 870,
    "ts": 7429770617541.208, "dur": 1.499,
    "args": {
      "External id": 3372,"Record function id": 0, "finished": true, "Ev Idx": 3371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770617563.694, "dur": 1.462,
    "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": 870, "tid": 870,
    "ts": 7429770617570.155, "dur": 11.876,
    "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": 870, "tid": 870,
    "ts": 7429770617585.370, "dur": 2.320,
    "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": 870, "tid": 870,
    "ts": 7429770617586.416, "dur": 0.489,
    "args": {
      "External id": 3376,"Record function id": 0, "finished": true, "Ev Idx": 3375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770617588.425, "dur": 1.552,
    "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": 870, "tid": 870,
    "ts": 7429770617589.155, "dur": 0.396,
    "args": {
      "External id": 3378,"Record function id": 0, "finished": true, "Ev Idx": 3377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770617590.612, "dur": 1.152,
    "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": 870, "tid": 870,
    "ts": 7429770617591.055, "dur": 0.359,
    "args": {
      "External id": 3380,"Record function id": 0, "finished": true, "Ev Idx": 3379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770617593.458, "dur": 8.166,
    "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": 870, "tid": 870,
    "ts": 7429770617607.635, "dur": 14.172,
    "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": 870, "tid": 870,
    "ts": 7429770617608.664, "dur": 0.401,
    "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": 870, "tid": 870,
    "ts": 7429770617609.747, "dur": 11.791,
    "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": 870, "tid": 870,
    "ts": 7429770617624.686, "dur": 0.347,
    "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": 870, "tid": 870,
    "ts": 7429770617636.374, "dur": 0.477,
    "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": 870, "tid": 870,
    "ts": 7429770617638.722, "dur": 44.081,
    "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": 870, "tid": 870,
    "ts": 7429770617639.236, "dur": 3.377,
    "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": 870, "tid": 870,
    "ts": 7429770617641.619, "dur": 0.382,
    "args": {
      "External id": 3389,"Record function id": 0, "finished": true, "Ev Idx": 3388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770617642.974, "dur": 2.846,
    "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": 870, "tid": 870,
    "ts": 7429770617643.532, "dur": 2.165,
    "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": 870, "tid": 870,
    "ts": 7429770617646.337, "dur": 1.511,
    "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": 870, "tid": 870,
    "ts": 7429770617647.225, "dur": 0.282,
    "args": {
      "External id": 3393,"Record function id": 0, "finished": true, "Ev Idx": 3392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770617648.137, "dur": 1.372,
    "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": 870, "tid": 870,
    "ts": 7429770617648.439, "dur": 1.014,
    "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": 870, "tid": 870,
    "ts": 7429770617649.956, "dur": 30.861,
    "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": 870, "tid": 870,
    "ts": 7429770617662.438, "dur": 1.247,
    "args": {
      "External id": 3397,"Record function id": 0, "finished": true, "Ev Idx": 3396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770617681.398, "dur": 1.105,
    "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": 870, "tid": 870,
    "ts": 7429770617685.023, "dur": 1.910,
    "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": 870, "tid": 870,
    "ts": 7429770617686.031, "dur": 0.314,
    "args": {
      "External id": 3400,"Record function id": 0, "finished": true, "Ev Idx": 3399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770617689.507, "dur": 12.537,
    "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": 870, "tid": 870,
    "ts": 7429770617690.043, "dur": 11.854,
    "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": 870, "tid": 870,
    "ts": 7429770617691.091, "dur": 2.256,
    "args": {
      "External id": 3403,"Record function id": 0, "finished": true, "Ev Idx": 3402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770617691.585, "dur": 1.588,
    "args": {
      "External id": 3404,"Record function id": 0, "finished": true, "Ev Idx": 3403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770617693.826, "dur": 7.623,
    "args": {
      "External id": 3405,"Record function id": 0, "finished": true, "Ev Idx": 3404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770617708.450, "dur": 1.816,
    "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": 870, "tid": 870,
    "ts": 7429770617708.988, "dur": 1.178,
    "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": 870, "tid": 870,
    "ts": 7429770617726.568, "dur": 38.974,
    "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": 870, "tid": 870,
    "ts": 7429770617727.589, "dur": 4.359,
    "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": 870, "tid": 870,
    "ts": 7429770617730.100, "dur": 1.155,
    "args": {
      "External id": 3410,"Record function id": 0, "finished": true, "Ev Idx": 3409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770617730.666, "dur": 0.447,
    "args": {
      "External id": 3411,"Record function id": 0, "finished": true, "Ev Idx": 3410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770617732.525, "dur": 32.854,
    "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": 870, "tid": 870,
    "ts": 7429770617732.938, "dur": 2.104,
    "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": 870, "tid": 870,
    "ts": 7429770617733.326, "dur": 1.611,
    "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": 870, "tid": 870,
    "ts": 7429770617735.650, "dur": 27.818,
    "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": 870, "tid": 870,
    "ts": 7429770617745.938, "dur": 1.160,
    "args": {
      "External id": 3416,"Record function id": 0, "finished": true, "Ev Idx": 3415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770617764.063, "dur": 1.034,
    "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": 870, "tid": 870,
    "ts": 7429770617771.223, "dur": 9.792,
    "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": 870, "tid": 870,
    "ts": 7429770617789.936, "dur": 0.249,
    "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": 870, "tid": 870,
    "ts": 7429770617792.525, "dur": 11.201,
    "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": 870, "tid": 870,
    "ts": 7429770617794.638, "dur": 0.416,
    "args": {
      "External id": 3421,"Record function id": 0, "finished": true, "Ev Idx": 3420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770617795.419, "dur": 0.099,
    "args": {
      "External id": 3422,"Record function id": 0, "finished": true, "Ev Idx": 3421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770617806.147, "dur": 13.379,
    "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": 870, "tid": 870,
    "ts": 7429770617823.552, "dur": 9.504,
    "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": 870, "tid": 870,
    "ts": 7429770617835.069, "dur": 11.019,
    "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": 870, "tid": 870,
    "ts": 7429770617847.983, "dur": 9.161,
    "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": 870, "tid": 870,
    "ts": 7429770617859.763, "dur": 0.276,
    "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": 870, "tid": 870,
    "ts": 7429770617860.931, "dur": 9.693,
    "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": 870, "tid": 870,
    "ts": 7429770617887.633, "dur": 40.051,
    "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": 870, "tid": 870,
    "ts": 7429770617888.443, "dur": 4.178,
    "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": 870, "tid": 870,
    "ts": 7429770617890.626, "dur": 1.334,
    "args": {
      "External id": 3431,"Record function id": 0, "finished": true, "Ev Idx": 3430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770617891.310, "dur": 0.459,
    "args": {
      "External id": 3432,"Record function id": 0, "finished": true, "Ev Idx": 3431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770617893.085, "dur": 34.456,
    "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": 870, "tid": 870,
    "ts": 7429770617893.620, "dur": 1.472,
    "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": 870, "tid": 870,
    "ts": 7429770617893.934, "dur": 1.032,
    "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": 870, "tid": 870,
    "ts": 7429770617895.445, "dur": 30.096,
    "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": 870, "tid": 870,
    "ts": 7429770617908.782, "dur": 1.235,
    "args": {
      "External id": 3437,"Record function id": 0, "finished": true, "Ev Idx": 3436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770617926.186, "dur": 1.066,
    "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": 870, "tid": 870,
    "ts": 7429770617936.405, "dur": 10.553,
    "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": 870, "tid": 870,
    "ts": 7429770617954.716, "dur": 35.426,
    "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": 870, "tid": 870,
    "ts": 7429770617955.054, "dur": 2.840,
    "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": 870, "tid": 870,
    "ts": 7429770617956.100, "dur": 1.072,
    "args": {
      "External id": 3442,"Record function id": 0, "finished": true, "Ev Idx": 3441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770617956.507, "dur": 0.492,
    "args": {
      "External id": 3443,"Record function id": 0, "finished": true, "Ev Idx": 3442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770617958.369, "dur": 31.671,
    "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": 870, "tid": 870,
    "ts": 7429770617958.775, "dur": 2.757,
    "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": 870, "tid": 870,
    "ts": 7429770617959.065, "dur": 2.361,
    "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": 870, "tid": 870,
    "ts": 7429770617961.851, "dur": 26.419,
    "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": 870, "tid": 870,
    "ts": 7429770617970.599, "dur": 1.131,
    "args": {
      "External id": 3448,"Record function id": 0, "finished": true, "Ev Idx": 3447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770617988.882, "dur": 0.915,
    "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": 870, "tid": 870,
    "ts": 7429770617991.920, "dur": 7.987,
    "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": 870, "tid": 870,
    "ts": 7429770618006.319, "dur": 44.067,
    "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": 870, "tid": 870,
    "ts": 7429770618006.692, "dur": 2.223,
    "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": 870, "tid": 870,
    "ts": 7429770618007.585, "dur": 0.857,
    "args": {
      "External id": 3453,"Record function id": 0, "finished": true, "Ev Idx": 3452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770618007.976, "dur": 0.341,
    "args": {
      "External id": 3454,"Record function id": 0, "finished": true, "Ev Idx": 3453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770618009.184, "dur": 41.103,
    "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": 870, "tid": 870,
    "ts": 7429770618009.735, "dur": 1.978,
    "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": 870, "tid": 870,
    "ts": 7429770618010.038, "dur": 1.573,
    "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": 870, "tid": 870,
    "ts": 7429770618011.994, "dur": 36.538,
    "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": 870, "tid": 870,
    "ts": 7429770618021.194, "dur": 1.185,
    "args": {
      "External id": 3459,"Record function id": 0, "finished": true, "Ev Idx": 3458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770618049.175, "dur": 0.880,
    "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": 870, "tid": 870,
    "ts": 7429770618052.790, "dur": 7.948,
    "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": 870, "tid": 870,
    "ts": 7429770618099.236, "dur": 0.380,
    "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": 870, "tid": 870,
    "ts": 7429770618102.159, "dur": 13.021,
    "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": 870, "tid": 870,
    "ts": 7429770618104.191, "dur": 0.554,
    "args": {
      "External id": 3464,"Record function id": 0, "finished": true, "Ev Idx": 3463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770618105.070, "dur": 0.093,
    "args": {
      "External id": 3465,"Record function id": 0, "finished": true, "Ev Idx": 3464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770618117.622, "dur": 13.167,
    "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": 870, "tid": 870,
    "ts": 7429770618135.253, "dur": 8.976,
    "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": 870, "tid": 870,
    "ts": 7429770618145.937, "dur": 8.049,
    "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": 870, "tid": 870,
    "ts": 7429770618155.530, "dur": 9.610,
    "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": 870, "tid": 870,
    "ts": 7429770618167.455, "dur": 0.236,
    "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": 870, "tid": 870,
    "ts": 7429770618168.504, "dur": 10.105,
    "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": 870, "tid": 870,
    "ts": 7429770618197.960, "dur": 112.749,
    "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": 870, "tid": 870,
    "ts": 7429770618198.550, "dur": 5.689,
    "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": 870, "tid": 870,
    "ts": 7429770618201.339, "dur": 1.634,
    "args": {
      "External id": 3474,"Record function id": 0, "finished": true, "Ev Idx": 3473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770618202.137, "dur": 0.655,
    "args": {
      "External id": 3475,"Record function id": 0, "finished": true, "Ev Idx": 3474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770618204.671, "dur": 105.884,
    "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": 870, "tid": 870,
    "ts": 7429770618205.246, "dur": 2.750,
    "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": 870, "tid": 870,
    "ts": 7429770618206.119, "dur": 1.738,
    "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": 870, "tid": 870,
    "ts": 7429770618208.348, "dur": 99.532,
    "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": 870, "tid": 870,
    "ts": 7429770618220.511, "dur": 66.435,
    "args": {
      "External id": 3480,"Record function id": 0, "finished": true, "Ev Idx": 3479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770618308.637, "dur": 1.442,
    "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": 870, "tid": 870,
    "ts": 7429770618314.725, "dur": 1.680,
    "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": 870, "tid": 870,
    "ts": 7429770618318.301, "dur": 2.726,
    "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": 870, "tid": 870,
    "ts": 7429770618319.509, "dur": 0.628,
    "args": {
      "External id": 3484,"Record function id": 0, "finished": true, "Ev Idx": 3483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770618330.977, "dur": 37.051,
    "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": 870, "tid": 870,
    "ts": 7429770618331.453, "dur": 2.727,
    "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": 870, "tid": 870,
    "ts": 7429770618332.790, "dur": 0.924,
    "args": {
      "External id": 3487,"Record function id": 0, "finished": true, "Ev Idx": 3486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770618333.323, "dur": 0.284,
    "args": {
      "External id": 3488,"Record function id": 0, "finished": true, "Ev Idx": 3487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770618334.510, "dur": 33.415,
    "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": 870, "tid": 870,
    "ts": 7429770618334.969, "dur": 1.323,
    "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": 870, "tid": 870,
    "ts": 7429770618335.305, "dur": 0.891,
    "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": 870, "tid": 870,
    "ts": 7429770618336.639, "dur": 29.515,
    "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": 870, "tid": 870,
    "ts": 7429770618349.547, "dur": 1.248,
    "args": {
      "External id": 3493,"Record function id": 0, "finished": true, "Ev Idx": 3492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770618366.704, "dur": 0.936,
    "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": 870, "tid": 870,
    "ts": 7429770618370.582, "dur": 1.219,
    "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": 870, "tid": 870,
    "ts": 7429770618373.510, "dur": 2.070,
    "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": 870, "tid": 870,
    "ts": 7429770618374.512, "dur": 0.577,
    "args": {
      "External id": 3497,"Record function id": 0, "finished": true, "Ev Idx": 3496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770618382.836, "dur": 32.906,
    "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": 870, "tid": 870,
    "ts": 7429770618383.247, "dur": 2.288,
    "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": 870, "tid": 870,
    "ts": 7429770618384.094, "dur": 0.995,
    "args": {
      "External id": 3500,"Record function id": 0, "finished": true, "Ev Idx": 3499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770618384.722, "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": 870, "tid": 870,
    "ts": 7429770618385.832, "dur": 29.812,
    "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": 870, "tid": 870,
    "ts": 7429770618386.271, "dur": 1.175,
    "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": 870, "tid": 870,
    "ts": 7429770618386.525, "dur": 0.814,
    "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": 870, "tid": 870,
    "ts": 7429770618387.794, "dur": 26.159,
    "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": 870, "tid": 870,
    "ts": 7429770618398.261, "dur": 1.249,
    "args": {
      "External id": 3506,"Record function id": 0, "finished": true, "Ev Idx": 3505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770618414.509, "dur": 0.900,
    "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": 870, "tid": 870,
    "ts": 7429770618418.342, "dur": 1.131,
    "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": 870, "tid": 870,
    "ts": 7429770618421.053, "dur": 1.701,
    "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": 870, "tid": 870,
    "ts": 7429770618421.936, "dur": 0.388,
    "args": {
      "External id": 3510,"Record function id": 0, "finished": true, "Ev Idx": 3509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770618425.762, "dur": 2.578,
    "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": 870, "tid": 870,
    "ts": 7429770618427.330, "dur": 0.493,
    "args": {
      "External id": 3512,"Record function id": 0, "finished": true, "Ev Idx": 3511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770618429.430, "dur": 1.261,
    "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": 870, "tid": 870,
    "ts": 7429770618429.874, "dur": 0.446,
    "args": {
      "External id": 3514,"Record function id": 0, "finished": true, "Ev Idx": 3513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770618431.778, "dur": 10.428,
    "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": 870, "tid": 870,
    "ts": 7429770618448.315, "dur": 4.710,
    "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": 870, "tid": 870,
    "ts": 7429770618450.764, "dur": 0.578,
    "args": {
      "External id": 3517,"Record function id": 0, "finished": true, "Ev Idx": 3516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770618455.344, "dur": 3.046,
    "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": 870, "tid": 870,
    "ts": 7429770618456.800, "dur": 0.589,
    "args": {
      "External id": 3519,"Record function id": 0, "finished": true, "Ev Idx": 3518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770618460.383, "dur": 12.430,
    "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": 870, "tid": 870,
    "ts": 7429770618477.500, "dur": 18.270,
    "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": 870, "tid": 870,
    "ts": 7429770618498.978, "dur": 8.419,
    "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": 870, "tid": 870,
    "ts": 7429770618509.030, "dur": 8.140,
    "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": 870, "tid": 870,
    "ts": 7429770618519.380, "dur": 8.677,
    "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": 870, "tid": 870,
    "ts": 7429770618530.558, "dur": 2.333,
    "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": 870, "tid": 870,
    "ts": 7429770618531.644, "dur": 0.590,
    "args": {
      "External id": 3526,"Record function id": 0, "finished": true, "Ev Idx": 3525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770618534.454, "dur": 1.978,
    "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": 870, "tid": 870,
    "ts": 7429770618535.577, "dur": 0.313,
    "args": {
      "External id": 3528,"Record function id": 0, "finished": true, "Ev Idx": 3527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770618537.374, "dur": 9.766,
    "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": 870, "tid": 870,
    "ts": 7429770618549.246, "dur": 11.061,
    "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": 870, "tid": 870,
    "ts": 7429770618562.837, "dur": 7.545,
    "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": 870, "tid": 870,
    "ts": 7429770618571.719, "dur": 6.437,
    "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": 870, "tid": 870,
    "ts": 7429770618602.960, "dur": 2.641,
    "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": 870, "tid": 870,
    "ts": 7429770618604.275, "dur": 0.432,
    "args": {
      "External id": 3534,"Record function id": 0, "finished": true, "Ev Idx": 3533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770618606.364, "dur": 1.833,
    "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": 870, "tid": 870,
    "ts": 7429770618607.283, "dur": 0.387,
    "args": {
      "External id": 3536,"Record function id": 0, "finished": true, "Ev Idx": 3535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770618608.870, "dur": 3.054,
    "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": 870, "tid": 870,
    "ts": 7429770618610.497, "dur": 0.727,
    "args": {
      "External id": 3538,"Record function id": 0, "finished": true, "Ev Idx": 3537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770618612.590, "dur": 2.956,
    "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": 870, "tid": 870,
    "ts": 7429770618614.623, "dur": 0.369,
    "args": {
      "External id": 3540,"Record function id": 0, "finished": true, "Ev Idx": 3539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770618616.172, "dur": 2.152,
    "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": 870, "tid": 870,
    "ts": 7429770618617.517, "dur": 0.335,
    "args": {
      "External id": 3542,"Record function id": 0, "finished": true, "Ev Idx": 3541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770618621.131, "dur": 3.129,
    "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": 870, "tid": 870,
    "ts": 7429770618623.092, "dur": 0.385,
    "args": {
      "External id": 3544,"Record function id": 0, "finished": true, "Ev Idx": 3543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770618626.741, "dur": 19.980,
    "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": 870, "tid": 870,
    "ts": 7429770618629.308, "dur": 15.130,
    "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": 870, "tid": 870,
    "ts": 7429770618631.136, "dur": 2.850,
    "args": {
      "External id": 3547,"Record function id": 0, "finished": true, "Ev Idx": 3546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770618632.202, "dur": 1.487,
    "args": {
      "External id": 3548,"Record function id": 0, "finished": true, "Ev Idx": 3547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770618634.785, "dur": 9.119,
    "args": {
      "External id": 3549,"Record function id": 0, "finished": true, "Ev Idx": 3548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770618645.127, "dur": 1.321,
    "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": 870, "tid": 870,
    "ts": 7429770618649.956, "dur": 3.065,
    "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": 870, "tid": 870,
    "ts": 7429770618652.029, "dur": 0.440,
    "args": {
      "External id": 3552,"Record function id": 0, "finished": true, "Ev Idx": 3551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770618653.635, "dur": 2.666,
    "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": 870, "tid": 870,
    "ts": 7429770618655.432, "dur": 0.404,
    "args": {
      "External id": 3554,"Record function id": 0, "finished": true, "Ev Idx": 3553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770618656.802, "dur": 1.605,
    "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": 870, "tid": 870,
    "ts": 7429770618657.547, "dur": 0.464,
    "args": {
      "External id": 3556,"Record function id": 0, "finished": true, "Ev Idx": 3555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770618658.985, "dur": 1.853,
    "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": 870, "tid": 870,
    "ts": 7429770618659.984, "dur": 0.324,
    "args": {
      "External id": 3558,"Record function id": 0, "finished": true, "Ev Idx": 3557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770618661.423, "dur": 1.715,
    "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": 870, "tid": 870,
    "ts": 7429770618662.386, "dur": 0.287,
    "args": {
      "External id": 3560,"Record function id": 0, "finished": true, "Ev Idx": 3559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770618664.575, "dur": 1.832,
    "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": 870, "tid": 870,
    "ts": 7429770618665.525, "dur": 0.442,
    "args": {
      "External id": 3562,"Record function id": 0, "finished": true, "Ev Idx": 3561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770618667.885, "dur": 12.614,
    "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": 870, "tid": 870,
    "ts": 7429770618668.748, "dur": 10.293,
    "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": 870, "tid": 870,
    "ts": 7429770618670.707, "dur": 1.871,
    "args": {
      "External id": 3565,"Record function id": 0, "finished": true, "Ev Idx": 3564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770618671.250, "dur": 1.200,
    "args": {
      "External id": 3566,"Record function id": 0, "finished": true, "Ev Idx": 3565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770618672.938, "dur": 5.791,
    "args": {
      "External id": 3567,"Record function id": 0, "finished": true, "Ev Idx": 3566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770618679.418, "dur": 0.926,
    "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": 870, "tid": 870,
    "ts": 7429770618682.556, "dur": 1.695,
    "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": 870, "tid": 870,
    "ts": 7429770618683.396, "dur": 0.352,
    "args": {
      "External id": 3570,"Record function id": 0, "finished": true, "Ev Idx": 3569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770618685.896, "dur": 98.804,
    "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": 870, "tid": 870,
    "ts": 7429770618686.430, "dur": 1.494,
    "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": 870, "tid": 870,
    "ts": 7429770618687.164, "dur": 0.311,
    "args": {
      "External id": 3573,"Record function id": 0, "finished": true, "Ev Idx": 3572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770618688.472, "dur": 4.311,
    "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": 870, "tid": 870,
    "ts": 7429770618690.492, "dur": 2.137,
    "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": 870, "tid": 870,
    "ts": 7429770618693.386, "dur": 2.438,
    "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": 870, "tid": 870,
    "ts": 7429770618695.157, "dur": 0.290,
    "args": {
      "External id": 3577,"Record function id": 0, "finished": true, "Ev Idx": 3576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770618696.117, "dur": 2.409,
    "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": 870, "tid": 870,
    "ts": 7429770618696.930, "dur": 1.491,
    "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": 870, "tid": 870,
    "ts": 7429770618699.175, "dur": 83.180,
    "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": 870, "tid": 870,
    "ts": 7429770618761.514, "dur": 1.249,
    "args": {
      "External id": 3581,"Record function id": 0, "finished": true, "Ev Idx": 3580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770618783.129, "dur": 1.190,
    "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": 870, "tid": 870,
    "ts": 7429770618788.732, "dur": 11.133,
    "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": 870, "tid": 870,
    "ts": 7429770618803.393, "dur": 35.937,
    "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": 870, "tid": 870,
    "ts": 7429770618837.936, "dur": 0.463,
    "args": {
      "External id": 3585,"Record function id": 0, "finished": true, "Ev Idx": 3584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770618840.047, "dur": 1.385,
    "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": 870, "tid": 870,
    "ts": 7429770618840.651, "dur": 0.331,
    "args": {
      "External id": 3587,"Record function id": 0, "finished": true, "Ev Idx": 3586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770618842.063, "dur": 0.990,
    "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": 870, "tid": 870,
    "ts": 7429770618842.532, "dur": 0.201,
    "args": {
      "External id": 3589,"Record function id": 0, "finished": true, "Ev Idx": 3588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770618844.784, "dur": 8.081,
    "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": 870, "tid": 870,
    "ts": 7429770618858.726, "dur": 14.076,
    "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": 870, "tid": 870,
    "ts": 7429770618859.732, "dur": 0.343,
    "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": 870, "tid": 870,
    "ts": 7429770618860.831, "dur": 11.697,
    "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": 870, "tid": 870,
    "ts": 7429770618876.008, "dur": 0.192,
    "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": 870, "tid": 870,
    "ts": 7429770618887.977, "dur": 0.473,
    "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": 870, "tid": 870,
    "ts": 7429770618890.228, "dur": 42.092,
    "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": 870, "tid": 870,
    "ts": 7429770618890.738, "dur": 2.700,
    "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": 870, "tid": 870,
    "ts": 7429770618892.407, "dur": 0.445,
    "args": {
      "External id": 3598,"Record function id": 0, "finished": true, "Ev Idx": 3597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770618893.850, "dur": 2.638,
    "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": 870, "tid": 870,
    "ts": 7429770618894.220, "dur": 2.141,
    "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": 870, "tid": 870,
    "ts": 7429770618896.886, "dur": 1.248,
    "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": 870, "tid": 870,
    "ts": 7429770618897.533, "dur": 0.226,
    "args": {
      "External id": 3602,"Record function id": 0, "finished": true, "Ev Idx": 3601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770618898.444, "dur": 1.177,
    "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": 870, "tid": 870,
    "ts": 7429770618898.684, "dur": 0.857,
    "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": 870, "tid": 870,
    "ts": 7429770618900.037, "dur": 30.343,
    "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": 870, "tid": 870,
    "ts": 7429770618913.344, "dur": 1.182,
    "args": {
      "External id": 3606,"Record function id": 0, "finished": true, "Ev Idx": 3605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770618930.981, "dur": 1.051,
    "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": 870, "tid": 870,
    "ts": 7429770618934.404, "dur": 1.952,
    "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": 870, "tid": 870,
    "ts": 7429770618935.454, "dur": 0.339,
    "args": {
      "External id": 3609,"Record function id": 0, "finished": true, "Ev Idx": 3608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770618939.157, "dur": 12.086,
    "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": 870, "tid": 870,
    "ts": 7429770618939.744, "dur": 11.348,
    "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": 870, "tid": 870,
    "ts": 7429770618940.840, "dur": 1.914,
    "args": {
      "External id": 3612,"Record function id": 0, "finished": true, "Ev Idx": 3611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770618941.259, "dur": 1.328,
    "args": {
      "External id": 3613,"Record function id": 0, "finished": true, "Ev Idx": 3612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770618943.209, "dur": 7.422,
    "args": {
      "External id": 3614,"Record function id": 0, "finished": true, "Ev Idx": 3613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770618957.800, "dur": 1.456,
    "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": 870, "tid": 870,
    "ts": 7429770618958.068, "dur": 1.101,
    "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": 870, "tid": 870,
    "ts": 7429770618974.875, "dur": 36.896,
    "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": 870, "tid": 870,
    "ts": 7429770618975.710, "dur": 3.616,
    "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": 870, "tid": 870,
    "ts": 7429770618977.858, "dur": 0.892,
    "args": {
      "External id": 3619,"Record function id": 0, "finished": true, "Ev Idx": 3618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770618978.272, "dur": 0.336,
    "args": {
      "External id": 3620,"Record function id": 0, "finished": true, "Ev Idx": 3619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770618979.705, "dur": 31.900,
    "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": 870, "tid": 870,
    "ts": 7429770618980.134, "dur": 1.891,
    "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": 870, "tid": 870,
    "ts": 7429770618980.379, "dur": 1.538,
    "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": 870, "tid": 870,
    "ts": 7429770618982.653, "dur": 27.063,
    "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": 870, "tid": 870,
    "ts": 7429770618992.959, "dur": 1.096,
    "args": {
      "External id": 3625,"Record function id": 0, "finished": true, "Ev Idx": 3624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770619010.307, "dur": 1.007,
    "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": 870, "tid": 870,
    "ts": 7429770619017.364, "dur": 8.781,
    "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": 870, "tid": 870,
    "ts": 7429770619036.385, "dur": 0.187,
    "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": 870, "tid": 870,
    "ts": 7429770619039.119, "dur": 11.761,
    "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": 870, "tid": 870,
    "ts": 7429770619041.056, "dur": 0.370,
    "args": {
      "External id": 3630,"Record function id": 0, "finished": true, "Ev Idx": 3629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770619041.991, "dur": 0.080,
    "args": {
      "External id": 3631,"Record function id": 0, "finished": true, "Ev Idx": 3630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770619053.712, "dur": 12.909,
    "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": 870, "tid": 870,
    "ts": 7429770619070.199, "dur": 8.279,
    "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": 870, "tid": 870,
    "ts": 7429770619080.493, "dur": 11.282,
    "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": 870, "tid": 870,
    "ts": 7429770619093.620, "dur": 10.228,
    "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": 870, "tid": 870,
    "ts": 7429770619106.404, "dur": 0.097,
    "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": 870, "tid": 870,
    "ts": 7429770619107.383, "dur": 8.178,
    "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": 870, "tid": 870,
    "ts": 7429770619144.861, "dur": 33.914,
    "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": 870, "tid": 870,
    "ts": 7429770619145.339, "dur": 2.742,
    "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": 870, "tid": 870,
    "ts": 7429770619146.322, "dur": 1.180,
    "args": {
      "External id": 3640,"Record function id": 0, "finished": true, "Ev Idx": 3639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770619146.900, "dur": 0.448,
    "args": {
      "External id": 3641,"Record function id": 0, "finished": true, "Ev Idx": 3640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770619148.474, "dur": 30.148,
    "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": 870, "tid": 870,
    "ts": 7429770619148.999, "dur": 1.613,
    "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": 870, "tid": 870,
    "ts": 7429770619149.307, "dur": 1.148,
    "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": 870, "tid": 870,
    "ts": 7429770619150.978, "dur": 25.583,
    "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": 870, "tid": 870,
    "ts": 7429770619159.494, "dur": 1.247,
    "args": {
      "External id": 3646,"Record function id": 0, "finished": true, "Ev Idx": 3645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770619177.156, "dur": 1.142,
    "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": 870, "tid": 870,
    "ts": 7429770619187.462, "dur": 11.064,
    "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": 870, "tid": 870,
    "ts": 7429770619206.299, "dur": 27.640,
    "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": 870, "tid": 870,
    "ts": 7429770619206.605, "dur": 2.191,
    "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": 870, "tid": 870,
    "ts": 7429770619207.429, "dur": 0.896,
    "args": {
      "External id": 3651,"Record function id": 0, "finished": true, "Ev Idx": 3650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770619207.842, "dur": 0.359,
    "args": {
      "External id": 3652,"Record function id": 0, "finished": true, "Ev Idx": 3651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770619209.128, "dur": 24.728,
    "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": 870, "tid": 870,
    "ts": 7429770619209.556, "dur": 1.362,
    "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": 870, "tid": 870,
    "ts": 7429770619209.830, "dur": 0.980,
    "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": 870, "tid": 870,
    "ts": 7429770619211.231, "dur": 20.807,
    "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": 870, "tid": 870,
    "ts": 7429770619218.122, "dur": 1.180,
    "args": {
      "External id": 3657,"Record function id": 0, "finished": true, "Ev Idx": 3656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770619232.572, "dur": 1.019,
    "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": 870, "tid": 870,
    "ts": 7429770619235.586, "dur": 7.631,
    "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": 870, "tid": 870,
    "ts": 7429770619249.172, "dur": 45.413,
    "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": 870, "tid": 870,
    "ts": 7429770619249.620, "dur": 2.099,
    "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": 870, "tid": 870,
    "ts": 7429770619250.395, "dur": 0.806,
    "args": {
      "External id": 3662,"Record function id": 0, "finished": true, "Ev Idx": 3661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770619250.745, "dur": 0.345,
    "args": {
      "External id": 3663,"Record function id": 0, "finished": true, "Ev Idx": 3662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770619252.017, "dur": 42.448,
    "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": 870, "tid": 870,
    "ts": 7429770619252.658, "dur": 1.229,
    "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": 870, "tid": 870,
    "ts": 7429770619252.893, "dur": 0.903,
    "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": 870, "tid": 870,
    "ts": 7429770619254.176, "dur": 38.550,
    "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": 870, "tid": 870,
    "ts": 7429770619263.289, "dur": 1.138,
    "args": {
      "External id": 3668,"Record function id": 0, "finished": true, "Ev Idx": 3667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770619293.330, "dur": 0.893,
    "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": 870, "tid": 870,
    "ts": 7429770619297.304, "dur": 8.244,
    "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": 870, "tid": 870,
    "ts": 7429770619324.424, "dur": 0.158,
    "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": 870, "tid": 870,
    "ts": 7429770619326.275, "dur": 8.584,
    "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": 870, "tid": 870,
    "ts": 7429770619327.485, "dur": 0.238,
    "args": {
      "External id": 3673,"Record function id": 0, "finished": true, "Ev Idx": 3672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770619327.990, "dur": 0.066,
    "args": {
      "External id": 3674,"Record function id": 0, "finished": true, "Ev Idx": 3673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770619337.442, "dur": 11.322,
    "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": 870, "tid": 870,
    "ts": 7429770619352.060, "dur": 8.955,
    "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": 870, "tid": 870,
    "ts": 7429770619362.547, "dur": 8.267,
    "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": 870, "tid": 870,
    "ts": 7429770619372.356, "dur": 8.005,
    "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": 870, "tid": 870,
    "ts": 7429770619382.888, "dur": 0.119,
    "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": 870, "tid": 870,
    "ts": 7429770619383.795, "dur": 8.963,
    "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": 870, "tid": 870,
    "ts": 7429770619410.921, "dur": 38.169,
    "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": 870, "tid": 870,
    "ts": 7429770619411.326, "dur": 4.258,
    "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": 870, "tid": 870,
    "ts": 7429770619413.026, "dur": 1.341,
    "args": {
      "External id": 3683,"Record function id": 0, "finished": true, "Ev Idx": 3682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770619413.778, "dur": 0.448,
    "args": {
      "External id": 3684,"Record function id": 0, "finished": true, "Ev Idx": 3683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770619415.944, "dur": 33.028,
    "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": 870, "tid": 870,
    "ts": 7429770619416.420, "dur": 1.734,
    "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": 870, "tid": 870,
    "ts": 7429770619416.708, "dur": 1.321,
    "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": 870, "tid": 870,
    "ts": 7429770619418.532, "dur": 28.472,
    "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": 870, "tid": 870,
    "ts": 7429770619430.498, "dur": 1.132,
    "args": {
      "External id": 3689,"Record function id": 0, "finished": true, "Ev Idx": 3688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770619447.609, "dur": 1.049,
    "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": 870, "tid": 870,
    "ts": 7429770619452.404, "dur": 1.410,
    "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": 870, "tid": 870,
    "ts": 7429770619455.643, "dur": 2.246,
    "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": 870, "tid": 870,
    "ts": 7429770619456.672, "dur": 0.489,
    "args": {
      "External id": 3693,"Record function id": 0, "finished": true, "Ev Idx": 3692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770619465.402, "dur": 31.467,
    "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": 870, "tid": 870,
    "ts": 7429770619465.778, "dur": 1.920,
    "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": 870, "tid": 870,
    "ts": 7429770619466.526, "dur": 0.742,
    "args": {
      "External id": 3696,"Record function id": 0, "finished": true, "Ev Idx": 3695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770619466.899, "dur": 0.281,
    "args": {
      "External id": 3697,"Record function id": 0, "finished": true, "Ev Idx": 3696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770619468.005, "dur": 28.769,
    "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": 870, "tid": 870,
    "ts": 7429770619468.416, "dur": 1.246,
    "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": 870, "tid": 870,
    "ts": 7429770619468.667, "dur": 0.877,
    "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": 870, "tid": 870,
    "ts": 7429770619470.002, "dur": 25.033,
    "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": 870, "tid": 870,
    "ts": 7429770619479.266, "dur": 1.175,
    "args": {
      "External id": 3702,"Record function id": 0, "finished": true, "Ev Idx": 3701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770619495.557, "dur": 0.970,
    "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": 870, "tid": 870,
    "ts": 7429770619498.964, "dur": 1.159,
    "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": 870, "tid": 870,
    "ts": 7429770619501.545, "dur": 1.540,
    "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": 870, "tid": 870,
    "ts": 7429770619502.269, "dur": 0.338,
    "args": {
      "External id": 3706,"Record function id": 0, "finished": true, "Ev Idx": 3705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770619509.905, "dur": 32.666,
    "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": 870, "tid": 870,
    "ts": 7429770619510.211, "dur": 1.977,
    "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": 870, "tid": 870,
    "ts": 7429770619510.993, "dur": 0.783,
    "args": {
      "External id": 3709,"Record function id": 0, "finished": true, "Ev Idx": 3708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770619511.354, "dur": 0.300,
    "args": {
      "External id": 3710,"Record function id": 0, "finished": true, "Ev Idx": 3709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770619512.471, "dur": 30.005,
    "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": 870, "tid": 870,
    "ts": 7429770619512.830, "dur": 1.302,
    "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": 870, "tid": 870,
    "ts": 7429770619513.090, "dur": 0.935,
    "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": 870, "tid": 870,
    "ts": 7429770619514.447, "dur": 26.385,
    "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": 870, "tid": 870,
    "ts": 7429770619525.948, "dur": 1.035,
    "args": {
      "External id": 3715,"Record function id": 0, "finished": true, "Ev Idx": 3714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770619541.335, "dur": 0.890,
    "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": 870, "tid": 870,
    "ts": 7429770619544.801, "dur": 1.110,
    "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": 870, "tid": 870,
    "ts": 7429770619547.399, "dur": 1.665,
    "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": 870, "tid": 870,
    "ts": 7429770619548.204, "dur": 0.377,
    "args": {
      "External id": 3719,"Record function id": 0, "finished": true, "Ev Idx": 3718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770619551.643, "dur": 2.202,
    "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": 870, "tid": 870,
    "ts": 7429770619552.980, "dur": 0.328,
    "args": {
      "External id": 3721,"Record function id": 0, "finished": true, "Ev Idx": 3720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770619554.895, "dur": 1.017,
    "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": 870, "tid": 870,
    "ts": 7429770619555.309, "dur": 0.295,
    "args": {
      "External id": 3723,"Record function id": 0, "finished": true, "Ev Idx": 3722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770619556.971, "dur": 12.297,
    "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": 870, "tid": 870,
    "ts": 7429770619574.009, "dur": 5.219,
    "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": 870, "tid": 870,
    "ts": 7429770619576.794, "dur": 1.179,
    "args": {
      "External id": 3726,"Record function id": 0, "finished": true, "Ev Idx": 3725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770619581.577, "dur": 1.768,
    "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": 870, "tid": 870,
    "ts": 7429770619582.447, "dur": 0.337,
    "args": {
      "External id": 3728,"Record function id": 0, "finished": true, "Ev Idx": 3727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770619585.172, "dur": 11.403,
    "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": 870, "tid": 870,
    "ts": 7429770619600.866, "dur": 16.827,
    "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": 870, "tid": 870,
    "ts": 7429770619620.748, "dur": 7.758,
    "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": 870, "tid": 870,
    "ts": 7429770619630.048, "dur": 9.390,
    "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": 870, "tid": 870,
    "ts": 7429770619641.616, "dur": 8.292,
    "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": 870, "tid": 870,
    "ts": 7429770619652.460, "dur": 2.317,
    "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": 870, "tid": 870,
    "ts": 7429770619653.656, "dur": 0.419,
    "args": {
      "External id": 3735,"Record function id": 0, "finished": true, "Ev Idx": 3734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770619656.294, "dur": 3.527,
    "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": 870, "tid": 870,
    "ts": 7429770619658.712, "dur": 0.566,
    "args": {
      "External id": 3737,"Record function id": 0, "finished": true, "Ev Idx": 3736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770619660.795, "dur": 8.254,
    "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": 870, "tid": 870,
    "ts": 7429770619671.028, "dur": 12.049,
    "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": 870, "tid": 870,
    "ts": 7429770619685.406, "dur": 8.832,
    "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": 870, "tid": 870,
    "ts": 7429770619695.540, "dur": 6.550,
    "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": 870, "tid": 870,
    "ts": 7429770619725.704, "dur": 3.428,
    "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": 870, "tid": 870,
    "ts": 7429770619727.887, "dur": 0.420,
    "args": {
      "External id": 3743,"Record function id": 0, "finished": true, "Ev Idx": 3742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770619729.842, "dur": 2.159,
    "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": 870, "tid": 870,
    "ts": 7429770619731.132, "dur": 0.320,
    "args": {
      "External id": 3745,"Record function id": 0, "finished": true, "Ev Idx": 3744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770619732.733, "dur": 2.443,
    "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": 870, "tid": 870,
    "ts": 7429770619733.714, "dur": 0.719,
    "args": {
      "External id": 3747,"Record function id": 0, "finished": true, "Ev Idx": 3746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770619735.982, "dur": 1.673,
    "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": 870, "tid": 870,
    "ts": 7429770619736.766, "dur": 0.404,
    "args": {
      "External id": 3749,"Record function id": 0, "finished": true, "Ev Idx": 3748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770619738.251, "dur": 1.506,
    "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": 870, "tid": 870,
    "ts": 7429770619738.980, "dur": 0.317,
    "args": {
      "External id": 3751,"Record function id": 0, "finished": true, "Ev Idx": 3750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770619742.300, "dur": 2.756,
    "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": 870, "tid": 870,
    "ts": 7429770619743.960, "dur": 0.433,
    "args": {
      "External id": 3753,"Record function id": 0, "finished": true, "Ev Idx": 3752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770619747.314, "dur": 19.530,
    "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": 870, "tid": 870,
    "ts": 7429770619749.619, "dur": 15.104,
    "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": 870, "tid": 870,
    "ts": 7429770619751.862, "dur": 3.184,
    "args": {
      "External id": 3756,"Record function id": 0, "finished": true, "Ev Idx": 3755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770619753.181, "dur": 1.624,
    "args": {
      "External id": 3757,"Record function id": 0, "finished": true, "Ev Idx": 3756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770619755.696, "dur": 8.593,
    "args": {
      "External id": 3758,"Record function id": 0, "finished": true, "Ev Idx": 3757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770619765.231, "dur": 1.323,
    "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": 870, "tid": 870,
    "ts": 7429770619770.083, "dur": 1.814,
    "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": 870, "tid": 870,
    "ts": 7429770619770.962, "dur": 0.376,
    "args": {
      "External id": 3761,"Record function id": 0, "finished": true, "Ev Idx": 3760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770619772.507, "dur": 1.482,
    "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": 870, "tid": 870,
    "ts": 7429770619773.227, "dur": 0.286,
    "args": {
      "External id": 3763,"Record function id": 0, "finished": true, "Ev Idx": 3762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770619774.480, "dur": 1.401,
    "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": 870, "tid": 870,
    "ts": 7429770619775.176, "dur": 0.296,
    "args": {
      "External id": 3765,"Record function id": 0, "finished": true, "Ev Idx": 3764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770619776.589, "dur": 2.483,
    "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": 870, "tid": 870,
    "ts": 7429770619778.253, "dur": 0.323,
    "args": {
      "External id": 3767,"Record function id": 0, "finished": true, "Ev Idx": 3766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770619779.634, "dur": 2.168,
    "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": 870, "tid": 870,
    "ts": 7429770619781.047, "dur": 0.285,
    "args": {
      "External id": 3769,"Record function id": 0, "finished": true, "Ev Idx": 3768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770619783.336, "dur": 1.567,
    "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": 870, "tid": 870,
    "ts": 7429770619784.122, "dur": 0.321,
    "args": {
      "External id": 3771,"Record function id": 0, "finished": true, "Ev Idx": 3770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770619786.278, "dur": 10.906,
    "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": 870, "tid": 870,
    "ts": 7429770619787.153, "dur": 8.536,
    "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": 870, "tid": 870,
    "ts": 7429770619788.006, "dur": 1.599,
    "args": {
      "External id": 3774,"Record function id": 0, "finished": true, "Ev Idx": 3773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770619788.333, "dur": 1.163,
    "args": {
      "External id": 3775,"Record function id": 0, "finished": true, "Ev Idx": 3774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770619789.899, "dur": 5.492,
    "args": {
      "External id": 3776,"Record function id": 0, "finished": true, "Ev Idx": 3775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770619796.054, "dur": 0.952,
    "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": 870, "tid": 870,
    "ts": 7429770619799.355, "dur": 2.586,
    "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": 870, "tid": 870,
    "ts": 7429770619801.015, "dur": 0.353,
    "args": {
      "External id": 3779,"Record function id": 0, "finished": true, "Ev Idx": 3778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770619803.409, "dur": 115.824,
    "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": 870, "tid": 870,
    "ts": 7429770619803.953, "dur": 2.187,
    "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": 870, "tid": 870,
    "ts": 7429770619805.344, "dur": 0.353,
    "args": {
      "External id": 3782,"Record function id": 0, "finished": true, "Ev Idx": 3781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770619806.499, "dur": 3.865,
    "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": 870, "tid": 870,
    "ts": 7429770619808.317, "dur": 1.882,
    "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": 870, "tid": 870,
    "ts": 7429770619810.907, "dur": 1.293,
    "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": 870, "tid": 870,
    "ts": 7429770619811.574, "dur": 0.266,
    "args": {
      "External id": 3786,"Record function id": 0, "finished": true, "Ev Idx": 3785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770619812.498, "dur": 1.772,
    "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": 870, "tid": 870,
    "ts": 7429770619813.323, "dur": 0.845,
    "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": 870, "tid": 870,
    "ts": 7429770619814.693, "dur": 102.060,
    "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": 870, "tid": 870,
    "ts": 7429770619893.766, "dur": 1.495,
    "args": {
      "External id": 3790,"Record function id": 0, "finished": true, "Ev Idx": 3789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770619917.533, "dur": 1.284,
    "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": 870, "tid": 870,
    "ts": 7429770619923.511, "dur": 11.617,
    "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": 870, "tid": 870,
    "ts": 7429770619938.789, "dur": 2.483,
    "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": 870, "tid": 870,
    "ts": 7429770619939.971, "dur": 0.491,
    "args": {
      "External id": 3794,"Record function id": 0, "finished": true, "Ev Idx": 3793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770619941.956, "dur": 1.293,
    "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": 870, "tid": 870,
    "ts": 7429770619942.494, "dur": 0.304,
    "args": {
      "External id": 3796,"Record function id": 0, "finished": true, "Ev Idx": 3795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770619943.903, "dur": 1.051,
    "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": 870, "tid": 870,
    "ts": 7429770619944.372, "dur": 0.238,
    "args": {
      "External id": 3798,"Record function id": 0, "finished": true, "Ev Idx": 3797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770619946.741, "dur": 10.252,
    "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": 870, "tid": 870,
    "ts": 7429770619962.894, "dur": 14.392,
    "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": 870, "tid": 870,
    "ts": 7429770619964.056, "dur": 0.330,
    "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": 870, "tid": 870,
    "ts": 7429770619965.078, "dur": 11.970,
    "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": 870, "tid": 870,
    "ts": 7429770619980.314, "dur": 0.198,
    "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": 870, "tid": 870,
    "ts": 7429770619991.867, "dur": 0.537,
    "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": 870, "tid": 870,
    "ts": 7429770619994.104, "dur": 44.405,
    "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": 870, "tid": 870,
    "ts": 7429770619995.254, "dur": 2.924,
    "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": 870, "tid": 870,
    "ts": 7429770619997.199, "dur": 0.391,
    "args": {
      "External id": 3807,"Record function id": 0, "finished": true, "Ev Idx": 3806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770619998.532, "dur": 3.485,
    "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": 870, "tid": 870,
    "ts": 7429770619999.043, "dur": 2.841,
    "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": 870, "tid": 870,
    "ts": 7429770620002.437, "dur": 1.381,
    "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": 870, "tid": 870,
    "ts": 7429770620003.148, "dur": 0.282,
    "args": {
      "External id": 3811,"Record function id": 0, "finished": true, "Ev Idx": 3810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770620004.148, "dur": 1.299,
    "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": 870, "tid": 870,
    "ts": 7429770620004.423, "dur": 0.941,
    "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": 870, "tid": 870,
    "ts": 7429770620005.904, "dur": 30.661,
    "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": 870, "tid": 870,
    "ts": 7429770620017.677, "dur": 1.226,
    "args": {
      "External id": 3815,"Record function id": 0, "finished": true, "Ev Idx": 3814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770620037.189, "dur": 1.021,
    "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": 870, "tid": 870,
    "ts": 7429770620040.693, "dur": 1.868,
    "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": 870, "tid": 870,
    "ts": 7429770620041.578, "dur": 0.310,
    "args": {
      "External id": 3818,"Record function id": 0, "finished": true, "Ev Idx": 3817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770620044.999, "dur": 12.715,
    "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": 870, "tid": 870,
    "ts": 7429770620045.642, "dur": 11.932,
    "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": 870, "tid": 870,
    "ts": 7429770620046.827, "dur": 2.364,
    "args": {
      "External id": 3821,"Record function id": 0, "finished": true, "Ev Idx": 3820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770620047.373, "dur": 1.657,
    "args": {
      "External id": 3822,"Record function id": 0, "finished": true, "Ev Idx": 3821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770620049.619, "dur": 7.459,
    "args": {
      "External id": 3823,"Record function id": 0, "finished": true, "Ev Idx": 3822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770620063.829, "dur": 1.688,
    "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": 870, "tid": 870,
    "ts": 7429770620064.167, "dur": 1.256,
    "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": 870, "tid": 870,
    "ts": 7429770620080.896, "dur": 39.017,
    "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": 870, "tid": 870,
    "ts": 7429770620081.655, "dur": 2.853,
    "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": 870, "tid": 870,
    "ts": 7429770620083.004, "dur": 0.942,
    "args": {
      "External id": 3828,"Record function id": 0, "finished": true, "Ev Idx": 3827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770620083.467, "dur": 0.344,
    "args": {
      "External id": 3829,"Record function id": 0, "finished": true, "Ev Idx": 3828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770620084.907, "dur": 34.841,
    "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": 870, "tid": 870,
    "ts": 7429770620085.366, "dur": 1.419,
    "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": 870, "tid": 870,
    "ts": 7429770620085.636, "dur": 1.051,
    "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": 870, "tid": 870,
    "ts": 7429770620087.264, "dur": 30.462,
    "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": 870, "tid": 870,
    "ts": 7429770620101.443, "dur": 1.172,
    "args": {
      "External id": 3834,"Record function id": 0, "finished": true, "Ev Idx": 3833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770620118.429, "dur": 1.015,
    "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": 870, "tid": 870,
    "ts": 7429770620125.463, "dur": 8.364,
    "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": 870, "tid": 870,
    "ts": 7429770620144.284, "dur": 0.166,
    "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": 870, "tid": 870,
    "ts": 7429770620146.726, "dur": 10.746,
    "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": 870, "tid": 870,
    "ts": 7429770620148.730, "dur": 0.404,
    "args": {
      "External id": 3839,"Record function id": 0, "finished": true, "Ev Idx": 3838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770620149.667, "dur": 0.085,
    "args": {
      "External id": 3840,"Record function id": 0, "finished": true, "Ev Idx": 3839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770620160.084, "dur": 14.685,
    "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": 870, "tid": 870,
    "ts": 7429770620178.113, "dur": 8.519,
    "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": 870, "tid": 870,
    "ts": 7429770620188.610, "dur": 9.241,
    "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": 870, "tid": 870,
    "ts": 7429770620199.500, "dur": 8.892,
    "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": 870, "tid": 870,
    "ts": 7429770620210.857, "dur": 0.119,
    "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": 870, "tid": 870,
    "ts": 7429770620211.942, "dur": 10.399,
    "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": 870, "tid": 870,
    "ts": 7429770620236.844, "dur": 39.360,
    "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": 870, "tid": 870,
    "ts": 7429770620237.449, "dur": 2.797,
    "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": 870, "tid": 870,
    "ts": 7429770620238.333, "dur": 1.262,
    "args": {
      "External id": 3849,"Record function id": 0, "finished": true, "Ev Idx": 3848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770620238.974, "dur": 0.429,
    "args": {
      "External id": 3850,"Record function id": 0, "finished": true, "Ev Idx": 3849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770620240.741, "dur": 35.318,
    "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": 870, "tid": 870,
    "ts": 7429770620241.214, "dur": 1.791,
    "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": 870, "tid": 870,
    "ts": 7429770620241.509, "dur": 1.385,
    "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": 870, "tid": 870,
    "ts": 7429770620243.378, "dur": 30.696,
    "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": 870, "tid": 870,
    "ts": 7429770620255.886, "dur": 1.179,
    "args": {
      "External id": 3855,"Record function id": 0, "finished": true, "Ev Idx": 3854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770620274.696, "dur": 1.089,
    "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": 870, "tid": 870,
    "ts": 7429770620284.146, "dur": 10.629,
    "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": 870, "tid": 870,
    "ts": 7429770620302.050, "dur": 36.091,
    "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": 870, "tid": 870,
    "ts": 7429770620302.534, "dur": 2.350,
    "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": 870, "tid": 870,
    "ts": 7429770620303.380, "dur": 1.016,
    "args": {
      "External id": 3860,"Record function id": 0, "finished": true, "Ev Idx": 3859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770620303.898, "dur": 0.385,
    "args": {
      "External id": 3861,"Record function id": 0, "finished": true, "Ev Idx": 3860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770620305.269, "dur": 32.748,
    "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": 870, "tid": 870,
    "ts": 7429770620305.860, "dur": 2.409,
    "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": 870, "tid": 870,
    "ts": 7429770620306.376, "dur": 1.770,
    "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": 870, "tid": 870,
    "ts": 7429770620308.718, "dur": 27.292,
    "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": 870, "tid": 870,
    "ts": 7429770620318.481, "dur": 1.402,
    "args": {
      "External id": 3866,"Record function id": 0, "finished": true, "Ev Idx": 3865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770620336.739, "dur": 1.034,
    "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": 870, "tid": 870,
    "ts": 7429770620339.924, "dur": 8.117,
    "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": 870, "tid": 870,
    "ts": 7429770620353.888, "dur": 47.240,
    "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": 870, "tid": 870,
    "ts": 7429770620354.243, "dur": 2.329,
    "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": 870, "tid": 870,
    "ts": 7429770620355.072, "dur": 1.005,
    "args": {
      "External id": 3871,"Record function id": 0, "finished": true, "Ev Idx": 3870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770620355.650, "dur": 0.311,
    "args": {
      "External id": 3872,"Record function id": 0, "finished": true, "Ev Idx": 3871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770620357.017, "dur": 43.994,
    "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": 870, "tid": 870,
    "ts": 7429770620357.576, "dur": 2.461,
    "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": 870, "tid": 870,
    "ts": 7429770620358.257, "dur": 1.667,
    "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": 870, "tid": 870,
    "ts": 7429770620360.484, "dur": 38.781,
    "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": 870, "tid": 870,
    "ts": 7429770620369.324, "dur": 1.046,
    "args": {
      "External id": 3877,"Record function id": 0, "finished": true, "Ev Idx": 3876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770620399.880, "dur": 0.902,
    "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": 870, "tid": 870,
    "ts": 7429770620403.507, "dur": 8.098,
    "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": 870, "tid": 870,
    "ts": 7429770620429.132, "dur": 0.178,
    "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": 870, "tid": 870,
    "ts": 7429770620430.959, "dur": 8.277,
    "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": 870, "tid": 870,
    "ts": 7429770620432.254, "dur": 0.285,
    "args": {
      "External id": 3882,"Record function id": 0, "finished": true, "Ev Idx": 3881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770620432.987, "dur": 0.093,
    "args": {
      "External id": 3883,"Record function id": 0, "finished": true, "Ev Idx": 3882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770620441.439, "dur": 11.327,
    "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": 870, "tid": 870,
    "ts": 7429770620455.876, "dur": 8.016,
    "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": 870, "tid": 870,
    "ts": 7429770620465.394, "dur": 7.336,
    "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": 870, "tid": 870,
    "ts": 7429770620474.086, "dur": 8.903,
    "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": 870, "tid": 870,
    "ts": 7429770620485.163, "dur": 0.130,
    "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": 870, "tid": 870,
    "ts": 7429770620485.973, "dur": 9.597,
    "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": 870, "tid": 870,
    "ts": 7429770620512.581, "dur": 33.096,
    "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": 870, "tid": 870,
    "ts": 7429770620513.066, "dur": 2.682,
    "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": 870, "tid": 870,
    "ts": 7429770620514.082, "dur": 1.070,
    "args": {
      "External id": 3892,"Record function id": 0, "finished": true, "Ev Idx": 3891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770620514.530, "dur": 0.445,
    "args": {
      "External id": 3893,"Record function id": 0, "finished": true, "Ev Idx": 3892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770620516.144, "dur": 29.416,
    "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": 870, "tid": 870,
    "ts": 7429770620516.643, "dur": 1.433,
    "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": 870, "tid": 870,
    "ts": 7429770620516.944, "dur": 1.016,
    "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": 870, "tid": 870,
    "ts": 7429770620518.401, "dur": 25.345,
    "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": 870, "tid": 870,
    "ts": 7429770620527.937, "dur": 1.092,
    "args": {
      "External id": 3898,"Record function id": 0, "finished": true, "Ev Idx": 3897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770620544.344, "dur": 0.964,
    "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": 870, "tid": 870,
    "ts": 7429770620548.667, "dur": 1.281,
    "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": 870, "tid": 870,
    "ts": 7429770620551.525, "dur": 2.311,
    "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": 870, "tid": 870,
    "ts": 7429770620552.659, "dur": 0.483,
    "args": {
      "External id": 3902,"Record function id": 0, "finished": true, "Ev Idx": 3901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770620560.519, "dur": 32.872,
    "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": 870, "tid": 870,
    "ts": 7429770620560.840, "dur": 1.846,
    "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": 870, "tid": 870,
    "ts": 7429770620561.547, "dur": 0.732,
    "args": {
      "External id": 3905,"Record function id": 0, "finished": true, "Ev Idx": 3904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770620561.922, "dur": 0.245,
    "args": {
      "External id": 3906,"Record function id": 0, "finished": true, "Ev Idx": 3905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770620562.983, "dur": 30.331,
    "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": 870, "tid": 870,
    "ts": 7429770620563.404, "dur": 2.108,
    "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": 870, "tid": 870,
    "ts": 7429770620563.660, "dur": 1.758,
    "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": 870, "tid": 870,
    "ts": 7429770620565.803, "dur": 25.808,
    "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": 870, "tid": 870,
    "ts": 7429770620575.308, "dur": 1.084,
    "args": {
      "External id": 3911,"Record function id": 0, "finished": true, "Ev Idx": 3910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770620592.137, "dur": 0.947,
    "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": 870, "tid": 870,
    "ts": 7429770620595.879, "dur": 1.132,
    "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": 870, "tid": 870,
    "ts": 7429770620598.316, "dur": 1.597,
    "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": 870, "tid": 870,
    "ts": 7429770620599.105, "dur": 0.351,
    "args": {
      "External id": 3915,"Record function id": 0, "finished": true, "Ev Idx": 3914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770620607.974, "dur": 30.350,
    "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": 870, "tid": 870,
    "ts": 7429770620608.291, "dur": 2.451,
    "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": 870, "tid": 870,
    "ts": 7429770620609.438, "dur": 0.917,
    "args": {
      "External id": 3918,"Record function id": 0, "finished": true, "Ev Idx": 3917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770620609.982, "dur": 0.252,
    "args": {
      "External id": 3919,"Record function id": 0, "finished": true, "Ev Idx": 3918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770620611.013, "dur": 27.229,
    "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": 870, "tid": 870,
    "ts": 7429770620611.417, "dur": 1.709,
    "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": 870, "tid": 870,
    "ts": 7429770620611.671, "dur": 1.350,
    "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": 870, "tid": 870,
    "ts": 7429770620613.419, "dur": 23.075,
    "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": 870, "tid": 870,
    "ts": 7429770620622.256, "dur": 1.004,
    "args": {
      "External id": 3924,"Record function id": 0, "finished": true, "Ev Idx": 3923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770620637.100, "dur": 0.897,
    "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": 870, "tid": 870,
    "ts": 7429770620640.938, "dur": 1.110,
    "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": 870, "tid": 870,
    "ts": 7429770620643.644, "dur": 1.618,
    "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": 870, "tid": 870,
    "ts": 7429770620644.491, "dur": 0.303,
    "args": {
      "External id": 3928,"Record function id": 0, "finished": true, "Ev Idx": 3927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770620648.021, "dur": 1.723,
    "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": 870, "tid": 870,
    "ts": 7429770620648.961, "dur": 0.271,
    "args": {
      "External id": 3930,"Record function id": 0, "finished": true, "Ev Idx": 3929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770620650.817, "dur": 1.036,
    "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": 870, "tid": 870,
    "ts": 7429770620651.260, "dur": 0.265,
    "args": {
      "External id": 3932,"Record function id": 0, "finished": true, "Ev Idx": 3931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770620652.833, "dur": 9.431,
    "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": 870, "tid": 870,
    "ts": 7429770620666.680, "dur": 3.166,
    "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": 870, "tid": 870,
    "ts": 7429770620668.398, "dur": 0.428,
    "args": {
      "External id": 3935,"Record function id": 0, "finished": true, "Ev Idx": 3934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770620672.061, "dur": 1.732,
    "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": 870, "tid": 870,
    "ts": 7429770620672.886, "dur": 0.328,
    "args": {
      "External id": 3937,"Record function id": 0, "finished": true, "Ev Idx": 3936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770620675.443, "dur": 10.095,
    "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": 870, "tid": 870,
    "ts": 7429770620689.571, "dur": 17.311,
    "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": 870, "tid": 870,
    "ts": 7429770620709.718, "dur": 8.343,
    "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": 870, "tid": 870,
    "ts": 7429770620719.925, "dur": 7.385,
    "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": 870, "tid": 870,
    "ts": 7429770620729.272, "dur": 6.728,
    "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": 870, "tid": 870,
    "ts": 7429770620738.176, "dur": 3.353,
    "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": 870, "tid": 870,
    "ts": 7429770620740.474, "dur": 0.415,
    "args": {
      "External id": 3944,"Record function id": 0, "finished": true, "Ev Idx": 3943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770620743.000, "dur": 2.012,
    "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": 870, "tid": 870,
    "ts": 7429770620744.174, "dur": 0.301,
    "args": {
      "External id": 3946,"Record function id": 0, "finished": true, "Ev Idx": 3945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770620745.872, "dur": 6.946,
    "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": 870, "tid": 870,
    "ts": 7429770620754.467, "dur": 11.792,
    "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": 870, "tid": 870,
    "ts": 7429770620768.791, "dur": 8.149,
    "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": 870, "tid": 870,
    "ts": 7429770620778.282, "dur": 6.514,
    "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": 870, "tid": 870,
    "ts": 7429770620806.961, "dur": 2.517,
    "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": 870, "tid": 870,
    "ts": 7429770620808.205, "dur": 0.429,
    "args": {
      "External id": 3952,"Record function id": 0, "finished": true, "Ev Idx": 3951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770620810.228, "dur": 1.817,
    "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": 870, "tid": 870,
    "ts": 7429770620811.190, "dur": 0.316,
    "args": {
      "External id": 3954,"Record function id": 0, "finished": true, "Ev Idx": 3953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770620813.006, "dur": 2.552,
    "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": 870, "tid": 870,
    "ts": 7429770620814.104, "dur": 0.886,
    "args": {
      "External id": 3956,"Record function id": 0, "finished": true, "Ev Idx": 3955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770620816.190, "dur": 3.061,
    "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": 870, "tid": 870,
    "ts": 7429770620818.294, "dur": 0.482,
    "args": {
      "External id": 3958,"Record function id": 0, "finished": true, "Ev Idx": 3957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770620819.875, "dur": 2.521,
    "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": 870, "tid": 870,
    "ts": 7429770620821.447, "dur": 0.502,
    "args": {
      "External id": 3960,"Record function id": 0, "finished": true, "Ev Idx": 3959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770620824.675, "dur": 2.624,
    "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": 870, "tid": 870,
    "ts": 7429770620826.246, "dur": 0.390,
    "args": {
      "External id": 3962,"Record function id": 0, "finished": true, "Ev Idx": 3961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770620829.453, "dur": 17.345,
    "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": 870, "tid": 870,
    "ts": 7429770620831.696, "dur": 13.058,
    "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": 870, "tid": 870,
    "ts": 7429770620832.965, "dur": 2.522,
    "args": {
      "External id": 3965,"Record function id": 0, "finished": true, "Ev Idx": 3964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770620833.744, "dur": 1.527,
    "args": {
      "External id": 3966,"Record function id": 0, "finished": true, "Ev Idx": 3965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770620836.057, "dur": 8.177,
    "args": {
      "External id": 3967,"Record function id": 0, "finished": true, "Ev Idx": 3966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770620845.296, "dur": 1.250,
    "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": 870, "tid": 870,
    "ts": 7429770620849.976, "dur": 2.121,
    "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": 870, "tid": 870,
    "ts": 7429770620851.053, "dur": 0.498,
    "args": {
      "External id": 3970,"Record function id": 0, "finished": true, "Ev Idx": 3969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770620852.752, "dur": 2.267,
    "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": 870, "tid": 870,
    "ts": 7429770620854.277, "dur": 0.276,
    "args": {
      "External id": 3972,"Record function id": 0, "finished": true, "Ev Idx": 3971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770620855.513, "dur": 2.006,
    "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": 870, "tid": 870,
    "ts": 7429770620856.802, "dur": 0.330,
    "args": {
      "External id": 3974,"Record function id": 0, "finished": true, "Ev Idx": 3973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770620858.082, "dur": 1.577,
    "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": 870, "tid": 870,
    "ts": 7429770620858.820, "dur": 0.311,
    "args": {
      "External id": 3976,"Record function id": 0, "finished": true, "Ev Idx": 3975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770620860.219, "dur": 1.576,
    "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": 870, "tid": 870,
    "ts": 7429770620861.053, "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": 870, "tid": 870,
    "ts": 7429770620863.297, "dur": 1.881,
    "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": 870, "tid": 870,
    "ts": 7429770620864.259, "dur": 0.481,
    "args": {
      "External id": 3980,"Record function id": 0, "finished": true, "Ev Idx": 3979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770620866.698, "dur": 13.328,
    "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": 870, "tid": 870,
    "ts": 7429770620867.606, "dur": 10.037,
    "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": 870, "tid": 870,
    "ts": 7429770620868.819, "dur": 1.754,
    "args": {
      "External id": 3983,"Record function id": 0, "finished": true, "Ev Idx": 3982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770620869.155, "dur": 1.322,
    "args": {
      "External id": 3984,"Record function id": 0, "finished": true, "Ev Idx": 3983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770620870.870, "dur": 6.440,
    "args": {
      "External id": 3985,"Record function id": 0, "finished": true, "Ev Idx": 3984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770620878.001, "dur": 1.863,
    "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": 870, "tid": 870,
    "ts": 7429770620881.974, "dur": 1.802,
    "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": 870, "tid": 870,
    "ts": 7429770620882.780, "dur": 0.472,
    "args": {
      "External id": 3988,"Record function id": 0, "finished": true, "Ev Idx": 3987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770620885.105, "dur": 108.533,
    "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": 870, "tid": 870,
    "ts": 7429770620885.662, "dur": 1.620,
    "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": 870, "tid": 870,
    "ts": 7429770620886.494, "dur": 0.331,
    "args": {
      "External id": 3991,"Record function id": 0, "finished": true, "Ev Idx": 3990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770620887.649, "dur": 3.470,
    "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": 870, "tid": 870,
    "ts": 7429770620889.286, "dur": 1.676,
    "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": 870, "tid": 870,
    "ts": 7429770620891.525, "dur": 1.898,
    "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": 870, "tid": 870,
    "ts": 7429770620892.610, "dur": 0.438,
    "args": {
      "External id": 3995,"Record function id": 0, "finished": true, "Ev Idx": 3994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770620893.744, "dur": 2.409,
    "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": 870, "tid": 870,
    "ts": 7429770620894.556, "dur": 1.499,
    "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": 870, "tid": 870,
    "ts": 7429770620896.587, "dur": 94.697,
    "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": 870, "tid": 870,
    "ts": 7429770620970.557, "dur": 1.494,
    "args": {
      "External id": 3999,"Record function id": 0, "finished": true, "Ev Idx": 3998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770620991.905, "dur": 1.358,
    "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": 870, "tid": 870,
    "ts": 7429770620997.719, "dur": 10.614,
    "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": 870, "tid": 870,
    "ts": 7429770621011.795, "dur": 2.344,
    "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": 870, "tid": 870,
    "ts": 7429770621012.861, "dur": 0.464,
    "args": {
      "External id": 4003,"Record function id": 0, "finished": true, "Ev Idx": 4002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770621014.884, "dur": 1.660,
    "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": 870, "tid": 870,
    "ts": 7429770621015.654, "dur": 0.408,
    "args": {
      "External id": 4005,"Record function id": 0, "finished": true, "Ev Idx": 4004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770621017.229, "dur": 1.325,
    "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": 870, "tid": 870,
    "ts": 7429770621017.854, "dur": 0.367,
    "args": {
      "External id": 4007,"Record function id": 0, "finished": true, "Ev Idx": 4006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770621020.174, "dur": 11.445,
    "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": 870, "tid": 870,
    "ts": 7429770621036.852, "dur": 13.252,
    "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": 870, "tid": 870,
    "ts": 7429770621037.927, "dur": 0.420,
    "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": 870, "tid": 870,
    "ts": 7429770621039.031, "dur": 10.839,
    "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": 870, "tid": 870,
    "ts": 7429770621053.187, "dur": 0.181,
    "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": 870, "tid": 870,
    "ts": 7429770621064.561, "dur": 0.479,
    "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": 870, "tid": 870,
    "ts": 7429770621066.773, "dur": 44.652,
    "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": 870, "tid": 870,
    "ts": 7429770621067.329, "dur": 3.259,
    "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": 870, "tid": 870,
    "ts": 7429770621069.682, "dur": 0.343,
    "args": {
      "External id": 4016,"Record function id": 0, "finished": true, "Ev Idx": 4015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770621070.984, "dur": 2.679,
    "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": 870, "tid": 870,
    "ts": 7429770621071.375, "dur": 2.170,
    "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": 870, "tid": 870,
    "ts": 7429770621074.114, "dur": 1.505,
    "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": 870, "tid": 870,
    "ts": 7429770621074.952, "dur": 0.296,
    "args": {
      "External id": 4020,"Record function id": 0, "finished": true, "Ev Idx": 4019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770621075.956, "dur": 1.534,
    "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": 870, "tid": 870,
    "ts": 7429770621076.377, "dur": 1.038,
    "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": 870, "tid": 870,
    "ts": 7429770621077.919, "dur": 31.552,
    "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": 870, "tid": 870,
    "ts": 7429770621091.339, "dur": 1.415,
    "args": {
      "External id": 4024,"Record function id": 0, "finished": true, "Ev Idx": 4023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770621110.006, "dur": 1.097,
    "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": 870, "tid": 870,
    "ts": 7429770621113.452, "dur": 2.054,
    "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": 870, "tid": 870,
    "ts": 7429770621114.591, "dur": 0.340,
    "args": {
      "External id": 4027,"Record function id": 0, "finished": true, "Ev Idx": 4026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770621118.123, "dur": 12.475,
    "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": 870, "tid": 870,
    "ts": 7429770621118.723, "dur": 11.730,
    "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": 870, "tid": 870,
    "ts": 7429770621119.835, "dur": 2.387,
    "args": {
      "External id": 4030,"Record function id": 0, "finished": true, "Ev Idx": 4029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770621120.305, "dur": 1.752,
    "args": {
      "External id": 4031,"Record function id": 0, "finished": true, "Ev Idx": 4030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770621122.622, "dur": 7.421,
    "args": {
      "External id": 4032,"Record function id": 0, "finished": true, "Ev Idx": 4031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770621136.648, "dur": 1.604,
    "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": 870, "tid": 870,
    "ts": 7429770621136.977, "dur": 1.177,
    "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": 870, "tid": 870,
    "ts": 7429770621152.795, "dur": 39.391,
    "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": 870, "tid": 870,
    "ts": 7429770621153.727, "dur": 2.874,
    "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": 870, "tid": 870,
    "ts": 7429770621155.039, "dur": 0.956,
    "args": {
      "External id": 4037,"Record function id": 0, "finished": true, "Ev Idx": 4036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770621155.505, "dur": 0.316,
    "args": {
      "External id": 4038,"Record function id": 0, "finished": true, "Ev Idx": 4037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770621156.983, "dur": 35.035,
    "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": 870, "tid": 870,
    "ts": 7429770621157.505, "dur": 2.519,
    "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": 870, "tid": 870,
    "ts": 7429770621157.899, "dur": 2.031,
    "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": 870, "tid": 870,
    "ts": 7429770621160.475, "dur": 29.564,
    "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": 870, "tid": 870,
    "ts": 7429770621172.523, "dur": 1.122,
    "args": {
      "External id": 4043,"Record function id": 0, "finished": true, "Ev Idx": 4042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770621190.711, "dur": 1.038,
    "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": 870, "tid": 870,
    "ts": 7429770621197.271, "dur": 8.576,
    "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": 870, "tid": 870,
    "ts": 7429770621214.553, "dur": 0.204,
    "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": 870, "tid": 870,
    "ts": 7429770621216.893, "dur": 11.932,
    "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": 870, "tid": 870,
    "ts": 7429770621218.690, "dur": 0.416,
    "args": {
      "External id": 4048,"Record function id": 0, "finished": true, "Ev Idx": 4047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770621219.422, "dur": 0.076,
    "args": {
      "External id": 4049,"Record function id": 0, "finished": true, "Ev Idx": 4048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770621231.598, "dur": 12.760,
    "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": 870, "tid": 870,
    "ts": 7429770621247.968, "dur": 7.825,
    "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": 870, "tid": 870,
    "ts": 7429770621257.765, "dur": 11.329,
    "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": 870, "tid": 870,
    "ts": 7429770621270.988, "dur": 8.778,
    "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": 870, "tid": 870,
    "ts": 7429770621282.334, "dur": 0.095,
    "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": 870, "tid": 870,
    "ts": 7429770621283.210, "dur": 8.009,
    "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": 870, "tid": 870,
    "ts": 7429770621305.599, "dur": 110.837,
    "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": 870, "tid": 870,
    "ts": 7429770621306.115, "dur": 4.241,
    "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": 870, "tid": 870,
    "ts": 7429770621308.498, "dur": 1.264,
    "args": {
      "External id": 4058,"Record function id": 0, "finished": true, "Ev Idx": 4057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770621309.136, "dur": 0.487,
    "args": {
      "External id": 4059,"Record function id": 0, "finished": true, "Ev Idx": 4058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770621310.826, "dur": 105.434,
    "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": 870, "tid": 870,
    "ts": 7429770621311.346, "dur": 2.207,
    "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": 870, "tid": 870,
    "ts": 7429770621311.815, "dur": 1.637,
    "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": 870, "tid": 870,
    "ts": 7429770621314.011, "dur": 99.593,
    "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": 870, "tid": 870,
    "ts": 7429770621326.009, "dur": 65.941,
    "args": {
      "External id": 4064,"Record function id": 0, "finished": true, "Ev Idx": 4063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770621414.455, "dur": 1.424,
    "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": 870, "tid": 870,
    "ts": 7429770621426.555, "dur": 11.154,
    "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": 870, "tid": 870,
    "ts": 7429770621445.942, "dur": 37.031,
    "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": 870, "tid": 870,
    "ts": 7429770621446.406, "dur": 3.555,
    "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": 870, "tid": 870,
    "ts": 7429770621447.672, "dur": 1.618,
    "args": {
      "External id": 4069,"Record function id": 0, "finished": true, "Ev Idx": 4068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770621448.683, "dur": 0.445,
    "args": {
      "External id": 4070,"Record function id": 0, "finished": true, "Ev Idx": 4069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770621450.337, "dur": 32.457,
    "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": 870, "tid": 870,
    "ts": 7429770621450.840, "dur": 2.703,
    "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": 870, "tid": 870,
    "ts": 7429770621451.185, "dur": 2.196,
    "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": 870, "tid": 870,
    "ts": 7429770621453.933, "dur": 27.065,
    "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": 870, "tid": 870,
    "ts": 7429770621465.014, "dur": 1.256,
    "args": {
      "External id": 4075,"Record function id": 0, "finished": true, "Ev Idx": 4074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770621481.587, "dur": 0.936,
    "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": 870, "tid": 870,
    "ts": 7429770621484.664, "dur": 8.288,
    "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": 870, "tid": 870,
    "ts": 7429770621499.032, "dur": 44.905,
    "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": 870, "tid": 870,
    "ts": 7429770621499.449, "dur": 2.408,
    "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": 870, "tid": 870,
    "ts": 7429770621500.350, "dur": 1.020,
    "args": {
      "External id": 4080,"Record function id": 0, "finished": true, "Ev Idx": 4079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770621500.854, "dur": 0.375,
    "args": {
      "External id": 4081,"Record function id": 0, "finished": true, "Ev Idx": 4080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770621502.156, "dur": 41.662,
    "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": 870, "tid": 870,
    "ts": 7429770621502.619, "dur": 1.330,
    "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": 870, "tid": 870,
    "ts": 7429770621502.878, "dur": 0.949,
    "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": 870, "tid": 870,
    "ts": 7429770621504.271, "dur": 37.769,
    "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": 870, "tid": 870,
    "ts": 7429770621512.692, "dur": 1.169,
    "args": {
      "External id": 4086,"Record function id": 0, "finished": true, "Ev Idx": 4085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770621542.640, "dur": 0.912,
    "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": 870, "tid": 870,
    "ts": 7429770621546.954, "dur": 8.990,
    "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": 870, "tid": 870,
    "ts": 7429770621575.068, "dur": 0.193,
    "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": 870, "tid": 870,
    "ts": 7429770621577.217, "dur": 9.985,
    "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": 870, "tid": 870,
    "ts": 7429770621578.768, "dur": 0.353,
    "args": {
      "External id": 4091,"Record function id": 0, "finished": true, "Ev Idx": 4090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770621579.527, "dur": 0.083,
    "args": {
      "External id": 4092,"Record function id": 0, "finished": true, "Ev Idx": 4091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770621589.777, "dur": 12.195,
    "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": 870, "tid": 870,
    "ts": 7429770621605.712, "dur": 8.420,
    "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": 870, "tid": 870,
    "ts": 7429770621615.866, "dur": 8.302,
    "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": 870, "tid": 870,
    "ts": 7429770621625.673, "dur": 9.286,
    "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": 870, "tid": 870,
    "ts": 7429770621675.394, "dur": 0.085,
    "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": 870, "tid": 870,
    "ts": 7429770621676.269, "dur": 9.852,
    "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": 870, "tid": 870,
    "ts": 7429770621705.186, "dur": 39.212,
    "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": 870, "tid": 870,
    "ts": 7429770621705.547, "dur": 4.024,
    "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": 870, "tid": 870,
    "ts": 7429770621707.691, "dur": 1.288,
    "args": {
      "External id": 4101,"Record function id": 0, "finished": true, "Ev Idx": 4100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770621708.278, "dur": 0.558,
    "args": {
      "External id": 4102,"Record function id": 0, "finished": true, "Ev Idx": 4101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770621709.901, "dur": 34.381,
    "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": 870, "tid": 870,
    "ts": 7429770621710.382, "dur": 1.428,
    "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": 870, "tid": 870,
    "ts": 7429770621710.680, "dur": 1.000,
    "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": 870, "tid": 870,
    "ts": 7429770621712.173, "dur": 30.160,
    "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": 870, "tid": 870,
    "ts": 7429770621725.837, "dur": 1.226,
    "args": {
      "External id": 4107,"Record function id": 0, "finished": true, "Ev Idx": 4106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770621742.902, "dur": 1.095,
    "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": 870, "tid": 870,
    "ts": 7429770621747.718, "dur": 1.518,
    "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": 870, "tid": 870,
    "ts": 7429770621751.129, "dur": 2.406,
    "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": 870, "tid": 870,
    "ts": 7429770621752.170, "dur": 0.571,
    "args": {
      "External id": 4111,"Record function id": 0, "finished": true, "Ev Idx": 4110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770621761.471, "dur": 35.144,
    "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": 870, "tid": 870,
    "ts": 7429770621761.771, "dur": 2.014,
    "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": 870, "tid": 870,
    "ts": 7429770621762.575, "dur": 0.766,
    "args": {
      "External id": 4114,"Record function id": 0, "finished": true, "Ev Idx": 4113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770621762.949, "dur": 0.279,
    "args": {
      "External id": 4115,"Record function id": 0, "finished": true, "Ev Idx": 4114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770621764.085, "dur": 32.450,
    "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": 870, "tid": 870,
    "ts": 7429770621764.471, "dur": 2.042,
    "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": 870, "tid": 870,
    "ts": 7429770621764.734, "dur": 1.671,
    "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": 870, "tid": 870,
    "ts": 7429770621766.822, "dur": 27.950,
    "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": 870, "tid": 870,
    "ts": 7429770621778.033, "dur": 1.125,
    "args": {
      "External id": 4120,"Record function id": 0, "finished": true, "Ev Idx": 4119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770621795.351, "dur": 0.913,
    "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": 870, "tid": 870,
    "ts": 7429770621798.914, "dur": 1.249,
    "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": 870, "tid": 870,
    "ts": 7429770621801.631, "dur": 1.608,
    "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": 870, "tid": 870,
    "ts": 7429770621802.405, "dur": 0.367,
    "args": {
      "External id": 4124,"Record function id": 0, "finished": true, "Ev Idx": 4123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770621824.426, "dur": 28.154,
    "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": 870, "tid": 870,
    "ts": 7429770621824.809, "dur": 2.081,
    "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": 870, "tid": 870,
    "ts": 7429770621825.677, "dur": 0.789,
    "args": {
      "External id": 4127,"Record function id": 0, "finished": true, "Ev Idx": 4126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770621826.075, "dur": 0.249,
    "args": {
      "External id": 4128,"Record function id": 0, "finished": true, "Ev Idx": 4127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770621827.163, "dur": 25.321,
    "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": 870, "tid": 870,
    "ts": 7429770621827.558, "dur": 1.220,
    "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": 870, "tid": 870,
    "ts": 7429770621827.815, "dur": 0.851,
    "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": 870, "tid": 870,
    "ts": 7429770621829.061, "dur": 21.678,
    "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": 870, "tid": 870,
    "ts": 7429770621836.590, "dur": 1.146,
    "args": {
      "External id": 4133,"Record function id": 0, "finished": true, "Ev Idx": 4132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770621851.303, "dur": 0.924,
    "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": 870, "tid": 870,
    "ts": 7429770621854.908, "dur": 1.141,
    "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": 870, "tid": 870,
    "ts": 7429770621857.691, "dur": 1.718,
    "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": 870, "tid": 870,
    "ts": 7429770621858.425, "dur": 0.422,
    "args": {
      "External id": 4137,"Record function id": 0, "finished": true, "Ev Idx": 4136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770621862.452, "dur": 2.227,
    "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": 870, "tid": 870,
    "ts": 7429770621863.868, "dur": 0.283,
    "args": {
      "External id": 4139,"Record function id": 0, "finished": true, "Ev Idx": 4138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770621865.765, "dur": 1.210,
    "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": 870, "tid": 870,
    "ts": 7429770621866.181, "dur": 0.290,
    "args": {
      "External id": 4141,"Record function id": 0, "finished": true, "Ev Idx": 4140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770621867.921, "dur": 12.434,
    "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": 870, "tid": 870,
    "ts": 7429770621885.929, "dur": 4.801,
    "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": 870, "tid": 870,
    "ts": 7429770621888.574, "dur": 0.615,
    "args": {
      "External id": 4144,"Record function id": 0, "finished": true, "Ev Idx": 4143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770621893.059, "dur": 1.823,
    "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": 870, "tid": 870,
    "ts": 7429770621893.965, "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": 870, "tid": 870,
    "ts": 7429770621896.712, "dur": 9.866,
    "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": 870, "tid": 870,
    "ts": 7429770621911.452, "dur": 17.698,
    "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": 870, "tid": 870,
    "ts": 7429770621932.519, "dur": 7.520,
    "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": 870, "tid": 870,
    "ts": 7429770621941.675, "dur": 7.939,
    "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": 870, "tid": 870,
    "ts": 7429770621951.734, "dur": 7.844,
    "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": 870, "tid": 870,
    "ts": 7429770621961.880, "dur": 3.560,
    "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": 870, "tid": 870,
    "ts": 7429770621964.242, "dur": 0.510,
    "args": {
      "External id": 4153,"Record function id": 0, "finished": true, "Ev Idx": 4152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770621966.975, "dur": 2.292,
    "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": 870, "tid": 870,
    "ts": 7429770621968.435, "dur": 0.293,
    "args": {
      "External id": 4155,"Record function id": 0, "finished": true, "Ev Idx": 4154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770621970.140, "dur": 6.935,
    "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": 870, "tid": 870,
    "ts": 7429770621978.888, "dur": 11.781,
    "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": 870, "tid": 870,
    "ts": 7429770621992.967, "dur": 8.802,
    "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": 870, "tid": 870,
    "ts": 7429770622003.017, "dur": 8.154,
    "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": 870, "tid": 870,
    "ts": 7429770622035.945, "dur": 3.337,
    "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": 870, "tid": 870,
    "ts": 7429770622037.981, "dur": 0.488,
    "args": {
      "External id": 4161,"Record function id": 0, "finished": true, "Ev Idx": 4160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770622040.007, "dur": 2.496,
    "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": 870, "tid": 870,
    "ts": 7429770622041.604, "dur": 0.341,
    "args": {
      "External id": 4163,"Record function id": 0, "finished": true, "Ev Idx": 4162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770622043.242, "dur": 2.486,
    "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": 870, "tid": 870,
    "ts": 7429770622044.239, "dur": 0.793,
    "args": {
      "External id": 4165,"Record function id": 0, "finished": true, "Ev Idx": 4164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770622046.403, "dur": 1.647,
    "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": 870, "tid": 870,
    "ts": 7429770622047.161, "dur": 0.389,
    "args": {
      "External id": 4167,"Record function id": 0, "finished": true, "Ev Idx": 4166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770622048.669, "dur": 2.204,
    "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": 870, "tid": 870,
    "ts": 7429770622050.102, "dur": 0.299,
    "args": {
      "External id": 4169,"Record function id": 0, "finished": true, "Ev Idx": 4168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770622053.932, "dur": 3.427,
    "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": 870, "tid": 870,
    "ts": 7429770622056.242, "dur": 0.421,
    "args": {
      "External id": 4171,"Record function id": 0, "finished": true, "Ev Idx": 4170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770622059.822, "dur": 41.500,
    "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": 870, "tid": 870,
    "ts": 7429770622062.473, "dur": 35.400,
    "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": 870, "tid": 870,
    "ts": 7429770622063.942, "dur": 2.897,
    "args": {
      "External id": 4174,"Record function id": 0, "finished": true, "Ev Idx": 4173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770622064.864, "dur": 1.736,
    "args": {
      "External id": 4175,"Record function id": 0, "finished": true, "Ev Idx": 4174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770622067.573, "dur": 29.106,
    "args": {
      "External id": 4176,"Record function id": 0, "finished": true, "Ev Idx": 4175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770622098.967, "dur": 1.938,
    "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": 870, "tid": 870,
    "ts": 7429770622106.203, "dur": 2.342,
    "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": 870, "tid": 870,
    "ts": 7429770622107.318, "dur": 0.542,
    "args": {
      "External id": 4179,"Record function id": 0, "finished": true, "Ev Idx": 4178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770622109.161, "dur": 2.210,
    "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": 870, "tid": 870,
    "ts": 7429770622110.604, "dur": 0.299,
    "args": {
      "External id": 4181,"Record function id": 0, "finished": true, "Ev Idx": 4180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770622111.926, "dur": 2.276,
    "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": 870, "tid": 870,
    "ts": 7429770622113.386, "dur": 0.376,
    "args": {
      "External id": 4183,"Record function id": 0, "finished": true, "Ev Idx": 4182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770622114.772, "dur": 2.102,
    "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": 870, "tid": 870,
    "ts": 7429770622115.997, "dur": 0.327,
    "args": {
      "External id": 4185,"Record function id": 0, "finished": true, "Ev Idx": 4184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770622117.428, "dur": 1.443,
    "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": 870, "tid": 870,
    "ts": 7429770622118.128, "dur": 0.284,
    "args": {
      "External id": 4187,"Record function id": 0, "finished": true, "Ev Idx": 4186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770622120.894, "dur": 2.075,
    "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": 870, "tid": 870,
    "ts": 7429770622122.165, "dur": 0.321,
    "args": {
      "External id": 4189,"Record function id": 0, "finished": true, "Ev Idx": 4188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770622124.614, "dur": 14.665,
    "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": 870, "tid": 870,
    "ts": 7429770622125.895, "dur": 11.274,
    "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": 870, "tid": 870,
    "ts": 7429770622126.774, "dur": 2.457,
    "args": {
      "External id": 4192,"Record function id": 0, "finished": true, "Ev Idx": 4191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770622127.230, "dur": 1.817,
    "args": {
      "External id": 4193,"Record function id": 0, "finished": true, "Ev Idx": 4192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770622129.560, "dur": 7.290,
    "args": {
      "External id": 4194,"Record function id": 0, "finished": true, "Ev Idx": 4193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770622137.540, "dur": 1.535,
    "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": 870, "tid": 870,
    "ts": 7429770622141.675, "dur": 2.070,
    "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": 870, "tid": 870,
    "ts": 7429770622142.787, "dur": 0.406,
    "args": {
      "External id": 4197,"Record function id": 0, "finished": true, "Ev Idx": 4196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770622145.684, "dur": 115.774,
    "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": 870, "tid": 870,
    "ts": 7429770622146.377, "dur": 1.690,
    "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": 870, "tid": 870,
    "ts": 7429770622147.260, "dur": 0.337,
    "args": {
      "External id": 4200,"Record function id": 0, "finished": true, "Ev Idx": 4199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770622148.628, "dur": 4.284,
    "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": 870, "tid": 870,
    "ts": 7429770622150.627, "dur": 2.133,
    "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": 870, "tid": 870,
    "ts": 7429770622153.412, "dur": 1.403,
    "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": 870, "tid": 870,
    "ts": 7429770622154.169, "dur": 0.284,
    "args": {
      "External id": 4204,"Record function id": 0, "finished": true, "Ev Idx": 4203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770622155.099, "dur": 2.740,
    "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": 870, "tid": 870,
    "ts": 7429770622155.953, "dur": 1.797,
    "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": 870, "tid": 870,
    "ts": 7429770622158.560, "dur": 100.251,
    "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": 870, "tid": 870,
    "ts": 7429770622234.270, "dur": 1.422,
    "args": {
      "External id": 4208,"Record function id": 0, "finished": true, "Ev Idx": 4207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770622259.706, "dur": 1.327,
    "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": 870, "tid": 870,
    "ts": 7429770622266.315, "dur": 11.720,
    "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": 870, "tid": 870,
    "ts": 7429770622281.763, "dur": 2.556,
    "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": 870, "tid": 870,
    "ts": 7429770622282.794, "dur": 0.599,
    "args": {
      "External id": 4212,"Record function id": 0, "finished": true, "Ev Idx": 4211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770622285.175, "dur": 1.703,
    "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": 870, "tid": 870,
    "ts": 7429770622286.151, "dur": 0.294,
    "args": {
      "External id": 4214,"Record function id": 0, "finished": true, "Ev Idx": 4213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770622287.634, "dur": 1.042,
    "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": 870, "tid": 870,
    "ts": 7429770622288.082, "dur": 0.242,
    "args": {
      "External id": 4216,"Record function id": 0, "finished": true, "Ev Idx": 4215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770622290.468, "dur": 9.403,
    "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": 870, "tid": 870,
    "ts": 7429770622311.203, "dur": 18.109,
    "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": 870, "tid": 870,
    "ts": 7429770622312.753, "dur": 0.438,
    "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": 870, "tid": 870,
    "ts": 7429770622314.145, "dur": 14.854,
    "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": 870, "tid": 870,
    "ts": 7429770622333.437, "dur": 0.266,
    "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": 870, "tid": 870,
    "ts": 7429770622347.272, "dur": 0.538,
    "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": 870, "tid": 870,
    "ts": 7429770622349.776, "dur": 47.551,
    "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": 870, "tid": 870,
    "ts": 7429770622350.507, "dur": 2.690,
    "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": 870, "tid": 870,
    "ts": 7429770622352.021, "dur": 0.506,
    "args": {
      "External id": 4225,"Record function id": 0, "finished": true, "Ev Idx": 4224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770622353.586, "dur": 3.861,
    "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": 870, "tid": 870,
    "ts": 7429770622354.367, "dur": 2.979,
    "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": 870, "tid": 870,
    "ts": 7429770622357.921, "dur": 1.732,
    "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": 870, "tid": 870,
    "ts": 7429770622359.001, "dur": 0.277,
    "args": {
      "External id": 4229,"Record function id": 0, "finished": true, "Ev Idx": 4228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770622359.988, "dur": 1.237,
    "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": 870, "tid": 870,
    "ts": 7429770622360.265, "dur": 0.875,
    "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": 870, "tid": 870,
    "ts": 7429770622361.712, "dur": 33.531,
    "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": 870, "tid": 870,
    "ts": 7429770622374.591, "dur": 1.643,
    "args": {
      "External id": 4233,"Record function id": 0, "finished": true, "Ev Idx": 4232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770622395.826, "dur": 1.182,
    "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": 870, "tid": 870,
    "ts": 7429770622399.566, "dur": 2.133,
    "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": 870, "tid": 870,
    "ts": 7429770622400.701, "dur": 0.397,
    "args": {
      "External id": 4236,"Record function id": 0, "finished": true, "Ev Idx": 4235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770622404.497, "dur": 15.103,
    "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": 870, "tid": 870,
    "ts": 7429770622405.323, "dur": 14.133,
    "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": 870, "tid": 870,
    "ts": 7429770622406.676, "dur": 3.641,
    "args": {
      "External id": 4239,"Record function id": 0, "finished": true, "Ev Idx": 4238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770622407.341, "dur": 2.789,
    "args": {
      "External id": 4240,"Record function id": 0, "finished": true, "Ev Idx": 4239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770622410.766, "dur": 8.185,
    "args": {
      "External id": 4241,"Record function id": 0, "finished": true, "Ev Idx": 4240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770622426.458, "dur": 1.555,
    "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": 870, "tid": 870,
    "ts": 7429770622426.744, "dur": 1.173,
    "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": 870, "tid": 870,
    "ts": 7429770622445.342, "dur": 40.228,
    "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": 870, "tid": 870,
    "ts": 7429770622446.598, "dur": 4.733,
    "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": 870, "tid": 870,
    "ts": 7429770622449.236, "dur": 1.359,
    "args": {
      "External id": 4246,"Record function id": 0, "finished": true, "Ev Idx": 4245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770622450.121, "dur": 0.357,
    "args": {
      "External id": 4247,"Record function id": 0, "finished": true, "Ev Idx": 4246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770622451.874, "dur": 33.540,
    "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": 870, "tid": 870,
    "ts": 7429770622452.464, "dur": 1.892,
    "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": 870, "tid": 870,
    "ts": 7429770622452.881, "dur": 1.383,
    "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": 870, "tid": 870,
    "ts": 7429770622454.899, "dur": 28.340,
    "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": 870, "tid": 870,
    "ts": 7429770622466.340, "dur": 1.180,
    "args": {
      "External id": 4252,"Record function id": 0, "finished": true, "Ev Idx": 4251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770622483.925, "dur": 1.193,
    "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": 870, "tid": 870,
    "ts": 7429770622491.513, "dur": 9.413,
    "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": 870, "tid": 870,
    "ts": 7429770622510.426, "dur": 0.173,
    "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": 870, "tid": 870,
    "ts": 7429770622513.311, "dur": 13.585,
    "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": 870, "tid": 870,
    "ts": 7429770622515.779, "dur": 0.610,
    "args": {
      "External id": 4257,"Record function id": 0, "finished": true, "Ev Idx": 4256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770622516.909, "dur": 0.095,
    "args": {
      "External id": 4258,"Record function id": 0, "finished": true, "Ev Idx": 4257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770622530.110, "dur": 15.522,
    "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": 870, "tid": 870,
    "ts": 7429770622549.433, "dur": 9.538,
    "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": 870, "tid": 870,
    "ts": 7429770622561.017, "dur": 10.271,
    "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": 870, "tid": 870,
    "ts": 7429770622573.125, "dur": 10.829,
    "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": 870, "tid": 870,
    "ts": 7429770622586.509, "dur": 0.105,
    "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": 870, "tid": 870,
    "ts": 7429770622587.512, "dur": 9.634,
    "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": 870, "tid": 870,
    "ts": 7429770622612.253, "dur": 40.892,
    "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": 870, "tid": 870,
    "ts": 7429770622612.800, "dur": 4.703,
    "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": 870, "tid": 870,
    "ts": 7429770622615.379, "dur": 1.322,
    "args": {
      "External id": 4267,"Record function id": 0, "finished": true, "Ev Idx": 4266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770622616.010, "dur": 0.544,
    "args": {
      "External id": 4268,"Record function id": 0, "finished": true, "Ev Idx": 4267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770622617.895, "dur": 35.091,
    "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": 870, "tid": 870,
    "ts": 7429770622618.366, "dur": 1.782,
    "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": 870, "tid": 870,
    "ts": 7429770622618.838, "dur": 1.198,
    "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": 870, "tid": 870,
    "ts": 7429770622620.505, "dur": 30.390,
    "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": 870, "tid": 870,
    "ts": 7429770622631.774, "dur": 1.366,
    "args": {
      "External id": 4273,"Record function id": 0, "finished": true, "Ev Idx": 4272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770622651.515, "dur": 1.159,
    "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": 870, "tid": 870,
    "ts": 7429770622660.800, "dur": 10.353,
    "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": 870, "tid": 870,
    "ts": 7429770622679.864, "dur": 39.538,
    "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": 870, "tid": 870,
    "ts": 7429770622680.512, "dur": 3.078,
    "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": 870, "tid": 870,
    "ts": 7429770622681.577, "dur": 1.421,
    "args": {
      "External id": 4278,"Record function id": 0, "finished": true, "Ev Idx": 4277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770622682.306, "dur": 0.571,
    "args": {
      "External id": 4279,"Record function id": 0, "finished": true, "Ev Idx": 4278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770622684.128, "dur": 35.154,
    "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": 870, "tid": 870,
    "ts": 7429770622684.728, "dur": 1.803,
    "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": 870, "tid": 870,
    "ts": 7429770622685.110, "dur": 1.336,
    "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": 870, "tid": 870,
    "ts": 7429770622686.845, "dur": 30.483,
    "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": 870, "tid": 870,
    "ts": 7429770622699.500, "dur": 1.221,
    "args": {
      "External id": 4284,"Record function id": 0, "finished": true, "Ev Idx": 4283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770622717.937, "dur": 1.094,
    "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": 870, "tid": 870,
    "ts": 7429770622720.990, "dur": 8.745,
    "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": 870, "tid": 870,
    "ts": 7429770622736.075, "dur": 48.532,
    "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": 870, "tid": 870,
    "ts": 7429770622736.628, "dur": 3.489,
    "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": 870, "tid": 870,
    "ts": 7429770622738.462, "dur": 1.200,
    "args": {
      "External id": 4289,"Record function id": 0, "finished": true, "Ev Idx": 4288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770622739.120, "dur": 0.418,
    "args": {
      "External id": 4290,"Record function id": 0, "finished": true, "Ev Idx": 4289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770622740.429, "dur": 44.058,
    "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": 870, "tid": 870,
    "ts": 7429770622740.957, "dur": 1.601,
    "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": 870, "tid": 870,
    "ts": 7429770622741.315, "dur": 1.131,
    "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": 870, "tid": 870,
    "ts": 7429770622742.871, "dur": 39.764,
    "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": 870, "tid": 870,
    "ts": 7429770622753.491, "dur": 1.059,
    "args": {
      "External id": 4295,"Record function id": 0, "finished": true, "Ev Idx": 4294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770622783.223, "dur": 0.984,
    "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": 870, "tid": 870,
    "ts": 7429770622787.118, "dur": 8.122,
    "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": 870, "tid": 870,
    "ts": 7429770622813.462, "dur": 0.250,
    "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": 870, "tid": 870,
    "ts": 7429770622815.439, "dur": 8.782,
    "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": 870, "tid": 870,
    "ts": 7429770622816.628, "dur": 0.277,
    "args": {
      "External id": 4300,"Record function id": 0, "finished": true, "Ev Idx": 4299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770622817.220, "dur": 0.277,
    "args": {
      "External id": 4301,"Record function id": 0, "finished": true, "Ev Idx": 4300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770622826.423, "dur": 12.883,
    "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": 870, "tid": 870,
    "ts": 7429770622843.176, "dur": 7.636,
    "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": 870, "tid": 870,
    "ts": 7429770622852.314, "dur": 7.986,
    "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": 870, "tid": 870,
    "ts": 7429770622861.843, "dur": 8.372,
    "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": 870, "tid": 870,
    "ts": 7429770622872.585, "dur": 0.441,
    "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": 870, "tid": 870,
    "ts": 7429770622873.731, "dur": 10.311,
    "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": 870, "tid": 870,
    "ts": 7429770622901.775, "dur": 38.279,
    "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": 870, "tid": 870,
    "ts": 7429770622902.130, "dur": 4.856,
    "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": 870, "tid": 870,
    "ts": 7429770622904.721, "dur": 1.361,
    "args": {
      "External id": 4310,"Record function id": 0, "finished": true, "Ev Idx": 4309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770622905.483, "dur": 0.473,
    "args": {
      "External id": 4311,"Record function id": 0, "finished": true, "Ev Idx": 4310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770622907.484, "dur": 32.455,
    "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": 870, "tid": 870,
    "ts": 7429770622908.135, "dur": 1.881,
    "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": 870, "tid": 870,
    "ts": 7429770622908.435, "dur": 1.474,
    "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": 870, "tid": 870,
    "ts": 7429770622910.333, "dur": 27.616,
    "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": 870, "tid": 870,
    "ts": 7429770622921.347, "dur": 1.316,
    "args": {
      "External id": 4316,"Record function id": 0, "finished": true, "Ev Idx": 4315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770622938.512, "dur": 1.151,
    "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": 870, "tid": 870,
    "ts": 7429770622943.099, "dur": 1.298,
    "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": 870, "tid": 870,
    "ts": 7429770622946.071, "dur": 1.955,
    "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": 870, "tid": 870,
    "ts": 7429770622947.026, "dur": 0.423,
    "args": {
      "External id": 4320,"Record function id": 0, "finished": true, "Ev Idx": 4319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770622954.730, "dur": 32.862,
    "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": 870, "tid": 870,
    "ts": 7429770622955.056, "dur": 2.092,
    "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": 870, "tid": 870,
    "ts": 7429770622955.921, "dur": 0.787,
    "args": {
      "External id": 4323,"Record function id": 0, "finished": true, "Ev Idx": 4322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770622956.323, "dur": 0.258,
    "args": {
      "External id": 4324,"Record function id": 0, "finished": true, "Ev Idx": 4323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770622957.466, "dur": 30.048,
    "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": 870, "tid": 870,
    "ts": 7429770622957.907, "dur": 1.333,
    "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": 870, "tid": 870,
    "ts": 7429770622958.317, "dur": 0.824,
    "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": 870, "tid": 870,
    "ts": 7429770622959.541, "dur": 26.095,
    "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": 870, "tid": 870,
    "ts": 7429770622968.526, "dur": 1.048,
    "args": {
      "External id": 4329,"Record function id": 0, "finished": true, "Ev Idx": 4328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770622986.377, "dur": 0.905,
    "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": 870, "tid": 870,
    "ts": 7429770622989.769, "dur": 1.092,
    "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": 870, "tid": 870,
    "ts": 7429770622992.218, "dur": 1.625,
    "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": 870, "tid": 870,
    "ts": 7429770622993.043, "dur": 0.318,
    "args": {
      "External id": 4333,"Record function id": 0, "finished": true, "Ev Idx": 4332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770623000.768, "dur": 32.786,
    "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": 870, "tid": 870,
    "ts": 7429770623001.088, "dur": 1.837,
    "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": 870, "tid": 870,
    "ts": 7429770623001.844, "dur": 0.701,
    "args": {
      "External id": 4336,"Record function id": 0, "finished": true, "Ev Idx": 4335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770623002.191, "dur": 0.240,
    "args": {
      "External id": 4337,"Record function id": 0, "finished": true, "Ev Idx": 4336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770623003.203, "dur": 30.254,
    "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": 870, "tid": 870,
    "ts": 7429770623003.627, "dur": 1.339,
    "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": 870, "tid": 870,
    "ts": 7429770623003.883, "dur": 0.977,
    "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": 870, "tid": 870,
    "ts": 7429770623005.267, "dur": 26.525,
    "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": 870, "tid": 870,
    "ts": 7429770623017.010, "dur": 1.007,
    "args": {
      "External id": 4342,"Record function id": 0, "finished": true, "Ev Idx": 4341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770623032.323, "dur": 0.906,
    "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": 870, "tid": 870,
    "ts": 7429770623035.837, "dur": 1.179,
    "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": 870, "tid": 870,
    "ts": 7429770623038.514, "dur": 1.783,
    "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": 870, "tid": 870,
    "ts": 7429770623039.366, "dur": 0.488,
    "args": {
      "External id": 4346,"Record function id": 0, "finished": true, "Ev Idx": 4345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770623043.132, "dur": 1.982,
    "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": 870, "tid": 870,
    "ts": 7429770623044.252, "dur": 0.326,
    "args": {
      "External id": 4348,"Record function id": 0, "finished": true, "Ev Idx": 4347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770623046.123, "dur": 1.087,
    "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": 870, "tid": 870,
    "ts": 7429770623046.584, "dur": 0.297,
    "args": {
      "External id": 4350,"Record function id": 0, "finished": true, "Ev Idx": 4349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770623048.280, "dur": 10.366,
    "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": 870, "tid": 870,
    "ts": 7429770623063.692, "dur": 3.451,
    "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": 870, "tid": 870,
    "ts": 7429770623065.532, "dur": 0.448,
    "args": {
      "External id": 4353,"Record function id": 0, "finished": true, "Ev Idx": 4352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770623069.395, "dur": 2.917,
    "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": 870, "tid": 870,
    "ts": 7429770623071.234, "dur": 0.483,
    "args": {
      "External id": 4355,"Record function id": 0, "finished": true, "Ev Idx": 4354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770623074.114, "dur": 11.424,
    "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": 870, "tid": 870,
    "ts": 7429770623089.659, "dur": 17.442,
    "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": 870, "tid": 870,
    "ts": 7429770623109.832, "dur": 9.201,
    "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": 870, "tid": 870,
    "ts": 7429770623120.621, "dur": 8.317,
    "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": 870, "tid": 870,
    "ts": 7429770623131.230, "dur": 6.853,
    "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": 870, "tid": 870,
    "ts": 7429770623140.333, "dur": 3.302,
    "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": 870, "tid": 870,
    "ts": 7429770623142.530, "dur": 0.435,
    "args": {
      "External id": 4362,"Record function id": 0, "finished": true, "Ev Idx": 4361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770623145.251, "dur": 2.649,
    "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": 870, "tid": 870,
    "ts": 7429770623147.033, "dur": 0.315,
    "args": {
      "External id": 4364,"Record function id": 0, "finished": true, "Ev Idx": 4363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770623148.897, "dur": 7.522,
    "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": 870, "tid": 870,
    "ts": 7429770623158.094, "dur": 9.319,
    "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": 870, "tid": 870,
    "ts": 7429770623169.631, "dur": 6.566,
    "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": 870, "tid": 870,
    "ts": 7429770623177.462, "dur": 6.996,
    "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": 870, "tid": 870,
    "ts": 7429770623207.622, "dur": 2.709,
    "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": 870, "tid": 870,
    "ts": 7429770623208.833, "dur": 0.657,
    "args": {
      "External id": 4370,"Record function id": 0, "finished": true, "Ev Idx": 4369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770623210.978, "dur": 2.130,
    "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": 870, "tid": 870,
    "ts": 7429770623212.243, "dur": 0.326,
    "args": {
      "External id": 4372,"Record function id": 0, "finished": true, "Ev Idx": 4371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770623213.720, "dur": 2.723,
    "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": 870, "tid": 870,
    "ts": 7429770623215.176, "dur": 0.748,
    "args": {
      "External id": 4374,"Record function id": 0, "finished": true, "Ev Idx": 4373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770623217.034, "dur": 1.650,
    "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": 870, "tid": 870,
    "ts": 7429770623217.771, "dur": 0.365,
    "args": {
      "External id": 4376,"Record function id": 0, "finished": true, "Ev Idx": 4375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770623219.274, "dur": 1.918,
    "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": 870, "tid": 870,
    "ts": 7429770623220.400, "dur": 0.322,
    "args": {
      "External id": 4378,"Record function id": 0, "finished": true, "Ev Idx": 4377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770623223.818, "dur": 3.075,
    "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": 870, "tid": 870,
    "ts": 7429770623225.871, "dur": 0.377,
    "args": {
      "External id": 4380,"Record function id": 0, "finished": true, "Ev Idx": 4379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770623229.008, "dur": 17.486,
    "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": 870, "tid": 870,
    "ts": 7429770623231.717, "dur": 12.255,
    "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": 870, "tid": 870,
    "ts": 7429770623232.943, "dur": 2.774,
    "args": {
      "External id": 4383,"Record function id": 0, "finished": true, "Ev Idx": 4382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770623233.839, "dur": 1.684,
    "args": {
      "External id": 4384,"Record function id": 0, "finished": true, "Ev Idx": 4383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770623236.331, "dur": 7.142,
    "args": {
      "External id": 4385,"Record function id": 0, "finished": true, "Ev Idx": 4384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770623244.655, "dur": 1.615,
    "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": 870, "tid": 870,
    "ts": 7429770623249.966, "dur": 2.067,
    "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": 870, "tid": 870,
    "ts": 7429770623251.113, "dur": 0.369,
    "args": {
      "External id": 4388,"Record function id": 0, "finished": true, "Ev Idx": 4387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770623252.649, "dur": 2.051,
    "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": 870, "tid": 870,
    "ts": 7429770623253.980, "dur": 0.253,
    "args": {
      "External id": 4390,"Record function id": 0, "finished": true, "Ev Idx": 4389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770623255.202, "dur": 2.367,
    "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": 870, "tid": 870,
    "ts": 7429770623256.848, "dur": 0.317,
    "args": {
      "External id": 4392,"Record function id": 0, "finished": true, "Ev Idx": 4391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770623258.139, "dur": 2.109,
    "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": 870, "tid": 870,
    "ts": 7429770623259.488, "dur": 0.278,
    "args": {
      "External id": 4394,"Record function id": 0, "finished": true, "Ev Idx": 4393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770623260.842, "dur": 1.461,
    "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": 870, "tid": 870,
    "ts": 7429770623261.579, "dur": 0.273,
    "args": {
      "External id": 4396,"Record function id": 0, "finished": true, "Ev Idx": 4395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770623263.745, "dur": 1.840,
    "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": 870, "tid": 870,
    "ts": 7429770623264.820, "dur": 0.328,
    "args": {
      "External id": 4398,"Record function id": 0, "finished": true, "Ev Idx": 4397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770623266.959, "dur": 11.917,
    "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": 870, "tid": 870,
    "ts": 7429770623267.870, "dur": 9.065,
    "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": 870, "tid": 870,
    "ts": 7429770623268.504, "dur": 1.651,
    "args": {
      "External id": 4401,"Record function id": 0, "finished": true, "Ev Idx": 4400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770623268.865, "dur": 1.184,
    "args": {
      "External id": 4402,"Record function id": 0, "finished": true, "Ev Idx": 4401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770623270.455, "dur": 6.148,
    "args": {
      "External id": 4403,"Record function id": 0, "finished": true, "Ev Idx": 4402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770623277.315, "dur": 1.404,
    "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": 870, "tid": 870,
    "ts": 7429770623280.858, "dur": 1.726,
    "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": 870, "tid": 870,
    "ts": 7429770623281.746, "dur": 0.364,
    "args": {
      "External id": 4406,"Record function id": 0, "finished": true, "Ev Idx": 4405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770623284.051, "dur": 111.513,
    "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": 870, "tid": 870,
    "ts": 7429770623284.603, "dur": 1.499,
    "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": 870, "tid": 870,
    "ts": 7429770623285.376, "dur": 0.307,
    "args": {
      "External id": 4409,"Record function id": 0, "finished": true, "Ev Idx": 4408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770623286.444, "dur": 3.658,
    "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": 870, "tid": 870,
    "ts": 7429770623288.154, "dur": 1.809,
    "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": 870, "tid": 870,
    "ts": 7429770623290.675, "dur": 1.384,
    "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": 870, "tid": 870,
    "ts": 7429770623291.398, "dur": 0.291,
    "args": {
      "External id": 4413,"Record function id": 0, "finished": true, "Ev Idx": 4412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770623292.511, "dur": 2.385,
    "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": 870, "tid": 870,
    "ts": 7429770623293.301, "dur": 1.504,
    "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": 870, "tid": 870,
    "ts": 7429770623295.456, "dur": 97.723,
    "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": 870, "tid": 870,
    "ts": 7429770623370.357, "dur": 1.316,
    "args": {
      "External id": 4417,"Record function id": 0, "finished": true, "Ev Idx": 4416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770623393.864, "dur": 1.290,
    "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": 870, "tid": 870,
    "ts": 7429770623400.139, "dur": 11.924,
    "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": 870, "tid": 870,
    "ts": 7429770623415.642, "dur": 2.728,
    "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": 870, "tid": 870,
    "ts": 7429770623416.841, "dur": 0.628,
    "args": {
      "External id": 4421,"Record function id": 0, "finished": true, "Ev Idx": 4420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770623419.049, "dur": 1.729,
    "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": 870, "tid": 870,
    "ts": 7429770623419.780, "dur": 0.563,
    "args": {
      "External id": 4423,"Record function id": 0, "finished": true, "Ev Idx": 4422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770623421.406, "dur": 1.185,
    "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": 870, "tid": 870,
    "ts": 7429770623422.023, "dur": 0.241,
    "args": {
      "External id": 4425,"Record function id": 0, "finished": true, "Ev Idx": 4424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770623424.541, "dur": 12.224,
    "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": 870, "tid": 870,
    "ts": 7429770623442.972, "dur": 13.973,
    "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": 870, "tid": 870,
    "ts": 7429770623444.115, "dur": 0.417,
    "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": 870, "tid": 870,
    "ts": 7429770623445.252, "dur": 11.459,
    "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": 870, "tid": 870,
    "ts": 7429770623459.764, "dur": 0.376,
    "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": 870, "tid": 870,
    "ts": 7429770623470.484, "dur": 0.667,
    "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": 870, "tid": 870,
    "ts": 7429770623472.979, "dur": 45.516,
    "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": 870, "tid": 870,
    "ts": 7429770623473.778, "dur": 2.725,
    "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": 870, "tid": 870,
    "ts": 7429770623475.482, "dur": 0.405,
    "args": {
      "External id": 4434,"Record function id": 0, "finished": true, "Ev Idx": 4433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770623477.040, "dur": 3.064,
    "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": 870, "tid": 870,
    "ts": 7429770623477.548, "dur": 2.442,
    "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": 870, "tid": 870,
    "ts": 7429770623480.575, "dur": 1.913,
    "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": 870, "tid": 870,
    "ts": 7429770623481.805, "dur": 0.306,
    "args": {
      "External id": 4438,"Record function id": 0, "finished": true, "Ev Idx": 4437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770623482.818, "dur": 1.618,
    "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": 870, "tid": 870,
    "ts": 7429770623483.263, "dur": 1.099,
    "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": 870, "tid": 870,
    "ts": 7429770623485.058, "dur": 31.424,
    "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": 870, "tid": 870,
    "ts": 7429770623497.095, "dur": 1.289,
    "args": {
      "External id": 4442,"Record function id": 0, "finished": true, "Ev Idx": 4441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770623517.030, "dur": 1.111,
    "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": 870, "tid": 870,
    "ts": 7429770623520.671, "dur": 2.065,
    "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": 870, "tid": 870,
    "ts": 7429770623521.800, "dur": 0.396,
    "args": {
      "External id": 4445,"Record function id": 0, "finished": true, "Ev Idx": 4444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770623525.120, "dur": 13.090,
    "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": 870, "tid": 870,
    "ts": 7429770623525.717, "dur": 12.356,
    "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": 870, "tid": 870,
    "ts": 7429770623527.007, "dur": 2.487,
    "args": {
      "External id": 4448,"Record function id": 0, "finished": true, "Ev Idx": 4447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770623527.700, "dur": 1.643,
    "args": {
      "External id": 4449,"Record function id": 0, "finished": true, "Ev Idx": 4448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770623529.930, "dur": 7.736,
    "args": {
      "External id": 4450,"Record function id": 0, "finished": true, "Ev Idx": 4449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770623544.469, "dur": 1.779,
    "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": 870, "tid": 870,
    "ts": 7429770623544.953, "dur": 1.202,
    "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": 870, "tid": 870,
    "ts": 7429770623562.579, "dur": 41.258,
    "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": 870, "tid": 870,
    "ts": 7429770623563.680, "dur": 3.203,
    "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": 870, "tid": 870,
    "ts": 7429770623565.172, "dur": 1.116,
    "args": {
      "External id": 4455,"Record function id": 0, "finished": true, "Ev Idx": 4454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770623565.734, "dur": 0.414,
    "args": {
      "External id": 4456,"Record function id": 0, "finished": true, "Ev Idx": 4455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770623567.369, "dur": 36.287,
    "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": 870, "tid": 870,
    "ts": 7429770623567.978, "dur": 2.137,
    "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": 870, "tid": 870,
    "ts": 7429770623568.432, "dur": 1.574,
    "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": 870, "tid": 870,
    "ts": 7429770623570.585, "dur": 30.921,
    "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": 870, "tid": 870,
    "ts": 7429770623585.118, "dur": 1.324,
    "args": {
      "External id": 4461,"Record function id": 0, "finished": true, "Ev Idx": 4460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770623602.193, "dur": 1.123,
    "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": 870, "tid": 870,
    "ts": 7429770623609.583, "dur": 8.593,
    "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": 870, "tid": 870,
    "ts": 7429770623627.594, "dur": 0.223,
    "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": 870, "tid": 870,
    "ts": 7429770623630.084, "dur": 10.834,
    "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": 870, "tid": 870,
    "ts": 7429770623631.903, "dur": 0.446,
    "args": {
      "External id": 4466,"Record function id": 0, "finished": true, "Ev Idx": 4465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770623632.869, "dur": 0.074,
    "args": {
      "External id": 4467,"Record function id": 0, "finished": true, "Ev Idx": 4466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770623643.747, "dur": 14.098,
    "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": 870, "tid": 870,
    "ts": 7429770623661.395, "dur": 8.791,
    "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": 870, "tid": 870,
    "ts": 7429770623672.700, "dur": 11.338,
    "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": 870, "tid": 870,
    "ts": 7429770623685.694, "dur": 8.958,
    "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": 870, "tid": 870,
    "ts": 7429770623697.254, "dur": 0.104,
    "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": 870, "tid": 870,
    "ts": 7429770623698.373, "dur": 9.086,
    "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": 870, "tid": 870,
    "ts": 7429770623722.100, "dur": 39.047,
    "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": 870, "tid": 870,
    "ts": 7429770623722.800, "dur": 3.179,
    "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": 870, "tid": 870,
    "ts": 7429770623724.034, "dur": 1.272,
    "args": {
      "External id": 4476,"Record function id": 0, "finished": true, "Ev Idx": 4475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770623724.655, "dur": 0.502,
    "args": {
      "External id": 4477,"Record function id": 0, "finished": true, "Ev Idx": 4476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770623726.713, "dur": 34.290,
    "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": 870, "tid": 870,
    "ts": 7429770623727.278, "dur": 2.062,
    "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": 870, "tid": 870,
    "ts": 7429770623727.640, "dur": 1.559,
    "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": 870, "tid": 870,
    "ts": 7429770623729.720, "dur": 29.324,
    "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": 870, "tid": 870,
    "ts": 7429770623740.800, "dur": 1.186,
    "args": {
      "External id": 4482,"Record function id": 0, "finished": true, "Ev Idx": 4481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770623759.678, "dur": 1.033,
    "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": 870, "tid": 870,
    "ts": 7429770623769.986, "dur": 11.293,
    "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": 870, "tid": 870,
    "ts": 7429770623789.366, "dur": 90.679,
    "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": 870, "tid": 870,
    "ts": 7429770623789.677, "dur": 2.672,
    "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": 870, "tid": 870,
    "ts": 7429770623790.714, "dur": 1.096,
    "args": {
      "External id": 4487,"Record function id": 0, "finished": true, "Ev Idx": 4486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770623791.112, "dur": 0.568,
    "args": {
      "External id": 4488,"Record function id": 0, "finished": true, "Ev Idx": 4487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770623792.686, "dur": 87.247,
    "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": 870, "tid": 870,
    "ts": 7429770623793.078, "dur": 3.584,
    "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": 870, "tid": 870,
    "ts": 7429770623793.381, "dur": 3.167,
    "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": 870, "tid": 870,
    "ts": 7429770623797.184, "dur": 80.195,
    "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": 870, "tid": 870,
    "ts": 7429770623807.797, "dur": 51.779,
    "args": {
      "External id": 4493,"Record function id": 0, "finished": true, "Ev Idx": 4492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770623878.273, "dur": 1.354,
    "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": 870, "tid": 870,
    "ts": 7429770623882.498, "dur": 9.331,
    "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": 870, "tid": 870,
    "ts": 7429770623899.294, "dur": 51.459,
    "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": 870, "tid": 870,
    "ts": 7429770623899.921, "dur": 3.190,
    "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": 870, "tid": 870,
    "ts": 7429770623901.044, "dur": 1.042,
    "args": {
      "External id": 4498,"Record function id": 0, "finished": true, "Ev Idx": 4497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770623901.514, "dur": 0.424,
    "args": {
      "External id": 4499,"Record function id": 0, "finished": true, "Ev Idx": 4498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770623903.445, "dur": 47.166,
    "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": 870, "tid": 870,
    "ts": 7429770623903.905, "dur": 2.207,
    "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": 870, "tid": 870,
    "ts": 7429770623904.191, "dur": 1.827,
    "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": 870, "tid": 870,
    "ts": 7429770623906.482, "dur": 42.343,
    "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": 870, "tid": 870,
    "ts": 7429770623916.597, "dur": 1.233,
    "args": {
      "External id": 4504,"Record function id": 0, "finished": true, "Ev Idx": 4503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770623949.415, "dur": 0.913,
    "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": 870, "tid": 870,
    "ts": 7429770623953.571, "dur": 9.717,
    "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": 870, "tid": 870,
    "ts": 7429770623983.561, "dur": 0.179,
    "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": 870, "tid": 870,
    "ts": 7429770623985.794, "dur": 10.376,
    "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": 870, "tid": 870,
    "ts": 7429770623987.176, "dur": 0.428,
    "args": {
      "External id": 4509,"Record function id": 0, "finished": true, "Ev Idx": 4508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770623987.924, "dur": 0.082,
    "args": {
      "External id": 4510,"Record function id": 0, "finished": true, "Ev Idx": 4509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770623998.915, "dur": 12.444,
    "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": 870, "tid": 870,
    "ts": 7429770624015.246, "dur": 7.806,
    "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": 870, "tid": 870,
    "ts": 7429770624024.702, "dur": 9.755,
    "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": 870, "tid": 870,
    "ts": 7429770624036.113, "dur": 10.456,
    "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": 870, "tid": 870,
    "ts": 7429770624049.074, "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": 870, "tid": 870,
    "ts": 7429770624050.155, "dur": 9.750,
    "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": 870, "tid": 870,
    "ts": 7429770624078.308, "dur": 43.996,
    "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": 870, "tid": 870,
    "ts": 7429770624079.109, "dur": 4.209,
    "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": 870, "tid": 870,
    "ts": 7429770624081.319, "dur": 1.347,
    "args": {
      "External id": 4519,"Record function id": 0, "finished": true, "Ev Idx": 4518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770624082.026, "dur": 0.469,
    "args": {
      "External id": 4520,"Record function id": 0, "finished": true, "Ev Idx": 4519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770624083.811, "dur": 38.367,
    "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": 870, "tid": 870,
    "ts": 7429770624084.590, "dur": 1.683,
    "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": 870, "tid": 870,
    "ts": 7429770624084.977, "dur": 1.186,
    "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": 870, "tid": 870,
    "ts": 7429770624086.609, "dur": 33.302,
    "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": 870, "tid": 870,
    "ts": 7429770624101.710, "dur": 1.246,
    "args": {
      "External id": 4525,"Record function id": 0, "finished": true, "Ev Idx": 4524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770624120.603, "dur": 1.249,
    "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": 870, "tid": 870,
    "ts": 7429770624125.892, "dur": 1.382,
    "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": 870, "tid": 870,
    "ts": 7429770624129.156, "dur": 2.466,
    "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": 870, "tid": 870,
    "ts": 7429770624130.409, "dur": 0.523,
    "args": {
      "External id": 4529,"Record function id": 0, "finished": true, "Ev Idx": 4528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770624138.859, "dur": 36.994,
    "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": 870, "tid": 870,
    "ts": 7429770624139.236, "dur": 2.037,
    "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": 870, "tid": 870,
    "ts": 7429770624140.056, "dur": 0.781,
    "args": {
      "External id": 4532,"Record function id": 0, "finished": true, "Ev Idx": 4531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770624140.445, "dur": 0.287,
    "args": {
      "External id": 4533,"Record function id": 0, "finished": true, "Ev Idx": 4532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770624142.030, "dur": 33.721,
    "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": 870, "tid": 870,
    "ts": 7429770624142.769, "dur": 1.748,
    "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": 870, "tid": 870,
    "ts": 7429770624143.348, "dur": 1.056,
    "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": 870, "tid": 870,
    "ts": 7429770624144.945, "dur": 29.122,
    "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": 870, "tid": 870,
    "ts": 7429770624156.709, "dur": 1.273,
    "args": {
      "External id": 4538,"Record function id": 0, "finished": true, "Ev Idx": 4537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770624174.587, "dur": 0.932,
    "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": 870, "tid": 870,
    "ts": 7429770624178.186, "dur": 1.138,
    "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": 870, "tid": 870,
    "ts": 7429770624180.863, "dur": 1.587,
    "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": 870, "tid": 870,
    "ts": 7429770624181.642, "dur": 0.349,
    "args": {
      "External id": 4542,"Record function id": 0, "finished": true, "Ev Idx": 4541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770624189.212, "dur": 35.409,
    "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": 870, "tid": 870,
    "ts": 7429770624189.516, "dur": 2.426,
    "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": 870, "tid": 870,
    "ts": 7429770624190.497, "dur": 1.066,
    "args": {
      "External id": 4545,"Record function id": 0, "finished": true, "Ev Idx": 4544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770624191.208, "dur": 0.250,
    "args": {
      "External id": 4546,"Record function id": 0, "finished": true, "Ev Idx": 4545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770624192.241, "dur": 32.283,
    "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": 870, "tid": 870,
    "ts": 7429770624192.910, "dur": 1.930,
    "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": 870, "tid": 870,
    "ts": 7429770624193.327, "dur": 1.391,
    "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": 870, "tid": 870,
    "ts": 7429770624195.253, "dur": 27.446,
    "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": 870, "tid": 870,
    "ts": 7429770624207.367, "dur": 1.285,
    "args": {
      "External id": 4551,"Record function id": 0, "finished": true, "Ev Idx": 4550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770624223.222, "dur": 1.080,
    "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": 870, "tid": 870,
    "ts": 7429770624226.889, "dur": 1.131,
    "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": 870, "tid": 870,
    "ts": 7429770624229.540, "dur": 1.704,
    "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": 870, "tid": 870,
    "ts": 7429770624230.400, "dur": 0.364,
    "args": {
      "External id": 4555,"Record function id": 0, "finished": true, "Ev Idx": 4554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770624234.271, "dur": 2.442,
    "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": 870, "tid": 870,
    "ts": 7429770624235.674, "dur": 0.504,
    "args": {
      "External id": 4557,"Record function id": 0, "finished": true, "Ev Idx": 4556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770624237.821, "dur": 1.006,
    "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": 870, "tid": 870,
    "ts": 7429770624238.231, "dur": 0.269,
    "args": {
      "External id": 4559,"Record function id": 0, "finished": true, "Ev Idx": 4558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770624240.072, "dur": 13.337,
    "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": 870, "tid": 870,
    "ts": 7429770624259.001, "dur": 5.042,
    "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": 870, "tid": 870,
    "ts": 7429770624262.001, "dur": 0.591,
    "args": {
      "External id": 4562,"Record function id": 0, "finished": true, "Ev Idx": 4561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770624266.403, "dur": 1.818,
    "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": 870, "tid": 870,
    "ts": 7429770624267.223, "dur": 0.355,
    "args": {
      "External id": 4564,"Record function id": 0, "finished": true, "Ev Idx": 4563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770624269.946, "dur": 11.145,
    "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": 870, "tid": 870,
    "ts": 7429770624285.447, "dur": 18.413,
    "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": 870, "tid": 870,
    "ts": 7429770624306.758, "dur": 7.762,
    "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": 870, "tid": 870,
    "ts": 7429770624316.029, "dur": 8.471,
    "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": 870, "tid": 870,
    "ts": 7429770624326.740, "dur": 7.102,
    "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": 870, "tid": 870,
    "ts": 7429770624336.396, "dur": 2.478,
    "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": 870, "tid": 870,
    "ts": 7429770624337.739, "dur": 0.412,
    "args": {
      "External id": 4571,"Record function id": 0, "finished": true, "Ev Idx": 4570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770624340.559, "dur": 3.110,
    "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": 870, "tid": 870,
    "ts": 7429770624342.817, "dur": 0.296,
    "args": {
      "External id": 4573,"Record function id": 0, "finished": true, "Ev Idx": 4572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770624344.801, "dur": 7.919,
    "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": 870, "tid": 870,
    "ts": 7429770624354.547, "dur": 10.280,
    "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": 870, "tid": 870,
    "ts": 7429770624367.359, "dur": 7.721,
    "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": 870, "tid": 870,
    "ts": 7429770624376.365, "dur": 6.230,
    "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": 870, "tid": 870,
    "ts": 7429770624406.795, "dur": 2.538,
    "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": 870, "tid": 870,
    "ts": 7429770624408.085, "dur": 0.447,
    "args": {
      "External id": 4579,"Record function id": 0, "finished": true, "Ev Idx": 4578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770624410.019, "dur": 2.015,
    "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": 870, "tid": 870,
    "ts": 7429770624410.973, "dur": 0.485,
    "args": {
      "External id": 4581,"Record function id": 0, "finished": true, "Ev Idx": 4580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770624413.159, "dur": 2.955,
    "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": 870, "tid": 870,
    "ts": 7429770624414.398, "dur": 0.972,
    "args": {
      "External id": 4583,"Record function id": 0, "finished": true, "Ev Idx": 4582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770624416.928, "dur": 2.771,
    "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": 870, "tid": 870,
    "ts": 7429770624418.778, "dur": 0.396,
    "args": {
      "External id": 4585,"Record function id": 0, "finished": true, "Ev Idx": 4584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770624420.348, "dur": 2.030,
    "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": 870, "tid": 870,
    "ts": 7429770624421.584, "dur": 0.316,
    "args": {
      "External id": 4587,"Record function id": 0, "finished": true, "Ev Idx": 4586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770624425.372, "dur": 3.947,
    "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": 870, "tid": 870,
    "ts": 7429770624428.234, "dur": 0.403,
    "args": {
      "External id": 4589,"Record function id": 0, "finished": true, "Ev Idx": 4588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770624431.650, "dur": 19.993,
    "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": 870, "tid": 870,
    "ts": 7429770624434.289, "dur": 15.397,
    "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": 870, "tid": 870,
    "ts": 7429770624436.516, "dur": 2.718,
    "args": {
      "External id": 4592,"Record function id": 0, "finished": true, "Ev Idx": 4591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770624437.318, "dur": 1.698,
    "args": {
      "External id": 4593,"Record function id": 0, "finished": true, "Ev Idx": 4592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770624440.065, "dur": 9.038,
    "args": {
      "External id": 4594,"Record function id": 0, "finished": true, "Ev Idx": 4593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770624450.154, "dur": 1.234,
    "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": 870, "tid": 870,
    "ts": 7429770624454.878, "dur": 2.341,
    "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": 870, "tid": 870,
    "ts": 7429770624456.159, "dur": 0.495,
    "args": {
      "External id": 4597,"Record function id": 0, "finished": true, "Ev Idx": 4596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770624457.835, "dur": 2.276,
    "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": 870, "tid": 870,
    "ts": 7429770624459.375, "dur": 0.275,
    "args": {
      "External id": 4599,"Record function id": 0, "finished": true, "Ev Idx": 4598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770624460.653, "dur": 2.665,
    "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": 870, "tid": 870,
    "ts": 7429770624462.572, "dur": 0.320,
    "args": {
      "External id": 4601,"Record function id": 0, "finished": true, "Ev Idx": 4600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770624464.112, "dur": 2.090,
    "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": 870, "tid": 870,
    "ts": 7429770624465.391, "dur": 0.328,
    "args": {
      "External id": 4603,"Record function id": 0, "finished": true, "Ev Idx": 4602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770624466.913, "dur": 1.412,
    "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": 870, "tid": 870,
    "ts": 7429770624467.579, "dur": 0.279,
    "args": {
      "External id": 4605,"Record function id": 0, "finished": true, "Ev Idx": 4604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770624469.834, "dur": 1.753,
    "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": 870, "tid": 870,
    "ts": 7429770624470.661, "dur": 0.489,
    "args": {
      "External id": 4607,"Record function id": 0, "finished": true, "Ev Idx": 4606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770624473.044, "dur": 46.137,
    "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": 870, "tid": 870,
    "ts": 7429770624507.215, "dur": 9.749,
    "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": 870, "tid": 870,
    "ts": 7429770624507.969, "dur": 1.783,
    "args": {
      "External id": 4610,"Record function id": 0, "finished": true, "Ev Idx": 4609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770624508.369, "dur": 1.264,
    "args": {
      "External id": 4611,"Record function id": 0, "finished": true, "Ev Idx": 4610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770624510.059, "dur": 6.597,
    "args": {
      "External id": 4612,"Record function id": 0, "finished": true, "Ev Idx": 4611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770624517.341, "dur": 1.682,
    "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": 870, "tid": 870,
    "ts": 7429770624521.175, "dur": 1.860,
    "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": 870, "tid": 870,
    "ts": 7429770624522.102, "dur": 0.378,
    "args": {
      "External id": 4615,"Record function id": 0, "finished": true, "Ev Idx": 4614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770624524.739, "dur": 102.649,
    "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": 870, "tid": 870,
    "ts": 7429770624525.359, "dur": 1.563,
    "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": 870, "tid": 870,
    "ts": 7429770624526.157, "dur": 0.348,
    "args": {
      "External id": 4618,"Record function id": 0, "finished": true, "Ev Idx": 4617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770624527.242, "dur": 3.870,
    "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": 870, "tid": 870,
    "ts": 7429770624529.072, "dur": 1.909,
    "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": 870, "tid": 870,
    "ts": 7429770624531.551, "dur": 1.382,
    "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": 870, "tid": 870,
    "ts": 7429770624532.242, "dur": 0.309,
    "args": {
      "External id": 4622,"Record function id": 0, "finished": true, "Ev Idx": 4621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770624533.248, "dur": 1.945,
    "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": 870, "tid": 870,
    "ts": 7429770624534.070, "dur": 1.036,
    "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": 870, "tid": 870,
    "ts": 7429770624535.819, "dur": 89.218,
    "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": 870, "tid": 870,
    "ts": 7429770624603.446, "dur": 1.378,
    "args": {
      "External id": 4626,"Record function id": 0, "finished": true, "Ev Idx": 4625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770624625.740, "dur": 1.300,
    "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": 870, "tid": 870,
    "ts": 7429770624631.606, "dur": 11.924,
    "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": 870, "tid": 870,
    "ts": 7429770624647.085, "dur": 2.328,
    "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": 870, "tid": 870,
    "ts": 7429770624648.044, "dur": 0.502,
    "args": {
      "External id": 4630,"Record function id": 0, "finished": true, "Ev Idx": 4629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770624650.092, "dur": 1.284,
    "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": 870, "tid": 870,
    "ts": 7429770624650.656, "dur": 0.304,
    "args": {
      "External id": 4632,"Record function id": 0, "finished": true, "Ev Idx": 4631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770624651.979, "dur": 1.026,
    "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": 870, "tid": 870,
    "ts": 7429770624652.436, "dur": 0.246,
    "args": {
      "External id": 4634,"Record function id": 0, "finished": true, "Ev Idx": 4633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770624654.945, "dur": 10.361,
    "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": 870, "tid": 870,
    "ts": 7429770624670.804, "dur": 16.627,
    "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": 870, "tid": 870,
    "ts": 7429770624671.967, "dur": 0.582,
    "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": 870, "tid": 870,
    "ts": 7429770624673.167, "dur": 13.987,
    "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": 870, "tid": 870,
    "ts": 7429770624690.384, "dur": 0.179,
    "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": 870, "tid": 870,
    "ts": 7429770624702.171, "dur": 0.509,
    "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": 870, "tid": 870,
    "ts": 7429770624704.393, "dur": 42.532,
    "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": 870, "tid": 870,
    "ts": 7429770624704.959, "dur": 3.004,
    "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": 870, "tid": 870,
    "ts": 7429770624707.108, "dur": 0.362,
    "args": {
      "External id": 4643,"Record function id": 0, "finished": true, "Ev Idx": 4642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770624708.336, "dur": 2.158,
    "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": 870, "tid": 870,
    "ts": 7429770624708.707, "dur": 1.667,
    "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": 870, "tid": 870,
    "ts": 7429770624710.908, "dur": 1.278,
    "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": 870, "tid": 870,
    "ts": 7429770624711.595, "dur": 0.233,
    "args": {
      "External id": 4647,"Record function id": 0, "finished": true, "Ev Idx": 4646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770624712.664, "dur": 1.351,
    "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": 870, "tid": 870,
    "ts": 7429770624713.045, "dur": 0.886,
    "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": 870, "tid": 870,
    "ts": 7429770624714.407, "dur": 30.607,
    "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": 870, "tid": 870,
    "ts": 7429770624727.887, "dur": 1.144,
    "args": {
      "External id": 4651,"Record function id": 0, "finished": true, "Ev Idx": 4650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770624745.578, "dur": 1.054,
    "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": 870, "tid": 870,
    "ts": 7429770624748.956, "dur": 2.005,
    "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": 870, "tid": 870,
    "ts": 7429770624749.995, "dur": 0.371,
    "args": {
      "External id": 4654,"Record function id": 0, "finished": true, "Ev Idx": 4653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770624753.339, "dur": 12.469,
    "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": 870, "tid": 870,
    "ts": 7429770624753.936, "dur": 11.733,
    "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": 870, "tid": 870,
    "ts": 7429770624754.894, "dur": 2.121,
    "args": {
      "External id": 4657,"Record function id": 0, "finished": true, "Ev Idx": 4656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770624755.346, "dur": 1.527,
    "args": {
      "External id": 4658,"Record function id": 0, "finished": true, "Ev Idx": 4657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770624757.400, "dur": 7.762,
    "args": {
      "External id": 4659,"Record function id": 0, "finished": true, "Ev Idx": 4658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770624771.700, "dur": 1.583,
    "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": 870, "tid": 870,
    "ts": 7429770624771.991, "dur": 1.184,
    "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": 870, "tid": 870,
    "ts": 7429770624789.117, "dur": 36.782,
    "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": 870, "tid": 870,
    "ts": 7429770624790.336, "dur": 4.128,
    "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": 870, "tid": 870,
    "ts": 7429770624792.811, "dur": 0.931,
    "args": {
      "External id": 4664,"Record function id": 0, "finished": true, "Ev Idx": 4663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770624793.257, "dur": 0.341,
    "args": {
      "External id": 4665,"Record function id": 0, "finished": true, "Ev Idx": 4664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770624794.821, "dur": 30.911,
    "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": 870, "tid": 870,
    "ts": 7429770624795.288, "dur": 1.985,
    "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": 870, "tid": 870,
    "ts": 7429770624795.575, "dur": 1.579,
    "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": 870, "tid": 870,
    "ts": 7429770624797.705, "dur": 25.778,
    "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": 870, "tid": 870,
    "ts": 7429770624807.328, "dur": 1.328,
    "args": {
      "External id": 4670,"Record function id": 0, "finished": true, "Ev Idx": 4669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770624824.402, "dur": 1.023,
    "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": 870, "tid": 870,
    "ts": 7429770624831.802, "dur": 9.605,
    "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": 870, "tid": 870,
    "ts": 7429770624851.536, "dur": 0.184,
    "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": 870, "tid": 870,
    "ts": 7429770624854.056, "dur": 12.230,
    "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": 870, "tid": 870,
    "ts": 7429770624855.768, "dur": 0.444,
    "args": {
      "External id": 4675,"Record function id": 0, "finished": true, "Ev Idx": 4674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770624856.561, "dur": 0.088,
    "args": {
      "External id": 4676,"Record function id": 0, "finished": true, "Ev Idx": 4675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770624868.876, "dur": 12.855,
    "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": 870, "tid": 870,
    "ts": 7429770624885.205, "dur": 9.398,
    "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": 870, "tid": 870,
    "ts": 7429770624896.421, "dur": 10.672,
    "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": 870, "tid": 870,
    "ts": 7429770624908.847, "dur": 8.816,
    "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": 870, "tid": 870,
    "ts": 7429770624919.963, "dur": 0.119,
    "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": 870, "tid": 870,
    "ts": 7429770624920.858, "dur": 8.549,
    "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": 870, "tid": 870,
    "ts": 7429770624946.240, "dur": 42.101,
    "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": 870, "tid": 870,
    "ts": 7429770624947.140, "dur": 3.951,
    "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": 870, "tid": 870,
    "ts": 7429770624949.220, "dur": 1.267,
    "args": {
      "External id": 4685,"Record function id": 0, "finished": true, "Ev Idx": 4684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770624949.877, "dur": 0.464,
    "args": {
      "External id": 4686,"Record function id": 0, "finished": true, "Ev Idx": 4685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770624951.445, "dur": 36.739,
    "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": 870, "tid": 870,
    "ts": 7429770624952.161, "dur": 2.553,
    "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": 870, "tid": 870,
    "ts": 7429770624952.645, "dur": 1.953,
    "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": 870, "tid": 870,
    "ts": 7429770624955.095, "dur": 31.197,
    "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": 870, "tid": 870,
    "ts": 7429770624969.127, "dur": 1.131,
    "args": {
      "External id": 4691,"Record function id": 0, "finished": true, "Ev Idx": 4690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770624986.973, "dur": 0.954,
    "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": 870, "tid": 870,
    "ts": 7429770624996.377, "dur": 11.027,
    "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": 870, "tid": 870,
    "ts": 7429770625014.525, "dur": 34.160,
    "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": 870, "tid": 870,
    "ts": 7429770625014.860, "dur": 2.352,
    "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": 870, "tid": 870,
    "ts": 7429770625015.654, "dur": 1.058,
    "args": {
      "External id": 4696,"Record function id": 0, "finished": true, "Ev Idx": 4695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770625016.164, "dur": 0.389,
    "args": {
      "External id": 4697,"Record function id": 0, "finished": true, "Ev Idx": 4696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770625017.571, "dur": 30.984,
    "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": 870, "tid": 870,
    "ts": 7429770625018.006, "dur": 2.431,
    "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": 870, "tid": 870,
    "ts": 7429770625018.297, "dur": 2.041,
    "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": 870, "tid": 870,
    "ts": 7429770625020.738, "dur": 25.916,
    "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": 870, "tid": 870,
    "ts": 7429770625029.410, "dur": 1.082,
    "args": {
      "External id": 4702,"Record function id": 0, "finished": true, "Ev Idx": 4701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770625047.452, "dur": 0.857,
    "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": 870, "tid": 870,
    "ts": 7429770625050.180, "dur": 7.935,
    "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": 870, "tid": 870,
    "ts": 7429770625063.887, "dur": 45.685,
    "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": 870, "tid": 870,
    "ts": 7429770625064.254, "dur": 2.305,
    "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": 870, "tid": 870,
    "ts": 7429770625065.151, "dur": 0.892,
    "args": {
      "External id": 4707,"Record function id": 0, "finished": true, "Ev Idx": 4706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770625065.555, "dur": 0.359,
    "args": {
      "External id": 4708,"Record function id": 0, "finished": true, "Ev Idx": 4707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770625066.854, "dur": 42.600,
    "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": 870, "tid": 870,
    "ts": 7429770625067.241, "dur": 2.095,
    "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": 870, "tid": 870,
    "ts": 7429770625067.701, "dur": 1.544,
    "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": 870, "tid": 870,
    "ts": 7429770625069.638, "dur": 38.106,
    "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": 870, "tid": 870,
    "ts": 7429770625078.296, "dur": 1.190,
    "args": {
      "External id": 4713,"Record function id": 0, "finished": true, "Ev Idx": 4712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770625108.305, "dur": 0.901,
    "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": 870, "tid": 870,
    "ts": 7429770625112.131, "dur": 8.562,
    "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": 870, "tid": 870,
    "ts": 7429770625138.423, "dur": 0.285,
    "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": 870, "tid": 870,
    "ts": 7429770625140.235, "dur": 10.532,
    "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": 870, "tid": 870,
    "ts": 7429770625141.378, "dur": 0.427,
    "args": {
      "External id": 4718,"Record function id": 0, "finished": true, "Ev Idx": 4717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770625142.107, "dur": 0.069,
    "args": {
      "External id": 4719,"Record function id": 0, "finished": true, "Ev Idx": 4718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770625152.929, "dur": 10.335,
    "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": 870, "tid": 870,
    "ts": 7429770625166.300, "dur": 7.546,
    "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": 870, "tid": 870,
    "ts": 7429770625175.268, "dur": 7.591,
    "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": 870, "tid": 870,
    "ts": 7429770625184.325, "dur": 8.786,
    "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": 870, "tid": 870,
    "ts": 7429770625195.249, "dur": 0.103,
    "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": 870, "tid": 870,
    "ts": 7429770625196.265, "dur": 10.171,
    "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": 870, "tid": 870,
    "ts": 7429770625224.223, "dur": 40.514,
    "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": 870, "tid": 870,
    "ts": 7429770625224.710, "dur": 3.881,
    "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": 870, "tid": 870,
    "ts": 7429770625226.602, "dur": 1.151,
    "args": {
      "External id": 4728,"Record function id": 0, "finished": true, "Ev Idx": 4727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770625227.084, "dur": 0.514,
    "args": {
      "External id": 4729,"Record function id": 0, "finished": true, "Ev Idx": 4728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770625229.066, "dur": 35.550,
    "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": 870, "tid": 870,
    "ts": 7429770625229.602, "dur": 1.477,
    "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": 870, "tid": 870,
    "ts": 7429770625229.927, "dur": 1.038,
    "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": 870, "tid": 870,
    "ts": 7429770625231.409, "dur": 31.391,
    "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": 870, "tid": 870,
    "ts": 7429770625245.387, "dur": 1.340,
    "args": {
      "External id": 4734,"Record function id": 0, "finished": true, "Ev Idx": 4733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770625263.438, "dur": 0.908,
    "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": 870, "tid": 870,
    "ts": 7429770625267.925, "dur": 1.306,
    "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": 870, "tid": 870,
    "ts": 7429770625271.063, "dur": 1.872,
    "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": 870, "tid": 870,
    "ts": 7429770625271.899, "dur": 0.444,
    "args": {
      "External id": 4738,"Record function id": 0, "finished": true, "Ev Idx": 4737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770625279.939, "dur": 33.110,
    "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": 870, "tid": 870,
    "ts": 7429770625280.276, "dur": 2.387,
    "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": 870, "tid": 870,
    "ts": 7429770625281.177, "dur": 0.978,
    "args": {
      "External id": 4741,"Record function id": 0, "finished": true, "Ev Idx": 4740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770625281.742, "dur": 0.298,
    "args": {
      "External id": 4742,"Record function id": 0, "finished": true, "Ev Idx": 4741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770625282.997, "dur": 29.956,
    "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": 870, "tid": 870,
    "ts": 7429770625283.414, "dur": 2.096,
    "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": 870, "tid": 870,
    "ts": 7429770625283.721, "dur": 1.696,
    "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": 870, "tid": 870,
    "ts": 7429770625285.806, "dur": 25.286,
    "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": 870, "tid": 870,
    "ts": 7429770625295.648, "dur": 1.033,
    "args": {
      "External id": 4747,"Record function id": 0, "finished": true, "Ev Idx": 4746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770625311.806, "dur": 0.910,
    "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": 870, "tid": 870,
    "ts": 7429770625315.359, "dur": 1.146,
    "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": 870, "tid": 870,
    "ts": 7429770625317.884, "dur": 1.534,
    "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": 870, "tid": 870,
    "ts": 7429770625318.640, "dur": 0.331,
    "args": {
      "External id": 4751,"Record function id": 0, "finished": true, "Ev Idx": 4750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770625326.136, "dur": 31.743,
    "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": 870, "tid": 870,
    "ts": 7429770625326.747, "dur": 2.040,
    "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": 870, "tid": 870,
    "ts": 7429770625327.707, "dur": 0.725,
    "args": {
      "External id": 4754,"Record function id": 0, "finished": true, "Ev Idx": 4753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770625328.052, "dur": 0.255,
    "args": {
      "External id": 4755,"Record function id": 0, "finished": true, "Ev Idx": 4754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770625329.056, "dur": 28.728,
    "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": 870, "tid": 870,
    "ts": 7429770625329.603, "dur": 1.760,
    "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": 870, "tid": 870,
    "ts": 7429770625329.860, "dur": 1.386,
    "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": 870, "tid": 870,
    "ts": 7429770625331.633, "dur": 24.481,
    "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": 870, "tid": 870,
    "ts": 7429770625341.676, "dur": 1.307,
    "args": {
      "External id": 4760,"Record function id": 0, "finished": true, "Ev Idx": 4759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770625356.661, "dur": 0.881,
    "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": 870, "tid": 870,
    "ts": 7429770625360.175, "dur": 1.099,
    "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": 870, "tid": 870,
    "ts": 7429770625362.692, "dur": 1.850,
    "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": 870, "tid": 870,
    "ts": 7429770625363.600, "dur": 0.483,
    "args": {
      "External id": 4764,"Record function id": 0, "finished": true, "Ev Idx": 4763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770625367.151, "dur": 2.600,
    "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": 870, "tid": 870,
    "ts": 7429770625368.711, "dur": 0.541,
    "args": {
      "External id": 4766,"Record function id": 0, "finished": true, "Ev Idx": 4765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770625370.864, "dur": 1.015,
    "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": 870, "tid": 870,
    "ts": 7429770625371.281, "dur": 0.265,
    "args": {
      "External id": 4768,"Record function id": 0, "finished": true, "Ev Idx": 4767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770625372.658, "dur": 9.688,
    "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": 870, "tid": 870,
    "ts": 7429770625386.644, "dur": 3.435,
    "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": 870, "tid": 870,
    "ts": 7429770625388.495, "dur": 0.435,
    "args": {
      "External id": 4771,"Record function id": 0, "finished": true, "Ev Idx": 4770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770625392.306, "dur": 1.923,
    "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": 870, "tid": 870,
    "ts": 7429770625393.092, "dur": 0.345,
    "args": {
      "External id": 4773,"Record function id": 0, "finished": true, "Ev Idx": 4772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770625395.871, "dur": 11.079,
    "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": 870, "tid": 870,
    "ts": 7429770625410.877, "dur": 17.644,
    "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": 870, "tid": 870,
    "ts": 7429770625431.143, "dur": 9.603,
    "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": 870, "tid": 870,
    "ts": 7429770625442.340, "dur": 7.360,
    "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": 870, "tid": 870,
    "ts": 7429770625451.881, "dur": 8.475,
    "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": 870, "tid": 870,
    "ts": 7429770625462.676, "dur": 2.983,
    "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": 870, "tid": 870,
    "ts": 7429770625464.585, "dur": 0.396,
    "args": {
      "External id": 4780,"Record function id": 0, "finished": true, "Ev Idx": 4779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770625467.272, "dur": 2.311,
    "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": 870, "tid": 870,
    "ts": 7429770625468.495, "dur": 0.515,
    "args": {
      "External id": 4782,"Record function id": 0, "finished": true, "Ev Idx": 4781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770625470.538, "dur": 7.622,
    "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": 870, "tid": 870,
    "ts": 7429770625479.844, "dur": 9.896,
    "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": 870, "tid": 870,
    "ts": 7429770625491.981, "dur": 7.508,
    "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": 870, "tid": 870,
    "ts": 7429770625500.771, "dur": 8.627,
    "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": 870, "tid": 870,
    "ts": 7429770625532.787, "dur": 2.507,
    "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": 870, "tid": 870,
    "ts": 7429770625534.103, "dur": 0.450,
    "args": {
      "External id": 4788,"Record function id": 0, "finished": true, "Ev Idx": 4787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770625535.960, "dur": 1.830,
    "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": 870, "tid": 870,
    "ts": 7429770625536.884, "dur": 0.344,
    "args": {
      "External id": 4790,"Record function id": 0, "finished": true, "Ev Idx": 4789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770625538.420, "dur": 2.962,
    "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": 870, "tid": 870,
    "ts": 7429770625540.118, "dur": 0.616,
    "args": {
      "External id": 4792,"Record function id": 0, "finished": true, "Ev Idx": 4791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770625542.042, "dur": 2.452,
    "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": 870, "tid": 870,
    "ts": 7429770625543.469, "dur": 0.494,
    "args": {
      "External id": 4794,"Record function id": 0, "finished": true, "Ev Idx": 4793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770625545.136, "dur": 2.319,
    "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": 870, "tid": 870,
    "ts": 7429770625546.672, "dur": 0.307,
    "args": {
      "External id": 4796,"Record function id": 0, "finished": true, "Ev Idx": 4795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770625550.015, "dur": 3.004,
    "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": 870, "tid": 870,
    "ts": 7429770625551.980, "dur": 0.383,
    "args": {
      "External id": 4798,"Record function id": 0, "finished": true, "Ev Idx": 4797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770625555.245, "dur": 17.121,
    "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": 870, "tid": 870,
    "ts": 7429770625557.261, "dur": 12.946,
    "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": 870, "tid": 870,
    "ts": 7429770625558.502, "dur": 2.506,
    "args": {
      "External id": 4801,"Record function id": 0, "finished": true, "Ev Idx": 4800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770625559.209, "dur": 1.595,
    "args": {
      "External id": 4802,"Record function id": 0, "finished": true, "Ev Idx": 4801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770625561.628, "dur": 8.003,
    "args": {
      "External id": 4803,"Record function id": 0, "finished": true, "Ev Idx": 4802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770625570.667, "dur": 1.415,
    "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": 870, "tid": 870,
    "ts": 7429770625575.473, "dur": 2.892,
    "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": 870, "tid": 870,
    "ts": 7429770625577.261, "dur": 0.527,
    "args": {
      "External id": 4806,"Record function id": 0, "finished": true, "Ev Idx": 4805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770625579.109, "dur": 2.501,
    "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": 870, "tid": 870,
    "ts": 7429770625580.744, "dur": 0.385,
    "args": {
      "External id": 4808,"Record function id": 0, "finished": true, "Ev Idx": 4807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770625582.168, "dur": 1.407,
    "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": 870, "tid": 870,
    "ts": 7429770625582.812, "dur": 0.318,
    "args": {
      "External id": 4810,"Record function id": 0, "finished": true, "Ev Idx": 4809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770625584.147, "dur": 1.668,
    "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": 870, "tid": 870,
    "ts": 7429770625585.008, "dur": 0.305,
    "args": {
      "External id": 4812,"Record function id": 0, "finished": true, "Ev Idx": 4811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770625586.391, "dur": 1.428,
    "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": 870, "tid": 870,
    "ts": 7429770625587.056, "dur": 0.295,
    "args": {
      "External id": 4814,"Record function id": 0, "finished": true, "Ev Idx": 4813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770625589.272, "dur": 1.769,
    "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": 870, "tid": 870,
    "ts": 7429770625590.252, "dur": 0.330,
    "args": {
      "External id": 4816,"Record function id": 0, "finished": true, "Ev Idx": 4815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770625592.411, "dur": 12.953,
    "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": 870, "tid": 870,
    "ts": 7429770625593.299, "dur": 10.280,
    "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": 870, "tid": 870,
    "ts": 7429770625594.643, "dur": 1.848,
    "args": {
      "External id": 4819,"Record function id": 0, "finished": true, "Ev Idx": 4818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770625595.173, "dur": 1.152,
    "args": {
      "External id": 4820,"Record function id": 0, "finished": true, "Ev Idx": 4819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770625596.828, "dur": 6.396,
    "args": {
      "External id": 4821,"Record function id": 0, "finished": true, "Ev Idx": 4820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770625603.976, "dur": 1.215,
    "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": 870, "tid": 870,
    "ts": 7429770625607.114, "dur": 1.597,
    "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": 870, "tid": 870,
    "ts": 7429770625607.874, "dur": 0.361,
    "args": {
      "External id": 4824,"Record function id": 0, "finished": true, "Ev Idx": 4823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770625610.202, "dur": 113.660,
    "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": 870, "tid": 870,
    "ts": 7429770625610.693, "dur": 1.755,
    "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": 870, "tid": 870,
    "ts": 7429770625611.695, "dur": 0.331,
    "args": {
      "External id": 4827,"Record function id": 0, "finished": true, "Ev Idx": 4826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770625612.786, "dur": 3.402,
    "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": 870, "tid": 870,
    "ts": 7429770625614.370, "dur": 1.686,
    "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": 870, "tid": 870,
    "ts": 7429770625616.653, "dur": 2.601,
    "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": 870, "tid": 870,
    "ts": 7429770625618.585, "dur": 0.285,
    "args": {
      "External id": 4831,"Record function id": 0, "finished": true, "Ev Idx": 4830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770625619.557, "dur": 2.285,
    "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": 870, "tid": 870,
    "ts": 7429770625620.306, "dur": 1.435,
    "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": 870, "tid": 870,
    "ts": 7429770625622.478, "dur": 98.645,
    "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": 870, "tid": 870,
    "ts": 7429770625697.405, "dur": 1.571,
    "args": {
      "External id": 4835,"Record function id": 0, "finished": true, "Ev Idx": 4834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770625722.116, "dur": 1.386,
    "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": 870, "tid": 870,
    "ts": 7429770625727.911, "dur": 11.474,
    "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": 870, "tid": 870,
    "ts": 7429770625742.934, "dur": 2.547,
    "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": 870, "tid": 870,
    "ts": 7429770625744.132, "dur": 0.404,
    "args": {
      "External id": 4839,"Record function id": 0, "finished": true, "Ev Idx": 4838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770625746.181, "dur": 1.639,
    "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": 870, "tid": 870,
    "ts": 7429770625746.971, "dur": 0.352,
    "args": {
      "External id": 4841,"Record function id": 0, "finished": true, "Ev Idx": 4840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770625748.475, "dur": 1.178,
    "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": 870, "tid": 870,
    "ts": 7429770625749.073, "dur": 0.229,
    "args": {
      "External id": 4843,"Record function id": 0, "finished": true, "Ev Idx": 4842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770625751.320, "dur": 11.555,
    "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": 870, "tid": 870,
    "ts": 7429770625768.573, "dur": 14.289,
    "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": 870, "tid": 870,
    "ts": 7429770625769.561, "dur": 0.562,
    "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": 870, "tid": 870,
    "ts": 7429770625770.841, "dur": 11.763,
    "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": 870, "tid": 870,
    "ts": 7429770625785.913, "dur": 0.407,
    "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": 870, "tid": 870,
    "ts": 7429770625797.275, "dur": 0.460,
    "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": 870, "tid": 870,
    "ts": 7429770625799.509, "dur": 47.194,
    "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": 870, "tid": 870,
    "ts": 7429770625800.490, "dur": 2.998,
    "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": 870, "tid": 870,
    "ts": 7429770625802.503, "dur": 0.395,
    "args": {
      "External id": 4852,"Record function id": 0, "finished": true, "Ev Idx": 4851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770625804.164, "dur": 3.164,
    "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": 870, "tid": 870,
    "ts": 7429770625804.896, "dur": 2.290,
    "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": 870, "tid": 870,
    "ts": 7429770625807.795, "dur": 1.805,
    "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": 870, "tid": 870,
    "ts": 7429770625808.785, "dur": 0.436,
    "args": {
      "External id": 4856,"Record function id": 0, "finished": true, "Ev Idx": 4855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770625809.942, "dur": 1.515,
    "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": 870, "tid": 870,
    "ts": 7429770625810.333, "dur": 1.023,
    "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": 870, "tid": 870,
    "ts": 7429770625811.961, "dur": 32.506,
    "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": 870, "tid": 870,
    "ts": 7429770625825.182, "dur": 1.362,
    "args": {
      "External id": 4860,"Record function id": 0, "finished": true, "Ev Idx": 4859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770625845.141, "dur": 1.282,
    "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": 870, "tid": 870,
    "ts": 7429770625848.813, "dur": 2.471,
    "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": 870, "tid": 870,
    "ts": 7429770625850.188, "dur": 0.473,
    "args": {
      "External id": 4863,"Record function id": 0, "finished": true, "Ev Idx": 4862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770625853.729, "dur": 13.172,
    "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": 870, "tid": 870,
    "ts": 7429770625854.377, "dur": 12.376,
    "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": 870, "tid": 870,
    "ts": 7429770625855.571, "dur": 2.205,
    "args": {
      "External id": 4866,"Record function id": 0, "finished": true, "Ev Idx": 4865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770625856.119, "dur": 1.488,
    "args": {
      "External id": 4867,"Record function id": 0, "finished": true, "Ev Idx": 4866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770625858.339, "dur": 7.976,
    "args": {
      "External id": 4868,"Record function id": 0, "finished": true, "Ev Idx": 4867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770625873.500, "dur": 1.742,
    "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": 870, "tid": 870,
    "ts": 7429770625873.985, "dur": 1.151,
    "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": 870, "tid": 870,
    "ts": 7429770625892.787, "dur": 39.101,
    "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": 870, "tid": 870,
    "ts": 7429770625893.902, "dur": 4.033,
    "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": 870, "tid": 870,
    "ts": 7429770625896.107, "dur": 1.249,
    "args": {
      "External id": 4873,"Record function id": 0, "finished": true, "Ev Idx": 4872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770625896.692, "dur": 0.517,
    "args": {
      "External id": 4874,"Record function id": 0, "finished": true, "Ev Idx": 4873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770625898.373, "dur": 33.342,
    "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": 870, "tid": 870,
    "ts": 7429770625898.779, "dur": 2.259,
    "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": 870, "tid": 870,
    "ts": 7429770625899.053, "dur": 1.893,
    "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": 870, "tid": 870,
    "ts": 7429770625901.522, "dur": 28.220,
    "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": 870, "tid": 870,
    "ts": 7429770625912.722, "dur": 1.078,
    "args": {
      "External id": 4879,"Record function id": 0, "finished": true, "Ev Idx": 4878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770625930.292, "dur": 1.105,
    "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": 870, "tid": 870,
    "ts": 7429770625937.880, "dur": 8.576,
    "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": 870, "tid": 870,
    "ts": 7429770625956.821, "dur": 0.347,
    "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": 870, "tid": 870,
    "ts": 7429770625959.346, "dur": 11.696,
    "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": 870, "tid": 870,
    "ts": 7429770625961.248, "dur": 0.586,
    "args": {
      "External id": 4884,"Record function id": 0, "finished": true, "Ev Idx": 4883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770625962.216, "dur": 0.203,
    "args": {
      "External id": 4885,"Record function id": 0, "finished": true, "Ev Idx": 4884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770625973.967, "dur": 14.648,
    "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": 870, "tid": 870,
    "ts": 7429770625992.225, "dur": 9.374,
    "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": 870, "tid": 870,
    "ts": 7429770626003.584, "dur": 10.715,
    "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": 870, "tid": 870,
    "ts": 7429770626016.063, "dur": 8.261,
    "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": 870, "tid": 870,
    "ts": 7429770626026.762, "dur": 0.099,
    "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": 870, "tid": 870,
    "ts": 7429770626027.866, "dur": 8.214,
    "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": 870, "tid": 870,
    "ts": 7429770626051.064, "dur": 74.633,
    "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": 870, "tid": 870,
    "ts": 7429770626051.830, "dur": 3.385,
    "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": 870, "tid": 870,
    "ts": 7429770626053.232, "dur": 1.388,
    "args": {
      "External id": 4894,"Record function id": 0, "finished": true, "Ev Idx": 4893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770626054.050, "dur": 0.425,
    "args": {
      "External id": 4895,"Record function id": 0, "finished": true, "Ev Idx": 4894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770626055.617, "dur": 69.878,
    "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": 870, "tid": 870,
    "ts": 7429770626056.151, "dur": 2.427,
    "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": 870, "tid": 870,
    "ts": 7429770626056.470, "dur": 1.994,
    "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": 870, "tid": 870,
    "ts": 7429770626058.945, "dur": 63.542,
    "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": 870, "tid": 870,
    "ts": 7429770626098.382, "dur": 2.543,
    "args": {
      "External id": 4900,"Record function id": 0, "finished": true, "Ev Idx": 4899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770626123.405, "dur": 1.644,
    "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": 870, "tid": 870,
    "ts": 7429770626136.680, "dur": 12.632,
    "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": 870, "tid": 870,
    "ts": 7429770626159.584, "dur": 39.511,
    "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": 870, "tid": 870,
    "ts": 7429770626160.404, "dur": 5.344,
    "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": 870, "tid": 870,
    "ts": 7429770626162.961, "dur": 1.965,
    "args": {
      "External id": 4905,"Record function id": 0, "finished": true, "Ev Idx": 4904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770626163.965, "dur": 0.795,
    "args": {
      "External id": 4906,"Record function id": 0, "finished": true, "Ev Idx": 4905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770626166.143, "dur": 32.832,
    "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": 870, "tid": 870,
    "ts": 7429770626166.634, "dur": 1.995,
    "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": 870, "tid": 870,
    "ts": 7429770626167.278, "dur": 1.242,
    "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": 870, "tid": 870,
    "ts": 7429770626168.961, "dur": 27.983,
    "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": 870, "tid": 870,
    "ts": 7429770626180.899, "dur": 1.504,
    "args": {
      "External id": 4911,"Record function id": 0, "finished": true, "Ev Idx": 4910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770626197.591, "dur": 1.075,
    "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": 870, "tid": 870,
    "ts": 7429770626201.039, "dur": 8.920,
    "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": 870, "tid": 870,
    "ts": 7429770626216.236, "dur": 48.752,
    "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": 870, "tid": 870,
    "ts": 7429770626216.803, "dur": 2.564,
    "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": 870, "tid": 870,
    "ts": 7429770626217.844, "dur": 1.040,
    "args": {
      "External id": 4916,"Record function id": 0, "finished": true, "Ev Idx": 4915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770626218.421, "dur": 0.332,
    "args": {
      "External id": 4917,"Record function id": 0, "finished": true, "Ev Idx": 4916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770626219.646, "dur": 45.225,
    "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": 870, "tid": 870,
    "ts": 7429770626220.230, "dur": 1.469,
    "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": 870, "tid": 870,
    "ts": 7429770626220.504, "dur": 1.083,
    "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": 870, "tid": 870,
    "ts": 7429770626222.006, "dur": 41.106,
    "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": 870, "tid": 870,
    "ts": 7429770626234.121, "dur": 1.070,
    "args": {
      "External id": 4922,"Record function id": 0, "finished": true, "Ev Idx": 4921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770626263.690, "dur": 0.907,
    "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": 870, "tid": 870,
    "ts": 7429770626267.863, "dur": 8.234,
    "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": 870, "tid": 870,
    "ts": 7429770626297.051, "dur": 0.294,
    "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": 870, "tid": 870,
    "ts": 7429770626299.491, "dur": 11.500,
    "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": 870, "tid": 870,
    "ts": 7429770626301.077, "dur": 0.428,
    "args": {
      "External id": 4927,"Record function id": 0, "finished": true, "Ev Idx": 4926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770626301.847, "dur": 0.134,
    "args": {
      "External id": 4928,"Record function id": 0, "finished": true, "Ev Idx": 4927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770626313.614, "dur": 13.051,
    "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": 870, "tid": 870,
    "ts": 7429770626330.592, "dur": 7.539,
    "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": 870, "tid": 870,
    "ts": 7429770626340.215, "dur": 9.858,
    "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": 870, "tid": 870,
    "ts": 7429770626351.609, "dur": 7.872,
    "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": 870, "tid": 870,
    "ts": 7429770626361.876, "dur": 0.108,
    "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": 870, "tid": 870,
    "ts": 7429770626362.716, "dur": 9.280,
    "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": 870, "tid": 870,
    "ts": 7429770626390.627, "dur": 39.646,
    "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": 870, "tid": 870,
    "ts": 7429770626391.303, "dur": 3.756,
    "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": 870, "tid": 870,
    "ts": 7429770626392.398, "dur": 1.696,
    "args": {
      "External id": 4937,"Record function id": 0, "finished": true, "Ev Idx": 4936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770626393.295, "dur": 0.628,
    "args": {
      "External id": 4938,"Record function id": 0, "finished": true, "Ev Idx": 4937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770626395.677, "dur": 34.479,
    "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": 870, "tid": 870,
    "ts": 7429770626396.225, "dur": 1.933,
    "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": 870, "tid": 870,
    "ts": 7429770626396.548, "dur": 1.495,
    "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": 870, "tid": 870,
    "ts": 7429770626398.525, "dur": 29.232,
    "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": 870, "tid": 870,
    "ts": 7429770626409.889, "dur": 1.241,
    "args": {
      "External id": 4943,"Record function id": 0, "finished": true, "Ev Idx": 4942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770626428.538, "dur": 1.325,
    "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": 870, "tid": 870,
    "ts": 7429770626433.615, "dur": 1.470,
    "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": 870, "tid": 870,
    "ts": 7429770626437.077, "dur": 2.516,
    "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": 870, "tid": 870,
    "ts": 7429770626438.198, "dur": 0.677,
    "args": {
      "External id": 4947,"Record function id": 0, "finished": true, "Ev Idx": 4946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770626447.141, "dur": 32.518,
    "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": 870, "tid": 870,
    "ts": 7429770626447.548, "dur": 2.185,
    "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": 870, "tid": 870,
    "ts": 7429770626448.508, "dur": 0.773,
    "args": {
      "External id": 4950,"Record function id": 0, "finished": true, "Ev Idx": 4949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770626448.879, "dur": 0.273,
    "args": {
      "External id": 4951,"Record function id": 0, "finished": true, "Ev Idx": 4950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770626450.039, "dur": 29.543,
    "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": 870, "tid": 870,
    "ts": 7429770626450.840, "dur": 1.383,
    "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": 870, "tid": 870,
    "ts": 7429770626451.273, "dur": 0.863,
    "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": 870, "tid": 870,
    "ts": 7429770626452.732, "dur": 25.067,
    "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": 870, "tid": 870,
    "ts": 7429770626462.267, "dur": 1.145,
    "args": {
      "External id": 4956,"Record function id": 0, "finished": true, "Ev Idx": 4955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770626478.356, "dur": 0.982,
    "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": 870, "tid": 870,
    "ts": 7429770626481.923, "dur": 1.156,
    "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": 870, "tid": 870,
    "ts": 7429770626484.513, "dur": 1.857,
    "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": 870, "tid": 870,
    "ts": 7429770626485.358, "dur": 0.504,
    "args": {
      "External id": 4960,"Record function id": 0, "finished": true, "Ev Idx": 4959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770626493.058, "dur": 32.782,
    "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": 870, "tid": 870,
    "ts": 7429770626493.476, "dur": 2.231,
    "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": 870, "tid": 870,
    "ts": 7429770626494.590, "dur": 0.727,
    "args": {
      "External id": 4963,"Record function id": 0, "finished": true, "Ev Idx": 4962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770626494.946, "dur": 0.266,
    "args": {
      "External id": 4964,"Record function id": 0, "finished": true, "Ev Idx": 4963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770626495.971, "dur": 29.712,
    "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": 870, "tid": 870,
    "ts": 7429770626496.346, "dur": 1.403,
    "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": 870, "tid": 870,
    "ts": 7429770626496.760, "dur": 0.877,
    "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": 870, "tid": 870,
    "ts": 7429770626498.061, "dur": 25.792,
    "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": 870, "tid": 870,
    "ts": 7429770626508.811, "dur": 1.094,
    "args": {
      "External id": 4969,"Record function id": 0, "finished": true, "Ev Idx": 4968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770626524.522, "dur": 0.940,
    "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": 870, "tid": 870,
    "ts": 7429770626528.281, "dur": 1.300,
    "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": 870, "tid": 870,
    "ts": 7429770626531.096, "dur": 1.506,
    "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": 870, "tid": 870,
    "ts": 7429770626531.832, "dur": 0.307,
    "args": {
      "External id": 4973,"Record function id": 0, "finished": true, "Ev Idx": 4972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770626535.570, "dur": 2.103,
    "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": 870, "tid": 870,
    "ts": 7429770626536.825, "dur": 0.295,
    "args": {
      "External id": 4975,"Record function id": 0, "finished": true, "Ev Idx": 4974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770626538.759, "dur": 0.955,
    "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": 870, "tid": 870,
    "ts": 7429770626539.136, "dur": 0.248,
    "args": {
      "External id": 4977,"Record function id": 0, "finished": true, "Ev Idx": 4976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770626540.648, "dur": 9.774,
    "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": 870, "tid": 870,
    "ts": 7429770626555.677, "dur": 4.829,
    "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": 870, "tid": 870,
    "ts": 7429770626557.908, "dur": 0.669,
    "args": {
      "External id": 4980,"Record function id": 0, "finished": true, "Ev Idx": 4979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770626562.862, "dur": 2.379,
    "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": 870, "tid": 870,
    "ts": 7429770626564.326, "dur": 0.303,
    "args": {
      "External id": 4982,"Record function id": 0, "finished": true, "Ev Idx": 4981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770626567.050, "dur": 11.452,
    "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": 870, "tid": 870,
    "ts": 7429770626582.882, "dur": 18.906,
    "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": 870, "tid": 870,
    "ts": 7429770626604.645, "dur": 8.762,
    "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": 870, "tid": 870,
    "ts": 7429770626615.169, "dur": 10.471,
    "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": 870, "tid": 870,
    "ts": 7429770626627.794, "dur": 7.817,
    "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": 870, "tid": 870,
    "ts": 7429770626638.008, "dur": 2.157,
    "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": 870, "tid": 870,
    "ts": 7429770626638.992, "dur": 0.459,
    "args": {
      "External id": 4989,"Record function id": 0, "finished": true, "Ev Idx": 4988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770626641.899, "dur": 1.838,
    "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": 870, "tid": 870,
    "ts": 7429770626642.891, "dur": 0.305,
    "args": {
      "External id": 4991,"Record function id": 0, "finished": true, "Ev Idx": 4990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770626644.674, "dur": 7.880,
    "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": 870, "tid": 870,
    "ts": 7429770626654.078, "dur": 9.397,
    "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": 870, "tid": 870,
    "ts": 7429770626665.912, "dur": 9.634,
    "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": 870, "tid": 870,
    "ts": 7429770626677.028, "dur": 6.928,
    "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": 870, "tid": 870,
    "ts": 7429770626708.062, "dur": 2.624,
    "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": 870, "tid": 870,
    "ts": 7429770626709.289, "dur": 0.521,
    "args": {
      "External id": 4997,"Record function id": 0, "finished": true, "Ev Idx": 4996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770626711.567, "dur": 2.835,
    "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": 870, "tid": 870,
    "ts": 7429770626713.562, "dur": 0.313,
    "args": {
      "External id": 4999,"Record function id": 0, "finished": true, "Ev Idx": 4998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770626715.117, "dur": 3.111,
    "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": 870, "tid": 870,
    "ts": 7429770626716.603, "dur": 0.910,
    "args": {
      "External id": 5001,"Record function id": 0, "finished": true, "Ev Idx": 5000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770626719.009, "dur": 2.319,
    "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": 870, "tid": 870,
    "ts": 7429770626720.422, "dur": 0.387,
    "args": {
      "External id": 5003,"Record function id": 0, "finished": true, "Ev Idx": 5002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770626721.976, "dur": 2.025,
    "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": 870, "tid": 870,
    "ts": 7429770626723.062, "dur": 0.484,
    "args": {
      "External id": 5005,"Record function id": 0, "finished": true, "Ev Idx": 5004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770626726.730, "dur": 2.925,
    "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": 870, "tid": 870,
    "ts": 7429770626728.521, "dur": 0.373,
    "args": {
      "External id": 5007,"Record function id": 0, "finished": true, "Ev Idx": 5006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770626732.041, "dur": 19.877,
    "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": 870, "tid": 870,
    "ts": 7429770626734.545, "dur": 14.467,
    "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": 870, "tid": 870,
    "ts": 7429770626736.108, "dur": 2.641,
    "args": {
      "External id": 5010,"Record function id": 0, "finished": true, "Ev Idx": 5009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770626736.828, "dur": 1.676,
    "args": {
      "External id": 5011,"Record function id": 0, "finished": true, "Ev Idx": 5010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770626739.353, "dur": 9.143,
    "args": {
      "External id": 5012,"Record function id": 0, "finished": true, "Ev Idx": 5011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770626749.601, "dur": 2.013,
    "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": 870, "tid": 870,
    "ts": 7429770626755.181, "dur": 2.148,
    "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": 870, "tid": 870,
    "ts": 7429770626756.398, "dur": 0.400,
    "args": {
      "External id": 5015,"Record function id": 0, "finished": true, "Ev Idx": 5014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770626757.948, "dur": 1.656,
    "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": 870, "tid": 870,
    "ts": 7429770626758.850, "dur": 0.290,
    "args": {
      "External id": 5017,"Record function id": 0, "finished": true, "Ev Idx": 5016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770626760.128, "dur": 1.571,
    "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": 870, "tid": 870,
    "ts": 7429770626760.905, "dur": 0.350,
    "args": {
      "External id": 5019,"Record function id": 0, "finished": true, "Ev Idx": 5018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770626762.256, "dur": 1.746,
    "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": 870, "tid": 870,
    "ts": 7429770626763.226, "dur": 0.291,
    "args": {
      "External id": 5021,"Record function id": 0, "finished": true, "Ev Idx": 5020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770626764.691, "dur": 2.055,
    "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": 870, "tid": 870,
    "ts": 7429770626765.841, "dur": 0.458,
    "args": {
      "External id": 5023,"Record function id": 0, "finished": true, "Ev Idx": 5022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770626768.490, "dur": 3.122,
    "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": 870, "tid": 870,
    "ts": 7429770626770.565, "dur": 0.619,
    "args": {
      "External id": 5025,"Record function id": 0, "finished": true, "Ev Idx": 5024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770626773.149, "dur": 13.574,
    "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": 870, "tid": 870,
    "ts": 7429770626774.101, "dur": 10.987,
    "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": 870, "tid": 870,
    "ts": 7429770626775.594, "dur": 1.900,
    "args": {
      "External id": 5028,"Record function id": 0, "finished": true, "Ev Idx": 5027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770626776.076, "dur": 1.299,
    "args": {
      "External id": 5029,"Record function id": 0, "finished": true, "Ev Idx": 5028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770626777.808, "dur": 6.934,
    "args": {
      "External id": 5030,"Record function id": 0, "finished": true, "Ev Idx": 5029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770626785.474, "dur": 1.091,
    "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": 870, "tid": 870,
    "ts": 7429770626788.594, "dur": 1.931,
    "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": 870, "tid": 870,
    "ts": 7429770626789.664, "dur": 0.347,
    "args": {
      "External id": 5033,"Record function id": 0, "finished": true, "Ev Idx": 5032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770626792.141, "dur": 152.807,
    "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": 870, "tid": 870,
    "ts": 7429770626792.674, "dur": 1.541,
    "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": 870, "tid": 870,
    "ts": 7429770626793.449, "dur": 0.343,
    "args": {
      "External id": 5036,"Record function id": 0, "finished": true, "Ev Idx": 5035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770626794.593, "dur": 5.376,
    "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": 870, "tid": 870,
    "ts": 7429770626796.534, "dur": 3.282,
    "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": 870, "tid": 870,
    "ts": 7429770626800.436, "dur": 2.102,
    "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": 870, "tid": 870,
    "ts": 7429770626801.898, "dur": 0.285,
    "args": {
      "External id": 5040,"Record function id": 0, "finished": true, "Ev Idx": 5039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770626802.847, "dur": 1.758,
    "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": 870, "tid": 870,
    "ts": 7429770626803.673, "dur": 0.827,
    "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": 870, "tid": 870,
    "ts": 7429770626805.185, "dur": 137.238,
    "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": 870, "tid": 870,
    "ts": 7429770626885.392, "dur": 33.982,
    "args": {
      "External id": 5044,"Record function id": 0, "finished": true, "Ev Idx": 5043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770626943.189, "dur": 1.368,
    "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": 870, "tid": 870,
    "ts": 7429770626949.295, "dur": 12.319,
    "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": 870, "tid": 870,
    "ts": 7429770626965.329, "dur": 2.627,
    "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": 870, "tid": 870,
    "ts": 7429770626966.544, "dur": 0.465,
    "args": {
      "External id": 5048,"Record function id": 0, "finished": true, "Ev Idx": 5047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770626968.758, "dur": 1.744,
    "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": 870, "tid": 870,
    "ts": 7429770626969.589, "dur": 0.304,
    "args": {
      "External id": 5050,"Record function id": 0, "finished": true, "Ev Idx": 5049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770626971.300, "dur": 1.246,
    "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": 870, "tid": 870,
    "ts": 7429770626971.953, "dur": 0.269,
    "args": {
      "External id": 5052,"Record function id": 0, "finished": true, "Ev Idx": 5051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770626974.403, "dur": 8.473,
    "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": 870, "tid": 870,
    "ts": 7429770626988.527, "dur": 15.668,
    "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": 870, "tid": 870,
    "ts": 7429770626989.581, "dur": 0.439,
    "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": 870, "tid": 870,
    "ts": 7429770626990.780, "dur": 13.193,
    "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": 870, "tid": 870,
    "ts": 7429770627006.893, "dur": 0.247,
    "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": 870, "tid": 870,
    "ts": 7429770627018.869, "dur": 0.533,
    "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": 870, "tid": 870,
    "ts": 7429770627021.335, "dur": 44.724,
    "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": 870, "tid": 870,
    "ts": 7429770627022.116, "dur": 2.330,
    "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": 870, "tid": 870,
    "ts": 7429770627023.516, "dur": 0.363,
    "args": {
      "External id": 5061,"Record function id": 0, "finished": true, "Ev Idx": 5060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770627024.830, "dur": 2.731,
    "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": 870, "tid": 870,
    "ts": 7429770627025.343, "dur": 2.088,
    "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": 870, "tid": 870,
    "ts": 7429770627028.005, "dur": 2.610,
    "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": 870, "tid": 870,
    "ts": 7429770627029.969, "dur": 0.275,
    "args": {
      "External id": 5065,"Record function id": 0, "finished": true, "Ev Idx": 5064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770627030.940, "dur": 2.055,
    "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": 870, "tid": 870,
    "ts": 7429770627031.308, "dur": 1.613,
    "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": 870, "tid": 870,
    "ts": 7429770627033.394, "dur": 30.714,
    "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": 870, "tid": 870,
    "ts": 7429770627045.288, "dur": 1.423,
    "args": {
      "External id": 5069,"Record function id": 0, "finished": true, "Ev Idx": 5068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770627064.688, "dur": 1.113,
    "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": 870, "tid": 870,
    "ts": 7429770627068.212, "dur": 2.256,
    "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": 870, "tid": 870,
    "ts": 7429770627069.491, "dur": 0.358,
    "args": {
      "External id": 5072,"Record function id": 0, "finished": true, "Ev Idx": 5071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770627072.828, "dur": 13.774,
    "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": 870, "tid": 870,
    "ts": 7429770627073.526, "dur": 12.925,
    "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": 870, "tid": 870,
    "ts": 7429770627074.696, "dur": 2.849,
    "args": {
      "External id": 5075,"Record function id": 0, "finished": true, "Ev Idx": 5074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770627075.656, "dur": 1.732,
    "args": {
      "External id": 5076,"Record function id": 0, "finished": true, "Ev Idx": 5075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770627077.983, "dur": 8.039,
    "args": {
      "External id": 5077,"Record function id": 0, "finished": true, "Ev Idx": 5076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770627092.783, "dur": 2.109,
    "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": 870, "tid": 870,
    "ts": 7429770627093.456, "dur": 1.335,
    "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": 870, "tid": 870,
    "ts": 7429770627110.793, "dur": 41.380,
    "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": 870, "tid": 870,
    "ts": 7429770627111.676, "dur": 4.221,
    "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": 870, "tid": 870,
    "ts": 7429770627113.905, "dur": 1.366,
    "args": {
      "External id": 5082,"Record function id": 0, "finished": true, "Ev Idx": 5081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770627114.664, "dur": 0.430,
    "args": {
      "External id": 5083,"Record function id": 0, "finished": true, "Ev Idx": 5082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770627116.305, "dur": 35.639,
    "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": 870, "tid": 870,
    "ts": 7429770627116.908, "dur": 2.505,
    "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": 870, "tid": 870,
    "ts": 7429770627117.367, "dur": 1.967,
    "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": 870, "tid": 870,
    "ts": 7429770627119.857, "dur": 30.093,
    "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": 870, "tid": 870,
    "ts": 7429770627132.857, "dur": 1.184,
    "args": {
      "External id": 5088,"Record function id": 0, "finished": true, "Ev Idx": 5087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770627150.562, "dur": 1.079,
    "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": 870, "tid": 870,
    "ts": 7429770627158.331, "dur": 9.859,
    "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": 870, "tid": 870,
    "ts": 7429770627177.414, "dur": 0.323,
    "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": 870, "tid": 870,
    "ts": 7429770627180.208, "dur": 12.101,
    "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": 870, "tid": 870,
    "ts": 7429770627182.275, "dur": 0.594,
    "args": {
      "External id": 5093,"Record function id": 0, "finished": true, "Ev Idx": 5092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770627183.281, "dur": 0.128,
    "args": {
      "External id": 5094,"Record function id": 0, "finished": true, "Ev Idx": 5093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770627194.888, "dur": 14.762,
    "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": 870, "tid": 870,
    "ts": 7429770627213.169, "dur": 8.007,
    "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": 870, "tid": 870,
    "ts": 7429770627223.158, "dur": 10.868,
    "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": 870, "tid": 870,
    "ts": 7429770627235.783, "dur": 8.683,
    "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": 870, "tid": 870,
    "ts": 7429770627246.956, "dur": 0.101,
    "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": 870, "tid": 870,
    "ts": 7429770627248.222, "dur": 8.918,
    "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": 870, "tid": 870,
    "ts": 7429770627272.192, "dur": 40.717,
    "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": 870, "tid": 870,
    "ts": 7429770627272.604, "dur": 3.562,
    "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": 870, "tid": 870,
    "ts": 7429770627273.940, "dur": 1.570,
    "args": {
      "External id": 5103,"Record function id": 0, "finished": true, "Ev Idx": 5102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770627274.772, "dur": 0.570,
    "args": {
      "External id": 5104,"Record function id": 0, "finished": true, "Ev Idx": 5103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770627276.523, "dur": 36.197,
    "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": 870, "tid": 870,
    "ts": 7429770627277.194, "dur": 2.063,
    "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": 870, "tid": 870,
    "ts": 7429770627277.621, "dur": 1.519,
    "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": 870, "tid": 870,
    "ts": 7429770627279.647, "dur": 31.077,
    "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": 870, "tid": 870,
    "ts": 7429770627293.203, "dur": 1.331,
    "args": {
      "External id": 5109,"Record function id": 0, "finished": true, "Ev Idx": 5108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770627311.322, "dur": 1.094,
    "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": 870, "tid": 870,
    "ts": 7429770627320.713, "dur": 10.664,
    "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": 870, "tid": 870,
    "ts": 7429770627338.921, "dur": 72.275,
    "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": 870, "tid": 870,
    "ts": 7429770627339.260, "dur": 2.247,
    "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": 870, "tid": 870,
    "ts": 7429770627340.097, "dur": 0.899,
    "args": {
      "External id": 5114,"Record function id": 0, "finished": true, "Ev Idx": 5113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770627340.539, "dur": 0.347,
    "args": {
      "External id": 5115,"Record function id": 0, "finished": true, "Ev Idx": 5114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770627341.839, "dur": 69.203,
    "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": 870, "tid": 870,
    "ts": 7429770627342.248, "dur": 3.242,
    "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": 870, "tid": 870,
    "ts": 7429770627342.674, "dur": 2.707,
    "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": 870, "tid": 870,
    "ts": 7429770627345.935, "dur": 27.693,
    "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": 870, "tid": 870,
    "ts": 7429770627354.288, "dur": 1.147,
    "args": {
      "External id": 5120,"Record function id": 0, "finished": true, "Ev Idx": 5119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770627409.718, "dur": 1.080,
    "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": 870, "tid": 870,
    "ts": 7429770627413.085, "dur": 8.808,
    "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": 870, "tid": 870,
    "ts": 7429770627428.368, "dur": 42.540,
    "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": 870, "tid": 870,
    "ts": 7429770627428.711, "dur": 2.322,
    "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": 870, "tid": 870,
    "ts": 7429770627429.660, "dur": 0.842,
    "args": {
      "External id": 5125,"Record function id": 0, "finished": true, "Ev Idx": 5124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770627430.047, "dur": 0.340,
    "args": {
      "External id": 5126,"Record function id": 0, "finished": true, "Ev Idx": 5125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770627431.323, "dur": 39.467,
    "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": 870, "tid": 870,
    "ts": 7429770627431.745, "dur": 1.332,
    "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": 870, "tid": 870,
    "ts": 7429770627432.020, "dur": 0.945,
    "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": 870, "tid": 870,
    "ts": 7429770627433.370, "dur": 35.713,
    "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": 870, "tid": 870,
    "ts": 7429770627440.421, "dur": 1.171,
    "args": {
      "External id": 5131,"Record function id": 0, "finished": true, "Ev Idx": 5130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770627469.629, "dur": 0.905,
    "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": 870, "tid": 870,
    "ts": 7429770627473.720, "dur": 8.718,
    "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": 870, "tid": 870,
    "ts": 7429770627499.999, "dur": 0.157,
    "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": 870, "tid": 870,
    "ts": 7429770627501.994, "dur": 9.931,
    "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": 870, "tid": 870,
    "ts": 7429770627503.268, "dur": 0.284,
    "args": {
      "External id": 5136,"Record function id": 0, "finished": true, "Ev Idx": 5135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770627503.856, "dur": 0.070,
    "args": {
      "External id": 5137,"Record function id": 0, "finished": true, "Ev Idx": 5136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770627514.206, "dur": 12.519,
    "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": 870, "tid": 870,
    "ts": 7429770627530.086, "dur": 7.274,
    "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": 870, "tid": 870,
    "ts": 7429770627538.934, "dur": 9.245,
    "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": 870, "tid": 870,
    "ts": 7429770627549.639, "dur": 7.351,
    "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": 870, "tid": 870,
    "ts": 7429770627559.355, "dur": 0.099,
    "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": 870, "tid": 870,
    "ts": 7429770627560.165, "dur": 9.101,
    "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": 870, "tid": 870,
    "ts": 7429770627587.188, "dur": 36.973,
    "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": 870, "tid": 870,
    "ts": 7429770627587.582, "dur": 3.174,
    "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": 870, "tid": 870,
    "ts": 7429770627588.489, "dur": 1.325,
    "args": {
      "External id": 5146,"Record function id": 0, "finished": true, "Ev Idx": 5145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770627589.133, "dur": 0.525,
    "args": {
      "External id": 5147,"Record function id": 0, "finished": true, "Ev Idx": 5146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770627591.094, "dur": 32.951,
    "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": 870, "tid": 870,
    "ts": 7429770627591.596, "dur": 3.056,
    "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": 870, "tid": 870,
    "ts": 7429770627591.892, "dur": 2.623,
    "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": 870, "tid": 870,
    "ts": 7429770627595.002, "dur": 27.101,
    "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": 870, "tid": 870,
    "ts": 7429770627604.940, "dur": 1.112,
    "args": {
      "External id": 5152,"Record function id": 0, "finished": true, "Ev Idx": 5151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770627622.701, "dur": 1.071,
    "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": 870, "tid": 870,
    "ts": 7429770627627.247, "dur": 1.323,
    "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": 870, "tid": 870,
    "ts": 7429770627630.373, "dur": 2.133,
    "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": 870, "tid": 870,
    "ts": 7429770627631.241, "dur": 0.465,
    "args": {
      "External id": 5156,"Record function id": 0, "finished": true, "Ev Idx": 5155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770627640.047, "dur": 31.372,
    "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": 870, "tid": 870,
    "ts": 7429770627640.329, "dur": 2.114,
    "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": 870, "tid": 870,
    "ts": 7429770627641.126, "dur": 0.875,
    "args": {
      "External id": 5159,"Record function id": 0, "finished": true, "Ev Idx": 5158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770627641.627, "dur": 0.279,
    "args": {
      "External id": 5160,"Record function id": 0, "finished": true, "Ev Idx": 5159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770627642.801, "dur": 28.521,
    "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": 870, "tid": 870,
    "ts": 7429770627643.201, "dur": 1.234,
    "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": 870, "tid": 870,
    "ts": 7429770627643.462, "dur": 0.875,
    "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": 870, "tid": 870,
    "ts": 7429770627644.762, "dur": 24.857,
    "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": 870, "tid": 870,
    "ts": 7429770627653.794, "dur": 1.108,
    "args": {
      "External id": 5165,"Record function id": 0, "finished": true, "Ev Idx": 5164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770627670.155, "dur": 0.923,
    "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": 870, "tid": 870,
    "ts": 7429770627673.748, "dur": 1.146,
    "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": 870, "tid": 870,
    "ts": 7429770627676.373, "dur": 1.595,
    "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": 870, "tid": 870,
    "ts": 7429770627677.134, "dur": 0.356,
    "args": {
      "External id": 5169,"Record function id": 0, "finished": true, "Ev Idx": 5168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770627684.738, "dur": 30.196,
    "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": 870, "tid": 870,
    "ts": 7429770627685.021, "dur": 1.886,
    "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": 870, "tid": 870,
    "ts": 7429770627685.766, "dur": 0.756,
    "args": {
      "External id": 5172,"Record function id": 0, "finished": true, "Ev Idx": 5171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770627686.132, "dur": 0.290,
    "args": {
      "External id": 5173,"Record function id": 0, "finished": true, "Ev Idx": 5172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770627687.190, "dur": 27.668,
    "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": 870, "tid": 870,
    "ts": 7429770627687.584, "dur": 1.244,
    "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": 870, "tid": 870,
    "ts": 7429770627687.852, "dur": 0.876,
    "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": 870, "tid": 870,
    "ts": 7429770627689.130, "dur": 24.100,
    "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": 870, "tid": 870,
    "ts": 7429770627697.854, "dur": 1.047,
    "args": {
      "External id": 5178,"Record function id": 0, "finished": true, "Ev Idx": 5177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770627713.746, "dur": 0.849,
    "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": 870, "tid": 870,
    "ts": 7429770627717.251, "dur": 1.151,
    "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": 870, "tid": 870,
    "ts": 7429770627719.974, "dur": 1.470,
    "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": 870, "tid": 870,
    "ts": 7429770627720.640, "dur": 0.327,
    "args": {
      "External id": 5182,"Record function id": 0, "finished": true, "Ev Idx": 5181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770627724.418, "dur": 2.033,
    "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": 870, "tid": 870,
    "ts": 7429770627725.616, "dur": 0.283,
    "args": {
      "External id": 5184,"Record function id": 0, "finished": true, "Ev Idx": 5183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770627727.526, "dur": 0.998,
    "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": 870, "tid": 870,
    "ts": 7429770627727.943, "dur": 0.274,
    "args": {
      "External id": 5186,"Record function id": 0, "finished": true, "Ev Idx": 5185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770627729.524, "dur": 10.210,
    "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": 870, "tid": 870,
    "ts": 7429770627744.663, "dur": 3.390,
    "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": 870, "tid": 870,
    "ts": 7429770627746.569, "dur": 0.425,
    "args": {
      "External id": 5189,"Record function id": 0, "finished": true, "Ev Idx": 5188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770627750.646, "dur": 3.690,
    "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": 870, "tid": 870,
    "ts": 7429770627752.555, "dur": 0.696,
    "args": {
      "External id": 5191,"Record function id": 0, "finished": true, "Ev Idx": 5190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770627756.282, "dur": 11.590,
    "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": 870, "tid": 870,
    "ts": 7429770627772.257, "dur": 16.205,
    "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": 870, "tid": 870,
    "ts": 7429770627791.593, "dur": 9.552,
    "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": 870, "tid": 870,
    "ts": 7429770627802.726, "dur": 7.112,
    "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": 870, "tid": 870,
    "ts": 7429770627812.120, "dur": 8.429,
    "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": 870, "tid": 870,
    "ts": 7429770627823.135, "dur": 2.499,
    "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": 870, "tid": 870,
    "ts": 7429770627824.593, "dur": 0.397,
    "args": {
      "External id": 5198,"Record function id": 0, "finished": true, "Ev Idx": 5197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770627827.466, "dur": 1.625,
    "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": 870, "tid": 870,
    "ts": 7429770627828.262, "dur": 0.310,
    "args": {
      "External id": 5200,"Record function id": 0, "finished": true, "Ev Idx": 5199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770627830.256, "dur": 7.320,
    "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": 870, "tid": 870,
    "ts": 7429770627839.247, "dur": 9.563,
    "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": 870, "tid": 870,
    "ts": 7429770627850.966, "dur": 7.882,
    "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": 870, "tid": 870,
    "ts": 7429770627860.134, "dur": 8.563,
    "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": 870, "tid": 870,
    "ts": 7429770627892.616, "dur": 2.254,
    "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": 870, "tid": 870,
    "ts": 7429770627893.719, "dur": 0.418,
    "args": {
      "External id": 5206,"Record function id": 0, "finished": true, "Ev Idx": 5205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770627895.494, "dur": 1.593,
    "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": 870, "tid": 870,
    "ts": 7429770627896.243, "dur": 0.307,
    "args": {
      "External id": 5208,"Record function id": 0, "finished": true, "Ev Idx": 5207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770627897.788, "dur": 3.451,
    "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": 870, "tid": 870,
    "ts": 7429770627899.831, "dur": 0.767,
    "args": {
      "External id": 5210,"Record function id": 0, "finished": true, "Ev Idx": 5209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770627901.838, "dur": 2.847,
    "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": 870, "tid": 870,
    "ts": 7429770627903.708, "dur": 0.468,
    "args": {
      "External id": 5212,"Record function id": 0, "finished": true, "Ev Idx": 5211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770627905.295, "dur": 1.965,
    "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": 870, "tid": 870,
    "ts": 7429770627906.467, "dur": 0.305,
    "args": {
      "External id": 5214,"Record function id": 0, "finished": true, "Ev Idx": 5213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770627909.756, "dur": 2.752,
    "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": 870, "tid": 870,
    "ts": 7429770627911.445, "dur": 0.386,
    "args": {
      "External id": 5216,"Record function id": 0, "finished": true, "Ev Idx": 5215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770627914.842, "dur": 17.232,
    "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": 870, "tid": 870,
    "ts": 7429770627917.263, "dur": 12.905,
    "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": 870, "tid": 870,
    "ts": 7429770627918.464, "dur": 2.645,
    "args": {
      "External id": 5219,"Record function id": 0, "finished": true, "Ev Idx": 5218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770627919.220, "dur": 1.689,
    "args": {
      "External id": 5220,"Record function id": 0, "finished": true, "Ev Idx": 5219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770627921.715, "dur": 7.887,
    "args": {
      "External id": 5221,"Record function id": 0, "finished": true, "Ev Idx": 5220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770627930.620, "dur": 1.232,
    "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": 870, "tid": 870,
    "ts": 7429770627935.495, "dur": 2.525,
    "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": 870, "tid": 870,
    "ts": 7429770627936.938, "dur": 0.527,
    "args": {
      "External id": 5224,"Record function id": 0, "finished": true, "Ev Idx": 5223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770627938.601, "dur": 2.167,
    "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": 870, "tid": 870,
    "ts": 7429770627939.999, "dur": 0.289,
    "args": {
      "External id": 5226,"Record function id": 0, "finished": true, "Ev Idx": 5225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770627941.293, "dur": 1.409,
    "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": 870, "tid": 870,
    "ts": 7429770627941.976, "dur": 0.306,
    "args": {
      "External id": 5228,"Record function id": 0, "finished": true, "Ev Idx": 5227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770627943.428, "dur": 1.461,
    "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": 870, "tid": 870,
    "ts": 7429770627944.116, "dur": 0.271,
    "args": {
      "External id": 5230,"Record function id": 0, "finished": true, "Ev Idx": 5229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770627945.464, "dur": 1.456,
    "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": 870, "tid": 870,
    "ts": 7429770627946.169, "dur": 0.273,
    "args": {
      "External id": 5232,"Record function id": 0, "finished": true, "Ev Idx": 5231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770627948.363, "dur": 1.521,
    "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": 870, "tid": 870,
    "ts": 7429770627949.116, "dur": 0.336,
    "args": {
      "External id": 5234,"Record function id": 0, "finished": true, "Ev Idx": 5233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770627951.336, "dur": 13.898,
    "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": 870, "tid": 870,
    "ts": 7429770627952.222, "dur": 11.551,
    "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": 870, "tid": 870,
    "ts": 7429770627954.068, "dur": 1.878,
    "args": {
      "External id": 5237,"Record function id": 0, "finished": true, "Ev Idx": 5236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770627954.377, "dur": 1.430,
    "args": {
      "External id": 5238,"Record function id": 0, "finished": true, "Ev Idx": 5237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770627956.275, "dur": 7.161,
    "args": {
      "External id": 5239,"Record function id": 0, "finished": true, "Ev Idx": 5238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770627964.165, "dur": 0.888,
    "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": 870, "tid": 870,
    "ts": 7429770627967.073, "dur": 1.645,
    "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": 870, "tid": 870,
    "ts": 7429770627967.904, "dur": 0.346,
    "args": {
      "External id": 5242,"Record function id": 0, "finished": true, "Ev Idx": 5241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770627970.138, "dur": 112.671,
    "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": 870, "tid": 870,
    "ts": 7429770627970.663, "dur": 1.479,
    "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": 870, "tid": 870,
    "ts": 7429770627971.393, "dur": 0.320,
    "args": {
      "External id": 5245,"Record function id": 0, "finished": true, "Ev Idx": 5244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770627972.482, "dur": 3.717,
    "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": 870, "tid": 870,
    "ts": 7429770627974.246, "dur": 1.844,
    "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": 870, "tid": 870,
    "ts": 7429770627976.658, "dur": 2.364,
    "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": 870, "tid": 870,
    "ts": 7429770627978.236, "dur": 0.425,
    "args": {
      "External id": 5249,"Record function id": 0, "finished": true, "Ev Idx": 5248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770627979.296, "dur": 2.153,
    "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": 870, "tid": 870,
    "ts": 7429770627980.106, "dur": 1.244,
    "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": 870, "tid": 870,
    "ts": 7429770627981.895, "dur": 98.569,
    "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": 870, "tid": 870,
    "ts": 7429770628058.348, "dur": 1.454,
    "args": {
      "External id": 5253,"Record function id": 0, "finished": true, "Ev Idx": 5252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770628081.225, "dur": 1.247,
    "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": 870, "tid": 870,
    "ts": 7429770628087.144, "dur": 11.931,
    "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": 870, "tid": 870,
    "ts": 7429770628102.728, "dur": 2.915,
    "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": 870, "tid": 870,
    "ts": 7429770628104.075, "dur": 0.589,
    "args": {
      "External id": 5257,"Record function id": 0, "finished": true, "Ev Idx": 5256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770628106.354, "dur": 1.344,
    "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": 870, "tid": 870,
    "ts": 7429770628106.922, "dur": 0.318,
    "args": {
      "External id": 5259,"Record function id": 0, "finished": true, "Ev Idx": 5258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770628108.355, "dur": 1.097,
    "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": 870, "tid": 870,
    "ts": 7429770628108.923, "dur": 0.214,
    "args": {
      "External id": 5261,"Record function id": 0, "finished": true, "Ev Idx": 5260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770628111.286, "dur": 11.156,
    "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": 870, "tid": 870,
    "ts": 7429770628128.576, "dur": 15.071,
    "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": 870, "tid": 870,
    "ts": 7429770628129.751, "dur": 0.427,
    "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": 870, "tid": 870,
    "ts": 7429770628130.876, "dur": 12.516,
    "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": 870, "tid": 870,
    "ts": 7429770628146.645, "dur": 0.296,
    "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": 870, "tid": 870,
    "ts": 7429770628157.915, "dur": 0.539,
    "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": 870, "tid": 870,
    "ts": 7429770628160.255, "dur": 45.414,
    "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": 870, "tid": 870,
    "ts": 7429770628160.814, "dur": 2.867,
    "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": 870, "tid": 870,
    "ts": 7429770628162.656, "dur": 0.421,
    "args": {
      "External id": 5270,"Record function id": 0, "finished": true, "Ev Idx": 5269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770628164.094, "dur": 2.745,
    "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": 870, "tid": 870,
    "ts": 7429770628164.459, "dur": 2.257,
    "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": 870, "tid": 870,
    "ts": 7429770628167.289, "dur": 1.545,
    "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": 870, "tid": 870,
    "ts": 7429770628167.973, "dur": 0.439,
    "args": {
      "External id": 5274,"Record function id": 0, "finished": true, "Ev Idx": 5273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770628169.343, "dur": 1.207,
    "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": 870, "tid": 870,
    "ts": 7429770628169.579, "dur": 0.897,
    "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": 870, "tid": 870,
    "ts": 7429770628171.026, "dur": 32.786,
    "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": 870, "tid": 870,
    "ts": 7429770628184.506, "dur": 1.636,
    "args": {
      "External id": 5278,"Record function id": 0, "finished": true, "Ev Idx": 5277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770628204.373, "dur": 1.004,
    "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": 870, "tid": 870,
    "ts": 7429770628207.891, "dur": 2.332,
    "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": 870, "tid": 870,
    "ts": 7429770628209.109, "dur": 0.447,
    "args": {
      "External id": 5281,"Record function id": 0, "finished": true, "Ev Idx": 5280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770628212.666, "dur": 13.107,
    "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": 870, "tid": 870,
    "ts": 7429770628213.298, "dur": 12.336,
    "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": 870, "tid": 870,
    "ts": 7429770628214.313, "dur": 2.291,
    "args": {
      "External id": 5284,"Record function id": 0, "finished": true, "Ev Idx": 5283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770628214.831, "dur": 1.606,
    "args": {
      "External id": 5285,"Record function id": 0, "finished": true, "Ev Idx": 5284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770628217.055, "dur": 8.144,
    "args": {
      "External id": 5286,"Record function id": 0, "finished": true, "Ev Idx": 5285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770628232.301, "dur": 1.896,
    "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": 870, "tid": 870,
    "ts": 7429770628232.613, "dur": 1.501,
    "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": 870, "tid": 870,
    "ts": 7429770628251.281, "dur": 38.134,
    "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": 870, "tid": 870,
    "ts": 7429770628251.898, "dur": 3.978,
    "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": 870, "tid": 870,
    "ts": 7429770628254.265, "dur": 1.014,
    "args": {
      "External id": 5291,"Record function id": 0, "finished": true, "Ev Idx": 5290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770628254.730, "dur": 0.415,
    "args": {
      "External id": 5292,"Record function id": 0, "finished": true, "Ev Idx": 5291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770628256.277, "dur": 32.956,
    "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": 870, "tid": 870,
    "ts": 7429770628256.917, "dur": 2.476,
    "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": 870, "tid": 870,
    "ts": 7429770628257.538, "dur": 1.768,
    "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": 870, "tid": 870,
    "ts": 7429770628260.122, "dur": 26.936,
    "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": 870, "tid": 870,
    "ts": 7429770628270.931, "dur": 1.079,
    "args": {
      "External id": 5297,"Record function id": 0, "finished": true, "Ev Idx": 5296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770628287.855, "dur": 1.029,
    "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": 870, "tid": 870,
    "ts": 7429770628295.554, "dur": 9.060,
    "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": 870, "tid": 870,
    "ts": 7429770628315.145, "dur": 0.218,
    "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": 870, "tid": 870,
    "ts": 7429770628317.550, "dur": 12.785,
    "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": 870, "tid": 870,
    "ts": 7429770628319.249, "dur": 0.426,
    "args": {
      "External id": 5302,"Record function id": 0, "finished": true, "Ev Idx": 5301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770628320.267, "dur": 0.098,
    "args": {
      "External id": 5303,"Record function id": 0, "finished": true, "Ev Idx": 5302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770628333.208, "dur": 16.135,
    "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": 870, "tid": 870,
    "ts": 7429770628353.168, "dur": 9.900,
    "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": 870, "tid": 870,
    "ts": 7429770628365.439, "dur": 9.721,
    "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": 870, "tid": 870,
    "ts": 7429770628377.003, "dur": 10.805,
    "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": 870, "tid": 870,
    "ts": 7429770628390.405, "dur": 0.107,
    "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": 870, "tid": 870,
    "ts": 7429770628391.386, "dur": 8.588,
    "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": 870, "tid": 870,
    "ts": 7429770628415.772, "dur": 38.491,
    "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": 870, "tid": 870,
    "ts": 7429770628416.463, "dur": 4.136,
    "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": 870, "tid": 870,
    "ts": 7429770628418.525, "dur": 1.181,
    "args": {
      "External id": 5312,"Record function id": 0, "finished": true, "Ev Idx": 5311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770628419.154, "dur": 0.418,
    "args": {
      "External id": 5313,"Record function id": 0, "finished": true, "Ev Idx": 5312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770628421.001, "dur": 33.090,
    "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": 870, "tid": 870,
    "ts": 7429770628421.516, "dur": 1.969,
    "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": 870, "tid": 870,
    "ts": 7429770628422.166, "dur": 1.169,
    "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": 870, "tid": 870,
    "ts": 7429770628423.824, "dur": 28.280,
    "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": 870, "tid": 870,
    "ts": 7429770628435.318, "dur": 1.161,
    "args": {
      "External id": 5318,"Record function id": 0, "finished": true, "Ev Idx": 5317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770628452.739, "dur": 0.990,
    "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": 870, "tid": 870,
    "ts": 7429770628462.495, "dur": 11.957,
    "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": 870, "tid": 870,
    "ts": 7429770628481.916, "dur": 35.173,
    "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": 870, "tid": 870,
    "ts": 7429770628482.287, "dur": 2.389,
    "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": 870, "tid": 870,
    "ts": 7429770628483.231, "dur": 0.883,
    "args": {
      "External id": 5323,"Record function id": 0, "finished": true, "Ev Idx": 5322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770628483.649, "dur": 0.347,
    "args": {
      "External id": 5324,"Record function id": 0, "finished": true, "Ev Idx": 5323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770628485.277, "dur": 31.666,
    "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": 870, "tid": 870,
    "ts": 7429770628485.938, "dur": 1.491,
    "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": 870, "tid": 870,
    "ts": 7429770628486.255, "dur": 1.052,
    "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": 870, "tid": 870,
    "ts": 7429770628487.728, "dur": 27.273,
    "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": 870, "tid": 870,
    "ts": 7429770628499.927, "dur": 1.309,
    "args": {
      "External id": 5329,"Record function id": 0, "finished": true, "Ev Idx": 5328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770628515.598, "dur": 1.071,
    "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": 870, "tid": 870,
    "ts": 7429770628519.008, "dur": 8.413,
    "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": 870, "tid": 870,
    "ts": 7429770628533.490, "dur": 43.506,
    "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": 870, "tid": 870,
    "ts": 7429770628534.037, "dur": 2.196,
    "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": 870, "tid": 870,
    "ts": 7429770628534.917, "dur": 0.887,
    "args": {
      "External id": 5334,"Record function id": 0, "finished": true, "Ev Idx": 5333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770628535.304, "dur": 0.374,
    "args": {
      "External id": 5335,"Record function id": 0, "finished": true, "Ev Idx": 5334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770628536.533, "dur": 40.377,
    "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": 870, "tid": 870,
    "ts": 7429770628536.934, "dur": 2.355,
    "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": 870, "tid": 870,
    "ts": 7429770628537.203, "dur": 1.963,
    "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": 870, "tid": 870,
    "ts": 7429770628539.613, "dur": 35.529,
    "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": 870, "tid": 870,
    "ts": 7429770628548.055, "dur": 1.084,
    "args": {
      "External id": 5340,"Record function id": 0, "finished": true, "Ev Idx": 5339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770628575.711, "dur": 0.914,
    "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": 870, "tid": 870,
    "ts": 7429770628579.597, "dur": 8.790,
    "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": 870, "tid": 870,
    "ts": 7429770628607.675, "dur": 0.188,
    "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": 870, "tid": 870,
    "ts": 7429770628609.581, "dur": 10.217,
    "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": 870, "tid": 870,
    "ts": 7429770628610.951, "dur": 0.240,
    "args": {
      "External id": 5345,"Record function id": 0, "finished": true, "Ev Idx": 5344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770628611.485, "dur": 0.067,
    "args": {
      "External id": 5346,"Record function id": 0, "finished": true, "Ev Idx": 5345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770628622.096, "dur": 10.558,
    "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": 870, "tid": 870,
    "ts": 7429770628635.749, "dur": 8.247,
    "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": 870, "tid": 870,
    "ts": 7429770628645.436, "dur": 7.435,
    "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": 870, "tid": 870,
    "ts": 7429770628654.298, "dur": 8.619,
    "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": 870, "tid": 870,
    "ts": 7429770628665.352, "dur": 0.112,
    "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": 870, "tid": 870,
    "ts": 7429770628666.223, "dur": 8.332,
    "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": 870, "tid": 870,
    "ts": 7429770628692.476, "dur": 37.746,
    "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": 870, "tid": 870,
    "ts": 7429770628693.039, "dur": 3.215,
    "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": 870, "tid": 870,
    "ts": 7429770628694.569, "dur": 1.079,
    "args": {
      "External id": 5355,"Record function id": 0, "finished": true, "Ev Idx": 5354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770628695.042, "dur": 0.462,
    "args": {
      "External id": 5356,"Record function id": 0, "finished": true, "Ev Idx": 5355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770628696.877, "dur": 33.190,
    "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": 870, "tid": 870,
    "ts": 7429770628697.508, "dur": 1.442,
    "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": 870, "tid": 870,
    "ts": 7429770628697.789, "dur": 1.037,
    "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": 870, "tid": 870,
    "ts": 7429770628699.291, "dur": 28.893,
    "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": 870, "tid": 870,
    "ts": 7429770628710.731, "dur": 1.193,
    "args": {
      "External id": 5361,"Record function id": 0, "finished": true, "Ev Idx": 5360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770628728.769, "dur": 1.038,
    "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": 870, "tid": 870,
    "ts": 7429770628733.391, "dur": 1.370,
    "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": 870, "tid": 870,
    "ts": 7429770628736.511, "dur": 2.120,
    "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": 870, "tid": 870,
    "ts": 7429770628737.507, "dur": 0.471,
    "args": {
      "External id": 5365,"Record function id": 0, "finished": true, "Ev Idx": 5364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770628745.235, "dur": 33.824,
    "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": 870, "tid": 870,
    "ts": 7429770628745.525, "dur": 1.969,
    "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": 870, "tid": 870,
    "ts": 7429770628746.248, "dur": 0.814,
    "args": {
      "External id": 5368,"Record function id": 0, "finished": true, "Ev Idx": 5367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770628746.642, "dur": 0.304,
    "args": {
      "External id": 5369,"Record function id": 0, "finished": true, "Ev Idx": 5368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770628747.784, "dur": 31.172,
    "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": 870, "tid": 870,
    "ts": 7429770628748.400, "dur": 1.850,
    "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": 870, "tid": 870,
    "ts": 7429770628748.674, "dur": 1.484,
    "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": 870, "tid": 870,
    "ts": 7429770628750.552, "dur": 26.647,
    "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": 870, "tid": 870,
    "ts": 7429770628760.727, "dur": 1.115,
    "args": {
      "External id": 5374,"Record function id": 0, "finished": true, "Ev Idx": 5373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770628777.761, "dur": 0.947,
    "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": 870, "tid": 870,
    "ts": 7429770628781.230, "dur": 1.342,
    "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": 870, "tid": 870,
    "ts": 7429770628785.222, "dur": 1.583,
    "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": 870, "tid": 870,
    "ts": 7429770628785.975, "dur": 0.358,
    "args": {
      "External id": 5378,"Record function id": 0, "finished": true, "Ev Idx": 5377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770628793.223, "dur": 34.021,
    "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": 870, "tid": 870,
    "ts": 7429770628793.516, "dur": 2.019,
    "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": 870, "tid": 870,
    "ts": 7429770628794.436, "dur": 0.716,
    "args": {
      "External id": 5381,"Record function id": 0, "finished": true, "Ev Idx": 5380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770628794.786, "dur": 0.261,
    "args": {
      "External id": 5382,"Record function id": 0, "finished": true, "Ev Idx": 5381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770628795.827, "dur": 31.317,
    "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": 870, "tid": 870,
    "ts": 7429770628796.268, "dur": 3.309,
    "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": 870, "tid": 870,
    "ts": 7429770628796.927, "dur": 2.540,
    "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": 870, "tid": 870,
    "ts": 7429770628799.969, "dur": 25.429,
    "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": 870, "tid": 870,
    "ts": 7429770628809.553, "dur": 1.043,
    "args": {
      "External id": 5387,"Record function id": 0, "finished": true, "Ev Idx": 5386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770628826.016, "dur": 0.863,
    "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": 870, "tid": 870,
    "ts": 7429770628829.472, "dur": 1.136,
    "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": 870, "tid": 870,
    "ts": 7429770628832.007, "dur": 1.562,
    "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": 870, "tid": 870,
    "ts": 7429770628832.712, "dur": 0.330,
    "args": {
      "External id": 5391,"Record function id": 0, "finished": true, "Ev Idx": 5390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770628836.172, "dur": 1.957,
    "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": 870, "tid": 870,
    "ts": 7429770628837.279, "dur": 0.321,
    "args": {
      "External id": 5393,"Record function id": 0, "finished": true, "Ev Idx": 5392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770628839.171, "dur": 1.020,
    "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": 870, "tid": 870,
    "ts": 7429770628839.559, "dur": 0.310,
    "args": {
      "External id": 5395,"Record function id": 0, "finished": true, "Ev Idx": 5394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770628840.984, "dur": 9.623,
    "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": 870, "tid": 870,
    "ts": 7429770628855.469, "dur": 3.503,
    "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": 870, "tid": 870,
    "ts": 7429770628857.320, "dur": 0.545,
    "args": {
      "External id": 5398,"Record function id": 0, "finished": true, "Ev Idx": 5397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770628861.263, "dur": 2.942,
    "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": 870, "tid": 870,
    "ts": 7429770628862.144, "dur": 1.162,
    "args": {
      "External id": 5400,"Record function id": 0, "finished": true, "Ev Idx": 5399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770628865.935, "dur": 13.275,
    "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": 870, "tid": 870,
    "ts": 7429770628883.148, "dur": 15.809,
    "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": 870, "tid": 870,
    "ts": 7429770628902.128, "dur": 9.380,
    "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": 870, "tid": 870,
    "ts": 7429770628913.111, "dur": 10.958,
    "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": 870, "tid": 870,
    "ts": 7429770628926.171, "dur": 6.832,
    "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": 870, "tid": 870,
    "ts": 7429770628935.436, "dur": 2.845,
    "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": 870, "tid": 870,
    "ts": 7429770628937.158, "dur": 0.413,
    "args": {
      "External id": 5407,"Record function id": 0, "finished": true, "Ev Idx": 5406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770628939.812, "dur": 2.422,
    "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": 870, "tid": 870,
    "ts": 7429770628941.396, "dur": 0.294,
    "args": {
      "External id": 5409,"Record function id": 0, "finished": true, "Ev Idx": 5408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770628943.123, "dur": 8.266,
    "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": 870, "tid": 870,
    "ts": 7429770628953.087, "dur": 13.415,
    "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": 870, "tid": 870,
    "ts": 7429770628968.809, "dur": 7.601,
    "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": 870, "tid": 870,
    "ts": 7429770628977.858, "dur": 9.568,
    "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": 870, "tid": 870,
    "ts": 7429770629010.658, "dur": 3.518,
    "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": 870, "tid": 870,
    "ts": 7429770629012.868, "dur": 0.458,
    "args": {
      "External id": 5415,"Record function id": 0, "finished": true, "Ev Idx": 5414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770629014.827, "dur": 2.289,
    "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": 870, "tid": 870,
    "ts": 7429770629016.281, "dur": 0.296,
    "args": {
      "External id": 5417,"Record function id": 0, "finished": true, "Ev Idx": 5416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770629017.773, "dur": 2.362,
    "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": 870, "tid": 870,
    "ts": 7429770629018.867, "dur": 0.612,
    "args": {
      "External id": 5419,"Record function id": 0, "finished": true, "Ev Idx": 5418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770629020.747, "dur": 1.842,
    "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": 870, "tid": 870,
    "ts": 7429770629021.669, "dur": 0.367,
    "args": {
      "External id": 5421,"Record function id": 0, "finished": true, "Ev Idx": 5420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770629023.150, "dur": 1.550,
    "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": 870, "tid": 870,
    "ts": 7429770629023.889, "dur": 0.333,
    "args": {
      "External id": 5423,"Record function id": 0, "finished": true, "Ev Idx": 5422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770629027.375, "dur": 4.067,
    "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": 870, "tid": 870,
    "ts": 7429770629030.333, "dur": 0.399,
    "args": {
      "External id": 5425,"Record function id": 0, "finished": true, "Ev Idx": 5424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770629033.823, "dur": 17.864,
    "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": 870, "tid": 870,
    "ts": 7429770629036.011, "dur": 13.636,
    "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": 870, "tid": 870,
    "ts": 7429770629037.800, "dur": 2.565,
    "args": {
      "External id": 5428,"Record function id": 0, "finished": true, "Ev Idx": 5427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770629038.532, "dur": 1.613,
    "args": {
      "External id": 5429,"Record function id": 0, "finished": true, "Ev Idx": 5428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770629040.921, "dur": 8.164,
    "args": {
      "External id": 5430,"Record function id": 0, "finished": true, "Ev Idx": 5429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770629050.150, "dur": 1.235,
    "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": 870, "tid": 870,
    "ts": 7429770629054.813, "dur": 1.884,
    "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": 870, "tid": 870,
    "ts": 7429770629055.694, "dur": 0.387,
    "args": {
      "External id": 5433,"Record function id": 0, "finished": true, "Ev Idx": 5432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770629057.329, "dur": 1.473,
    "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": 870, "tid": 870,
    "ts": 7429770629058.037, "dur": 0.280,
    "args": {
      "External id": 5435,"Record function id": 0, "finished": true, "Ev Idx": 5434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770629059.298, "dur": 2.829,
    "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": 870, "tid": 870,
    "ts": 7429770629061.219, "dur": 0.463,
    "args": {
      "External id": 5437,"Record function id": 0, "finished": true, "Ev Idx": 5436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770629062.686, "dur": 2.058,
    "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": 870, "tid": 870,
    "ts": 7429770629063.953, "dur": 0.318,
    "args": {
      "External id": 5439,"Record function id": 0, "finished": true, "Ev Idx": 5438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770629065.476, "dur": 1.446,
    "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": 870, "tid": 870,
    "ts": 7429770629066.172, "dur": 0.285,
    "args": {
      "External id": 5441,"Record function id": 0, "finished": true, "Ev Idx": 5440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770629068.762, "dur": 1.725,
    "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": 870, "tid": 870,
    "ts": 7429770629069.602, "dur": 0.462,
    "args": {
      "External id": 5443,"Record function id": 0, "finished": true, "Ev Idx": 5442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770629071.931, "dur": 13.579,
    "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": 870, "tid": 870,
    "ts": 7429770629072.882, "dur": 9.855,
    "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": 870, "tid": 870,
    "ts": 7429770629073.705, "dur": 1.769,
    "args": {
      "External id": 5446,"Record function id": 0, "finished": true, "Ev Idx": 5445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770629074.125, "dur": 1.232,
    "args": {
      "External id": 5447,"Record function id": 0, "finished": true, "Ev Idx": 5446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770629075.794, "dur": 6.628,
    "args": {
      "External id": 5448,"Record function id": 0, "finished": true, "Ev Idx": 5447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770629083.135, "dur": 2.231,
    "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": 870, "tid": 870,
    "ts": 7429770629087.579, "dur": 2.274,
    "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": 870, "tid": 870,
    "ts": 7429770629088.980, "dur": 0.385,
    "args": {
      "External id": 5451,"Record function id": 0, "finished": true, "Ev Idx": 5450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770629091.196, "dur": 108.795,
    "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": 870, "tid": 870,
    "ts": 7429770629091.691, "dur": 1.736,
    "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": 870, "tid": 870,
    "ts": 7429770629092.664, "dur": 0.335,
    "args": {
      "External id": 5454,"Record function id": 0, "finished": true, "Ev Idx": 5453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770629093.773, "dur": 3.474,
    "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": 870, "tid": 870,
    "ts": 7429770629095.444, "dur": 1.664,
    "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": 870, "tid": 870,
    "ts": 7429770629097.663, "dur": 1.405,
    "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": 870, "tid": 870,
    "ts": 7429770629098.385, "dur": 0.294,
    "args": {
      "External id": 5458,"Record function id": 0, "finished": true, "Ev Idx": 5457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770629099.387, "dur": 1.737,
    "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": 870, "tid": 870,
    "ts": 7429770629100.151, "dur": 0.877,
    "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": 870, "tid": 870,
    "ts": 7429770629101.649, "dur": 95.954,
    "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": 870, "tid": 870,
    "ts": 7429770629175.917, "dur": 1.253,
    "args": {
      "External id": 5462,"Record function id": 0, "finished": true, "Ev Idx": 5461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770629198.342, "dur": 1.308,
    "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": 870, "tid": 870,
    "ts": 7429770629204.022, "dur": 10.737,
    "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": 870, "tid": 870,
    "ts": 7429770629218.415, "dur": 2.348,
    "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": 870, "tid": 870,
    "ts": 7429770629219.501, "dur": 0.449,
    "args": {
      "External id": 5466,"Record function id": 0, "finished": true, "Ev Idx": 5465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770629221.529, "dur": 1.633,
    "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": 870, "tid": 870,
    "ts": 7429770629222.337, "dur": 0.376,
    "args": {
      "External id": 5468,"Record function id": 0, "finished": true, "Ev Idx": 5467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770629223.973, "dur": 1.200,
    "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": 870, "tid": 870,
    "ts": 7429770629224.543, "dur": 0.263,
    "args": {
      "External id": 5470,"Record function id": 0, "finished": true, "Ev Idx": 5469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770629226.939, "dur": 7.973,
    "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": 870, "tid": 870,
    "ts": 7429770629240.570, "dur": 14.501,
    "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": 870, "tid": 870,
    "ts": 7429770629241.795, "dur": 0.399,
    "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": 870, "tid": 870,
    "ts": 7429770629242.831, "dur": 12.027,
    "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": 870, "tid": 870,
    "ts": 7429770629258.042, "dur": 0.177,
    "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": 870, "tid": 870,
    "ts": 7429770629268.873, "dur": 0.488,
    "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": 870, "tid": 870,
    "ts": 7429770629271.186, "dur": 44.311,
    "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": 870, "tid": 870,
    "ts": 7429770629271.906, "dur": 3.198,
    "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": 870, "tid": 870,
    "ts": 7429770629274.173, "dur": 0.392,
    "args": {
      "External id": 5479,"Record function id": 0, "finished": true, "Ev Idx": 5478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770629275.501, "dur": 3.590,
    "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": 870, "tid": 870,
    "ts": 7429770629276.066, "dur": 2.904,
    "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": 870, "tid": 870,
    "ts": 7429770629279.667, "dur": 2.408,
    "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": 870, "tid": 870,
    "ts": 7429770629281.145, "dur": 0.524,
    "args": {
      "External id": 5483,"Record function id": 0, "finished": true, "Ev Idx": 5482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770629282.389, "dur": 1.546,
    "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": 870, "tid": 870,
    "ts": 7429770629282.761, "dur": 1.102,
    "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": 870, "tid": 870,
    "ts": 7429770629284.375, "dur": 29.246,
    "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": 870, "tid": 870,
    "ts": 7429770629296.070, "dur": 1.290,
    "args": {
      "External id": 5487,"Record function id": 0, "finished": true, "Ev Idx": 5486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770629314.159, "dur": 1.070,
    "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": 870, "tid": 870,
    "ts": 7429770629317.504, "dur": 1.862,
    "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": 870, "tid": 870,
    "ts": 7429770629318.502, "dur": 0.305,
    "args": {
      "External id": 5490,"Record function id": 0, "finished": true, "Ev Idx": 5489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770629321.733, "dur": 12.242,
    "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": 870, "tid": 870,
    "ts": 7429770629322.342, "dur": 11.487,
    "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": 870, "tid": 870,
    "ts": 7429770629323.390, "dur": 2.320,
    "args": {
      "External id": 5493,"Record function id": 0, "finished": true, "Ev Idx": 5492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770629323.829, "dur": 1.692,
    "args": {
      "External id": 5494,"Record function id": 0, "finished": true, "Ev Idx": 5493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770629326.122, "dur": 7.247,
    "args": {
      "External id": 5495,"Record function id": 0, "finished": true, "Ev Idx": 5494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770629339.590, "dur": 1.495,
    "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": 870, "tid": 870,
    "ts": 7429770629339.913, "dur": 1.070,
    "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": 870, "tid": 870,
    "ts": 7429770629355.652, "dur": 37.645,
    "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": 870, "tid": 870,
    "ts": 7429770629356.239, "dur": 3.083,
    "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": 870, "tid": 870,
    "ts": 7429770629357.865, "dur": 0.940,
    "args": {
      "External id": 5500,"Record function id": 0, "finished": true, "Ev Idx": 5499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770629358.320, "dur": 0.366,
    "args": {
      "External id": 5501,"Record function id": 0, "finished": true, "Ev Idx": 5500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770629359.740, "dur": 33.378,
    "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": 870, "tid": 870,
    "ts": 7429770629360.353, "dur": 1.775,
    "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": 870, "tid": 870,
    "ts": 7429770629360.653, "dur": 1.371,
    "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": 870, "tid": 870,
    "ts": 7429770629362.568, "dur": 28.664,
    "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": 870, "tid": 870,
    "ts": 7429770629375.297, "dur": 1.335,
    "args": {
      "External id": 5506,"Record function id": 0, "finished": true, "Ev Idx": 5505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770629391.867, "dur": 0.943,
    "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": 870, "tid": 870,
    "ts": 7429770629398.624, "dur": 8.732,
    "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": 870, "tid": 870,
    "ts": 7429770629416.404, "dur": 0.207,
    "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": 870, "tid": 870,
    "ts": 7429770629418.588, "dur": 11.695,
    "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": 870, "tid": 870,
    "ts": 7429770629420.259, "dur": 0.492,
    "args": {
      "External id": 5511,"Record function id": 0, "finished": true, "Ev Idx": 5510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770629421.131, "dur": 0.244,
    "args": {
      "External id": 5512,"Record function id": 0, "finished": true, "Ev Idx": 5511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770629432.827, "dur": 14.307,
    "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": 870, "tid": 870,
    "ts": 7429770629450.418, "dur": 8.548,
    "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": 870, "tid": 870,
    "ts": 7429770629460.810, "dur": 10.753,
    "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": 870, "tid": 870,
    "ts": 7429770629473.307, "dur": 8.363,
    "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": 870, "tid": 870,
    "ts": 7429770629484.193, "dur": 0.107,
    "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": 870, "tid": 870,
    "ts": 7429770629485.244, "dur": 8.260,
    "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": 870, "tid": 870,
    "ts": 7429770629507.603, "dur": 37.379,
    "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": 870, "tid": 870,
    "ts": 7429770629508.065, "dur": 2.923,
    "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": 870, "tid": 870,
    "ts": 7429770629509.118, "dur": 1.230,
    "args": {
      "External id": 5521,"Record function id": 0, "finished": true, "Ev Idx": 5520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770629509.834, "dur": 0.401,
    "args": {
      "External id": 5522,"Record function id": 0, "finished": true, "Ev Idx": 5521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770629511.501, "dur": 33.343,
    "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": 870, "tid": 870,
    "ts": 7429770629512.005, "dur": 1.527,
    "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": 870, "tid": 870,
    "ts": 7429770629512.303, "dur": 1.104,
    "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": 870, "tid": 870,
    "ts": 7429770629513.894, "dur": 29.006,
    "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": 870, "tid": 870,
    "ts": 7429770629527.182, "dur": 1.117,
    "args": {
      "External id": 5527,"Record function id": 0, "finished": true, "Ev Idx": 5526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770629543.456, "dur": 1.128,
    "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": 870, "tid": 870,
    "ts": 7429770629552.453, "dur": 11.374,
    "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": 870, "tid": 870,
    "ts": 7429770629571.144, "dur": 85.684,
    "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": 870, "tid": 870,
    "ts": 7429770629571.429, "dur": 2.600,
    "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": 870, "tid": 870,
    "ts": 7429770629572.585, "dur": 0.967,
    "args": {
      "External id": 5532,"Record function id": 0, "finished": true, "Ev Idx": 5531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770629573.039, "dur": 0.380,
    "args": {
      "External id": 5533,"Record function id": 0, "finished": true, "Ev Idx": 5532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770629574.353, "dur": 82.306,
    "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": 870, "tid": 870,
    "ts": 7429770629574.772, "dur": 2.482,
    "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": 870, "tid": 870,
    "ts": 7429770629575.049, "dur": 2.074,
    "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": 870, "tid": 870,
    "ts": 7429770629577.589, "dur": 76.793,
    "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": 870, "tid": 870,
    "ts": 7429770629585.894, "dur": 48.980,
    "args": {
      "External id": 5538,"Record function id": 0, "finished": true, "Ev Idx": 5537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770629655.036, "dur": 1.317,
    "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": 870, "tid": 870,
    "ts": 7429770629659.299, "dur": 8.641,
    "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": 870, "tid": 870,
    "ts": 7429770629675.059, "dur": 50.574,
    "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": 870, "tid": 870,
    "ts": 7429770629675.744, "dur": 3.166,
    "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": 870, "tid": 870,
    "ts": 7429770629677.093, "dur": 1.233,
    "args": {
      "External id": 5543,"Record function id": 0, "finished": true, "Ev Idx": 5542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770629677.518, "dur": 0.632,
    "args": {
      "External id": 5544,"Record function id": 0, "finished": true, "Ev Idx": 5543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770629679.371, "dur": 46.135,
    "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": 870, "tid": 870,
    "ts": 7429770629679.959, "dur": 1.559,
    "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": 870, "tid": 870,
    "ts": 7429770629680.412, "dur": 1.000,
    "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": 870, "tid": 870,
    "ts": 7429770629681.830, "dur": 41.693,
    "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": 870, "tid": 870,
    "ts": 7429770629694.821, "dur": 1.120,
    "args": {
      "External id": 5549,"Record function id": 0, "finished": true, "Ev Idx": 5548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770629724.127, "dur": 1.082,
    "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": 870, "tid": 870,
    "ts": 7429770629728.540, "dur": 7.526,
    "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": 870, "tid": 870,
    "ts": 7429770629754.928, "dur": 0.355,
    "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": 870, "tid": 870,
    "ts": 7429770629757.380, "dur": 10.967,
    "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": 870, "tid": 870,
    "ts": 7429770629758.833, "dur": 0.423,
    "args": {
      "External id": 5554,"Record function id": 0, "finished": true, "Ev Idx": 5553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770629759.590, "dur": 0.265,
    "args": {
      "External id": 5555,"Record function id": 0, "finished": true, "Ev Idx": 5554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770629770.768, "dur": 11.225,
    "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": 870, "tid": 870,
    "ts": 7429770629785.348, "dur": 7.345,
    "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": 870, "tid": 870,
    "ts": 7429770629794.340, "dur": 9.458,
    "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": 870, "tid": 870,
    "ts": 7429770629805.263, "dur": 7.503,
    "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": 870, "tid": 870,
    "ts": 7429770629814.890, "dur": 0.265,
    "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": 870, "tid": 870,
    "ts": 7429770629815.935, "dur": 8.665,
    "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": 870, "tid": 870,
    "ts": 7429770629842.597, "dur": 37.260,
    "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": 870, "tid": 870,
    "ts": 7429770629843.027, "dur": 3.339,
    "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": 870, "tid": 870,
    "ts": 7429770629844.036, "dur": 1.266,
    "args": {
      "External id": 5564,"Record function id": 0, "finished": true, "Ev Idx": 5563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770629844.680, "dur": 0.458,
    "args": {
      "External id": 5565,"Record function id": 0, "finished": true, "Ev Idx": 5564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770629846.731, "dur": 33.004,
    "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": 870, "tid": 870,
    "ts": 7429770629847.409, "dur": 2.649,
    "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": 870, "tid": 870,
    "ts": 7429770629848.042, "dur": 1.878,
    "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": 870, "tid": 870,
    "ts": 7429770629850.420, "dur": 27.392,
    "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": 870, "tid": 870,
    "ts": 7429770629860.728, "dur": 1.218,
    "args": {
      "External id": 5570,"Record function id": 0, "finished": true, "Ev Idx": 5569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770629878.374, "dur": 1.088,
    "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": 870, "tid": 870,
    "ts": 7429770629882.898, "dur": 1.316,
    "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": 870, "tid": 870,
    "ts": 7429770629885.958, "dur": 2.328,
    "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": 870, "tid": 870,
    "ts": 7429770629887.124, "dur": 0.467,
    "args": {
      "External id": 5574,"Record function id": 0, "finished": true, "Ev Idx": 5573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770629896.093, "dur": 30.844,
    "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": 870, "tid": 870,
    "ts": 7429770629896.586, "dur": 2.196,
    "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": 870, "tid": 870,
    "ts": 7429770629897.571, "dur": 0.760,
    "args": {
      "External id": 5577,"Record function id": 0, "finished": true, "Ev Idx": 5576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770629897.967, "dur": 0.269,
    "args": {
      "External id": 5578,"Record function id": 0, "finished": true, "Ev Idx": 5577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770629899.065, "dur": 27.794,
    "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": 870, "tid": 870,
    "ts": 7429770629899.445, "dur": 1.233,
    "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": 870, "tid": 870,
    "ts": 7429770629899.725, "dur": 0.834,
    "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": 870, "tid": 870,
    "ts": 7429770629900.979, "dur": 24.132,
    "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": 870, "tid": 870,
    "ts": 7429770629909.455, "dur": 1.123,
    "args": {
      "External id": 5583,"Record function id": 0, "finished": true, "Ev Idx": 5582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770629925.612, "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": 870, "tid": 870,
    "ts": 7429770629929.150, "dur": 1.124,
    "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": 870, "tid": 870,
    "ts": 7429770629931.680, "dur": 1.648,
    "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": 870, "tid": 870,
    "ts": 7429770629932.542, "dur": 0.302,
    "args": {
      "External id": 5587,"Record function id": 0, "finished": true, "Ev Idx": 5586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770629940.515, "dur": 32.569,
    "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": 870, "tid": 870,
    "ts": 7429770629940.979, "dur": 2.103,
    "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": 870, "tid": 870,
    "ts": 7429770629941.699, "dur": 0.956,
    "args": {
      "External id": 5590,"Record function id": 0, "finished": true, "Ev Idx": 5589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770629942.087, "dur": 0.444,
    "args": {
      "External id": 5591,"Record function id": 0, "finished": true, "Ev Idx": 5590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770629943.350, "dur": 29.655,
    "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": 870, "tid": 870,
    "ts": 7429770629943.877, "dur": 1.301,
    "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": 870, "tid": 870,
    "ts": 7429770629944.144, "dur": 0.932,
    "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": 870, "tid": 870,
    "ts": 7429770629945.627, "dur": 25.694,
    "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": 870, "tid": 870,
    "ts": 7429770629956.720, "dur": 1.182,
    "args": {
      "External id": 5596,"Record function id": 0, "finished": true, "Ev Idx": 5595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770629971.830, "dur": 0.925,
    "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": 870, "tid": 870,
    "ts": 7429770629975.313, "dur": 1.036,
    "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": 870, "tid": 870,
    "ts": 7429770629977.821, "dur": 1.670,
    "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": 870, "tid": 870,
    "ts": 7429770629978.656, "dur": 0.385,
    "args": {
      "External id": 5600,"Record function id": 0, "finished": true, "Ev Idx": 5599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770629982.336, "dur": 2.341,
    "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": 870, "tid": 870,
    "ts": 7429770629983.811, "dur": 0.332,
    "args": {
      "External id": 5602,"Record function id": 0, "finished": true, "Ev Idx": 5601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770629985.738, "dur": 1.237,
    "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": 870, "tid": 870,
    "ts": 7429770629986.297, "dur": 0.276,
    "args": {
      "External id": 5604,"Record function id": 0, "finished": true, "Ev Idx": 5603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770629987.779, "dur": 9.325,
    "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": 870, "tid": 870,
    "ts": 7429770630002.083, "dur": 4.726,
    "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": 870, "tid": 870,
    "ts": 7429770630004.489, "dur": 0.888,
    "args": {
      "External id": 5607,"Record function id": 0, "finished": true, "Ev Idx": 5606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770630008.919, "dur": 2.139,
    "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": 870, "tid": 870,
    "ts": 7429770630010.084, "dur": 0.325,
    "args": {
      "External id": 5609,"Record function id": 0, "finished": true, "Ev Idx": 5608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770630012.826, "dur": 11.471,
    "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": 870, "tid": 870,
    "ts": 7429770630028.379, "dur": 17.885,
    "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": 870, "tid": 870,
    "ts": 7429770630049.257, "dur": 8.876,
    "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": 870, "tid": 870,
    "ts": 7429770630059.760, "dur": 8.840,
    "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": 870, "tid": 870,
    "ts": 7429770630090.465, "dur": 12.331,
    "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": 870, "tid": 870,
    "ts": 7429770630106.287, "dur": 3.044,
    "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": 870, "tid": 870,
    "ts": 7429770630107.700, "dur": 0.581,
    "args": {
      "External id": 5616,"Record function id": 0, "finished": true, "Ev Idx": 5615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770630111.078, "dur": 1.724,
    "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": 870, "tid": 870,
    "ts": 7429770630111.920, "dur": 0.326,
    "args": {
      "External id": 5618,"Record function id": 0, "finished": true, "Ev Idx": 5617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770630114.072, "dur": 8.968,
    "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": 870, "tid": 870,
    "ts": 7429770630125.342, "dur": 10.823,
    "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": 870, "tid": 870,
    "ts": 7429770630138.572, "dur": 8.108,
    "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": 870, "tid": 870,
    "ts": 7429770630148.164, "dur": 8.308,
    "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": 870, "tid": 870,
    "ts": 7429770630181.558, "dur": 2.574,
    "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": 870, "tid": 870,
    "ts": 7429770630182.910, "dur": 0.454,
    "args": {
      "External id": 5624,"Record function id": 0, "finished": true, "Ev Idx": 5623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770630184.814, "dur": 2.726,
    "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": 870, "tid": 870,
    "ts": 7429770630186.406, "dur": 0.589,
    "args": {
      "External id": 5626,"Record function id": 0, "finished": true, "Ev Idx": 5625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770630188.279, "dur": 3.563,
    "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": 870, "tid": 870,
    "ts": 7429770630190.351, "dur": 0.740,
    "args": {
      "External id": 5628,"Record function id": 0, "finished": true, "Ev Idx": 5627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770630192.843, "dur": 1.804,
    "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": 870, "tid": 870,
    "ts": 7429770630193.748, "dur": 0.345,
    "args": {
      "External id": 5630,"Record function id": 0, "finished": true, "Ev Idx": 5629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770630195.399, "dur": 1.699,
    "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": 870, "tid": 870,
    "ts": 7429770630196.331, "dur": 0.309,
    "args": {
      "External id": 5632,"Record function id": 0, "finished": true, "Ev Idx": 5631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770630224.348, "dur": 3.263,
    "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": 870, "tid": 870,
    "ts": 7429770630226.425, "dur": 0.410,
    "args": {
      "External id": 5634,"Record function id": 0, "finished": true, "Ev Idx": 5633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770630229.913, "dur": 21.844,
    "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": 870, "tid": 870,
    "ts": 7429770630232.649, "dur": 16.884,
    "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": 870, "tid": 870,
    "ts": 7429770630234.513, "dur": 3.279,
    "args": {
      "External id": 5637,"Record function id": 0, "finished": true, "Ev Idx": 5636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770630235.451, "dur": 1.976,
    "args": {
      "External id": 5638,"Record function id": 0, "finished": true, "Ev Idx": 5637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770630238.472, "dur": 10.472,
    "args": {
      "External id": 5639,"Record function id": 0, "finished": true, "Ev Idx": 5638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770630250.103, "dur": 1.397,
    "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": 870, "tid": 870,
    "ts": 7429770630255.030, "dur": 1.763,
    "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": 870, "tid": 870,
    "ts": 7429770630255.847, "dur": 0.341,
    "args": {
      "External id": 5642,"Record function id": 0, "finished": true, "Ev Idx": 5641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770630257.430, "dur": 1.501,
    "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": 870, "tid": 870,
    "ts": 7429770630258.184, "dur": 0.261,
    "args": {
      "External id": 5644,"Record function id": 0, "finished": true, "Ev Idx": 5643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770630259.487, "dur": 1.475,
    "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": 870, "tid": 870,
    "ts": 7429770630260.220, "dur": 0.302,
    "args": {
      "External id": 5646,"Record function id": 0, "finished": true, "Ev Idx": 5645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770630261.529, "dur": 1.545,
    "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": 870, "tid": 870,
    "ts": 7429770630262.252, "dur": 0.309,
    "args": {
      "External id": 5648,"Record function id": 0, "finished": true, "Ev Idx": 5647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770630263.668, "dur": 2.312,
    "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": 870, "tid": 870,
    "ts": 7429770630265.234, "dur": 0.287,
    "args": {
      "External id": 5650,"Record function id": 0, "finished": true, "Ev Idx": 5649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770630267.654, "dur": 1.589,
    "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": 870, "tid": 870,
    "ts": 7429770630268.468, "dur": 0.324,
    "args": {
      "External id": 5652,"Record function id": 0, "finished": true, "Ev Idx": 5651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770630270.684, "dur": 10.598,
    "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": 870, "tid": 870,
    "ts": 7429770630271.582, "dur": 8.241,
    "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": 870, "tid": 870,
    "ts": 7429770630272.201, "dur": 1.724,
    "args": {
      "External id": 5655,"Record function id": 0, "finished": true, "Ev Idx": 5654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770630272.576, "dur": 1.227,
    "args": {
      "External id": 5656,"Record function id": 0, "finished": true, "Ev Idx": 5655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770630274.276, "dur": 5.225,
    "args": {
      "External id": 5657,"Record function id": 0, "finished": true, "Ev Idx": 5656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770630280.220, "dur": 0.892,
    "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": 870, "tid": 870,
    "ts": 7429770630283.623, "dur": 2.139,
    "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": 870, "tid": 870,
    "ts": 7429770630284.848, "dur": 0.341,
    "args": {
      "External id": 5660,"Record function id": 0, "finished": true, "Ev Idx": 5659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770630287.585, "dur": 116.572,
    "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": 870, "tid": 870,
    "ts": 7429770630288.347, "dur": 2.171,
    "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": 870, "tid": 870,
    "ts": 7429770630289.722, "dur": 0.349,
    "args": {
      "External id": 5663,"Record function id": 0, "finished": true, "Ev Idx": 5662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770630290.907, "dur": 4.347,
    "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": 870, "tid": 870,
    "ts": 7429770630292.896, "dur": 2.225,
    "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": 870, "tid": 870,
    "ts": 7429770630295.747, "dur": 1.371,
    "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": 870, "tid": 870,
    "ts": 7429770630296.439, "dur": 0.302,
    "args": {
      "External id": 5667,"Record function id": 0, "finished": true, "Ev Idx": 5666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770630297.420, "dur": 1.790,
    "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": 870, "tid": 870,
    "ts": 7429770630298.201, "dur": 0.897,
    "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": 870, "tid": 870,
    "ts": 7429770630299.915, "dur": 101.575,
    "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": 870, "tid": 870,
    "ts": 7429770630373.398, "dur": 1.489,
    "args": {
      "External id": 5671,"Record function id": 0, "finished": true, "Ev Idx": 5670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770630402.273, "dur": 1.456,
    "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": 870, "tid": 870,
    "ts": 7429770630409.358, "dur": 12.454,
    "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": 870, "tid": 870,
    "ts": 7429770630425.485, "dur": 2.253,
    "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": 870, "tid": 870,
    "ts": 7429770630426.295, "dur": 0.527,
    "args": {
      "External id": 5675,"Record function id": 0, "finished": true, "Ev Idx": 5674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770630428.449, "dur": 1.351,
    "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": 870, "tid": 870,
    "ts": 7429770630429.004, "dur": 0.320,
    "args": {
      "External id": 5677,"Record function id": 0, "finished": true, "Ev Idx": 5676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770630430.478, "dur": 0.991,
    "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": 870, "tid": 870,
    "ts": 7429770630430.919, "dur": 0.217,
    "args": {
      "External id": 5679,"Record function id": 0, "finished": true, "Ev Idx": 5678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770630433.334, "dur": 10.921,
    "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": 870, "tid": 870,
    "ts": 7429770630450.902, "dur": 15.810,
    "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": 870, "tid": 870,
    "ts": 7429770630452.029, "dur": 0.395,
    "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": 870, "tid": 870,
    "ts": 7429770630453.270, "dur": 13.223,
    "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": 870, "tid": 870,
    "ts": 7429770630470.202, "dur": 0.239,
    "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": 870, "tid": 870,
    "ts": 7429770630483.412, "dur": 0.489,
    "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": 870, "tid": 870,
    "ts": 7429770630485.700, "dur": 43.610,
    "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": 870, "tid": 870,
    "ts": 7429770630486.416, "dur": 2.067,
    "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": 870, "tid": 870,
    "ts": 7429770630487.556, "dur": 0.342,
    "args": {
      "External id": 5688,"Record function id": 0, "finished": true, "Ev Idx": 5687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770630488.989, "dur": 2.961,
    "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": 870, "tid": 870,
    "ts": 7429770630489.479, "dur": 2.346,
    "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": 870, "tid": 870,
    "ts": 7429770630492.453, "dur": 1.771,
    "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": 870, "tid": 870,
    "ts": 7429770630493.580, "dur": 0.264,
    "args": {
      "External id": 5692,"Record function id": 0, "finished": true, "Ev Idx": 5691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770630494.537, "dur": 2.289,
    "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": 870, "tid": 870,
    "ts": 7429770630494.787, "dur": 1.963,
    "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": 870, "tid": 870,
    "ts": 7429770630497.376, "dur": 29.885,
    "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": 870, "tid": 870,
    "ts": 7429770630508.245, "dur": 1.313,
    "args": {
      "External id": 5696,"Record function id": 0, "finished": true, "Ev Idx": 5695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770630527.874, "dur": 1.140,
    "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": 870, "tid": 870,
    "ts": 7429770630531.388, "dur": 2.127,
    "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": 870, "tid": 870,
    "ts": 7429770630532.546, "dur": 0.315,
    "args": {
      "External id": 5699,"Record function id": 0, "finished": true, "Ev Idx": 5698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770630536.105, "dur": 12.361,
    "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": 870, "tid": 870,
    "ts": 7429770630536.647, "dur": 11.667,
    "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": 870, "tid": 870,
    "ts": 7429770630537.685, "dur": 2.011,
    "args": {
      "External id": 5702,"Record function id": 0, "finished": true, "Ev Idx": 5701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770630538.122, "dur": 1.405,
    "args": {
      "External id": 5703,"Record function id": 0, "finished": true, "Ev Idx": 5702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770630540.186, "dur": 7.669,
    "args": {
      "External id": 5704,"Record function id": 0, "finished": true, "Ev Idx": 5703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770630554.881, "dur": 1.612,
    "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": 870, "tid": 870,
    "ts": 7429770630555.234, "dur": 1.158,
    "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": 870, "tid": 870,
    "ts": 7429770630575.940, "dur": 39.247,
    "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": 870, "tid": 870,
    "ts": 7429770630576.715, "dur": 3.627,
    "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": 870, "tid": 870,
    "ts": 7429770630578.771, "dur": 0.949,
    "args": {
      "External id": 5709,"Record function id": 0, "finished": true, "Ev Idx": 5708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770630579.208, "dur": 0.391,
    "args": {
      "External id": 5710,"Record function id": 0, "finished": true, "Ev Idx": 5709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770630580.801, "dur": 34.137,
    "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": 870, "tid": 870,
    "ts": 7429770630581.394, "dur": 1.577,
    "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": 870, "tid": 870,
    "ts": 7429770630581.663, "dur": 1.209,
    "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": 870, "tid": 870,
    "ts": 7429770630583.495, "dur": 29.213,
    "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": 870, "tid": 870,
    "ts": 7429770630595.348, "dur": 1.348,
    "args": {
      "External id": 5715,"Record function id": 0, "finished": true, "Ev Idx": 5714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770630613.388, "dur": 1.155,
    "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": 870, "tid": 870,
    "ts": 7429770630622.061, "dur": 10.763,
    "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": 870, "tid": 870,
    "ts": 7429770630643.669, "dur": 0.221,
    "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": 870, "tid": 870,
    "ts": 7429770630646.665, "dur": 14.788,
    "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": 870, "tid": 870,
    "ts": 7429770630648.601, "dur": 0.415,
    "args": {
      "External id": 5720,"Record function id": 0, "finished": true, "Ev Idx": 5719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770630649.438, "dur": 0.078,
    "args": {
      "External id": 5721,"Record function id": 0, "finished": true, "Ev Idx": 5720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770630664.494, "dur": 15.820,
    "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": 870, "tid": 870,
    "ts": 7429770630684.358, "dur": 9.011,
    "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": 870, "tid": 870,
    "ts": 7429770630695.591, "dur": 10.035,
    "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": 870, "tid": 870,
    "ts": 7429770630707.492, "dur": 9.174,
    "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": 870, "tid": 870,
    "ts": 7429770630719.429, "dur": 0.122,
    "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": 870, "tid": 870,
    "ts": 7429770630720.542, "dur": 9.684,
    "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": 870, "tid": 870,
    "ts": 7429770630746.269, "dur": 37.267,
    "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": 870, "tid": 870,
    "ts": 7429770630746.808, "dur": 3.070,
    "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": 870, "tid": 870,
    "ts": 7429770630747.920, "dur": 1.349,
    "args": {
      "External id": 5730,"Record function id": 0, "finished": true, "Ev Idx": 5729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770630748.602, "dur": 0.530,
    "args": {
      "External id": 5731,"Record function id": 0, "finished": true, "Ev Idx": 5730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770630750.343, "dur": 33.038,
    "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": 870, "tid": 870,
    "ts": 7429770630750.821, "dur": 2.087,
    "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": 870, "tid": 870,
    "ts": 7429770630751.147, "dur": 1.658,
    "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": 870, "tid": 870,
    "ts": 7429770630753.266, "dur": 28.053,
    "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": 870, "tid": 870,
    "ts": 7429770630763.568, "dur": 1.199,
    "args": {
      "External id": 5736,"Record function id": 0, "finished": true, "Ev Idx": 5735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770630781.981, "dur": 1.097,
    "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": 870, "tid": 870,
    "ts": 7429770630792.050, "dur": 11.504,
    "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": 870, "tid": 870,
    "ts": 7429770630810.748, "dur": 33.122,
    "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": 870, "tid": 870,
    "ts": 7429770630811.174, "dur": 2.470,
    "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": 870, "tid": 870,
    "ts": 7429770630812.238, "dur": 0.874,
    "args": {
      "External id": 5741,"Record function id": 0, "finished": true, "Ev Idx": 5740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770630812.678, "dur": 0.319,
    "args": {
      "External id": 5742,"Record function id": 0, "finished": true, "Ev Idx": 5741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770630813.997, "dur": 29.789,
    "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": 870, "tid": 870,
    "ts": 7429770630814.378, "dur": 2.118,
    "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": 870, "tid": 870,
    "ts": 7429770630814.695, "dur": 1.672,
    "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": 870, "tid": 870,
    "ts": 7429770630816.832, "dur": 25.156,
    "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": 870, "tid": 870,
    "ts": 7429770630826.427, "dur": 1.222,
    "args": {
      "External id": 5747,"Record function id": 0, "finished": true, "Ev Idx": 5746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770630842.570, "dur": 0.950,
    "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": 870, "tid": 870,
    "ts": 7429770630845.504, "dur": 8.171,
    "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": 870, "tid": 870,
    "ts": 7429770630859.123, "dur": 43.239,
    "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": 870, "tid": 870,
    "ts": 7429770630859.477, "dur": 2.231,
    "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": 870, "tid": 870,
    "ts": 7429770630860.318, "dur": 0.924,
    "args": {
      "External id": 5752,"Record function id": 0, "finished": true, "Ev Idx": 5751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770630860.836, "dur": 0.290,
    "args": {
      "External id": 5753,"Record function id": 0, "finished": true, "Ev Idx": 5752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770630862.004, "dur": 40.241,
    "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": 870, "tid": 870,
    "ts": 7429770630862.518, "dur": 1.279,
    "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": 870, "tid": 870,
    "ts": 7429770630862.783, "dur": 0.913,
    "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": 870, "tid": 870,
    "ts": 7429770630864.083, "dur": 36.351,
    "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": 870, "tid": 870,
    "ts": 7429770630873.357, "dur": 1.082,
    "args": {
      "External id": 5758,"Record function id": 0, "finished": true, "Ev Idx": 5757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770630901.017, "dur": 0.981,
    "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": 870, "tid": 870,
    "ts": 7429770630905.167, "dur": 8.398,
    "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": 870, "tid": 870,
    "ts": 7429770630932.975, "dur": 0.189,
    "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": 870, "tid": 870,
    "ts": 7429770630934.992, "dur": 10.657,
    "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": 870, "tid": 870,
    "ts": 7429770630936.391, "dur": 0.225,
    "args": {
      "External id": 5763,"Record function id": 0, "finished": true, "Ev Idx": 5762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770630936.983, "dur": 0.066,
    "args": {
      "External id": 5764,"Record function id": 0, "finished": true, "Ev Idx": 5763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770630948.069, "dur": 11.005,
    "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": 870, "tid": 870,
    "ts": 7429770630962.400, "dur": 7.594,
    "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": 870, "tid": 870,
    "ts": 7429770630971.485, "dur": 8.609,
    "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": 870, "tid": 870,
    "ts": 7429770630981.644, "dur": 7.655,
    "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": 870, "tid": 870,
    "ts": 7429770630991.956, "dur": 0.104,
    "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": 870, "tid": 870,
    "ts": 7429770630992.800, "dur": 7.471,
    "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": 870, "tid": 870,
    "ts": 7429770631017.768, "dur": 37.106,
    "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": 870, "tid": 870,
    "ts": 7429770631018.160, "dur": 3.257,
    "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": 870, "tid": 870,
    "ts": 7429770631019.213, "dur": 1.206,
    "args": {
      "External id": 5773,"Record function id": 0, "finished": true, "Ev Idx": 5772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770631019.835, "dur": 0.404,
    "args": {
      "External id": 5774,"Record function id": 0, "finished": true, "Ev Idx": 5773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770631021.747, "dur": 33.007,
    "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": 870, "tid": 870,
    "ts": 7429770631022.270, "dur": 2.072,
    "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": 870, "tid": 870,
    "ts": 7429770631022.697, "dur": 1.527,
    "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": 870, "tid": 870,
    "ts": 7429770631024.699, "dur": 28.019,
    "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": 870, "tid": 870,
    "ts": 7429770631034.284, "dur": 1.166,
    "args": {
      "External id": 5779,"Record function id": 0, "finished": true, "Ev Idx": 5778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770631053.366, "dur": 1.101,
    "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": 870, "tid": 870,
    "ts": 7429770631058.157, "dur": 1.385,
    "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": 870, "tid": 870,
    "ts": 7429770631061.234, "dur": 2.156,
    "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": 870, "tid": 870,
    "ts": 7429770631062.296, "dur": 0.422,
    "args": {
      "External id": 5783,"Record function id": 0, "finished": true, "Ev Idx": 5782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770631069.924, "dur": 30.364,
    "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": 870, "tid": 870,
    "ts": 7429770631070.241, "dur": 2.067,
    "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": 870, "tid": 870,
    "ts": 7429770631071.084, "dur": 0.775,
    "args": {
      "External id": 5786,"Record function id": 0, "finished": true, "Ev Idx": 5785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770631071.466, "dur": 0.275,
    "args": {
      "External id": 5787,"Record function id": 0, "finished": true, "Ev Idx": 5786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770631072.590, "dur": 27.602,
    "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": 870, "tid": 870,
    "ts": 7429770631072.978, "dur": 1.404,
    "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": 870, "tid": 870,
    "ts": 7429770631073.402, "dur": 0.875,
    "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": 870, "tid": 870,
    "ts": 7429770631074.696, "dur": 23.825,
    "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": 870, "tid": 870,
    "ts": 7429770631083.117, "dur": 1.052,
    "args": {
      "External id": 5792,"Record function id": 0, "finished": true, "Ev Idx": 5791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770631099.086, "dur": 0.884,
    "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": 870, "tid": 870,
    "ts": 7429770631102.739, "dur": 1.154,
    "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": 870, "tid": 870,
    "ts": 7429770631105.261, "dur": 1.557,
    "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": 870, "tid": 870,
    "ts": 7429770631105.978, "dur": 0.366,
    "args": {
      "External id": 5796,"Record function id": 0, "finished": true, "Ev Idx": 5795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770631113.465, "dur": 32.303,
    "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": 870, "tid": 870,
    "ts": 7429770631113.783, "dur": 1.839,
    "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": 870, "tid": 870,
    "ts": 7429770631114.476, "dur": 0.733,
    "args": {
      "External id": 5799,"Record function id": 0, "finished": true, "Ev Idx": 5798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770631114.827, "dur": 0.285,
    "args": {
      "External id": 5800,"Record function id": 0, "finished": true, "Ev Idx": 5799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770631115.927, "dur": 29.762,
    "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": 870, "tid": 870,
    "ts": 7429770631116.277, "dur": 1.442,
    "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": 870, "tid": 870,
    "ts": 7429770631116.812, "dur": 0.788,
    "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": 870, "tid": 870,
    "ts": 7429770631118.027, "dur": 26.045,
    "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": 870, "tid": 870,
    "ts": 7429770631129.409, "dur": 1.074,
    "args": {
      "External id": 5805,"Record function id": 0, "finished": true, "Ev Idx": 5804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770631144.587, "dur": 0.864,
    "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": 870, "tid": 870,
    "ts": 7429770631148.075, "dur": 1.082,
    "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": 870, "tid": 870,
    "ts": 7429770631150.499, "dur": 1.552,
    "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": 870, "tid": 870,
    "ts": 7429770631151.219, "dur": 0.396,
    "args": {
      "External id": 5809,"Record function id": 0, "finished": true, "Ev Idx": 5808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770631154.716, "dur": 1.971,
    "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": 870, "tid": 870,
    "ts": 7429770631155.804, "dur": 0.352,
    "args": {
      "External id": 5811,"Record function id": 0, "finished": true, "Ev Idx": 5810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770631157.747, "dur": 0.987,
    "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": 870, "tid": 870,
    "ts": 7429770631158.151, "dur": 0.275,
    "args": {
      "External id": 5813,"Record function id": 0, "finished": true, "Ev Idx": 5812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770631159.547, "dur": 10.767,
    "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": 870, "tid": 870,
    "ts": 7429770631174.780, "dur": 3.444,
    "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": 870, "tid": 870,
    "ts": 7429770631176.630, "dur": 0.497,
    "args": {
      "External id": 5816,"Record function id": 0, "finished": true, "Ev Idx": 5815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770631180.292, "dur": 1.804,
    "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": 870, "tid": 870,
    "ts": 7429770631181.118, "dur": 0.347,
    "args": {
      "External id": 5818,"Record function id": 0, "finished": true, "Ev Idx": 5817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770631183.886, "dur": 11.415,
    "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": 870, "tid": 870,
    "ts": 7429770631199.056, "dur": 17.011,
    "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": 870, "tid": 870,
    "ts": 7429770631218.745, "dur": 9.184,
    "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": 870, "tid": 870,
    "ts": 7429770631229.446, "dur": 9.391,
    "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": 870, "tid": 870,
    "ts": 7429770631241.147, "dur": 7.089,
    "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": 870, "tid": 870,
    "ts": 7429770631250.484, "dur": 2.348,
    "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": 870, "tid": 870,
    "ts": 7429770631251.791, "dur": 0.409,
    "args": {
      "External id": 5825,"Record function id": 0, "finished": true, "Ev Idx": 5824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770631254.393, "dur": 2.117,
    "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": 870, "tid": 870,
    "ts": 7429770631255.607, "dur": 0.388,
    "args": {
      "External id": 5827,"Record function id": 0, "finished": true, "Ev Idx": 5826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770631257.383, "dur": 7.074,
    "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": 870, "tid": 870,
    "ts": 7429770631266.083, "dur": 9.563,
    "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": 870, "tid": 870,
    "ts": 7429770631277.797, "dur": 6.892,
    "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": 870, "tid": 870,
    "ts": 7429770631285.995, "dur": 7.950,
    "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": 870, "tid": 870,
    "ts": 7429770631316.308, "dur": 2.439,
    "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": 870, "tid": 870,
    "ts": 7429770631317.486, "dur": 0.450,
    "args": {
      "External id": 5833,"Record function id": 0, "finished": true, "Ev Idx": 5832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770631319.450, "dur": 2.282,
    "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": 870, "tid": 870,
    "ts": 7429770631320.919, "dur": 0.311,
    "args": {
      "External id": 5835,"Record function id": 0, "finished": true, "Ev Idx": 5834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770631322.510, "dur": 2.331,
    "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": 870, "tid": 870,
    "ts": 7429770631323.644, "dur": 0.504,
    "args": {
      "External id": 5837,"Record function id": 0, "finished": true, "Ev Idx": 5836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770631325.497, "dur": 1.914,
    "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": 870, "tid": 870,
    "ts": 7429770631326.456, "dur": 0.468,
    "args": {
      "External id": 5839,"Record function id": 0, "finished": true, "Ev Idx": 5838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770631328.043, "dur": 2.529,
    "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": 870, "tid": 870,
    "ts": 7429770631329.674, "dur": 0.441,
    "args": {
      "External id": 5841,"Record function id": 0, "finished": true, "Ev Idx": 5840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770631333.096, "dur": 3.210,
    "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": 870, "tid": 870,
    "ts": 7429770631335.252, "dur": 0.414,
    "args": {
      "External id": 5843,"Record function id": 0, "finished": true, "Ev Idx": 5842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770631338.618, "dur": 17.708,
    "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": 870, "tid": 870,
    "ts": 7429770631340.988, "dur": 13.502,
    "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": 870, "tid": 870,
    "ts": 7429770631342.168, "dur": 2.270,
    "args": {
      "External id": 5846,"Record function id": 0, "finished": true, "Ev Idx": 5845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770631342.819, "dur": 1.405,
    "args": {
      "External id": 5847,"Record function id": 0, "finished": true, "Ev Idx": 5846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770631344.940, "dur": 9.048,
    "args": {
      "External id": 5848,"Record function id": 0, "finished": true, "Ev Idx": 5847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770631354.940, "dur": 1.119,
    "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": 870, "tid": 870,
    "ts": 7429770631359.507, "dur": 1.691,
    "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": 870, "tid": 870,
    "ts": 7429770631360.286, "dur": 0.350,
    "args": {
      "External id": 5851,"Record function id": 0, "finished": true, "Ev Idx": 5850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770631361.840, "dur": 2.804,
    "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": 870, "tid": 870,
    "ts": 7429770631363.823, "dur": 0.352,
    "args": {
      "External id": 5853,"Record function id": 0, "finished": true, "Ev Idx": 5852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770631365.166, "dur": 1.809,
    "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": 870, "tid": 870,
    "ts": 7429770631366.238, "dur": 0.318,
    "args": {
      "External id": 5855,"Record function id": 0, "finished": true, "Ev Idx": 5854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770631367.531, "dur": 1.797,
    "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": 870, "tid": 870,
    "ts": 7429770631368.526, "dur": 0.296,
    "args": {
      "External id": 5857,"Record function id": 0, "finished": true, "Ev Idx": 5856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770631370.028, "dur": 1.552,
    "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": 870, "tid": 870,
    "ts": 7429770631370.841, "dur": 0.283,
    "args": {
      "External id": 5859,"Record function id": 0, "finished": true, "Ev Idx": 5858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770631373.209, "dur": 1.765,
    "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": 870, "tid": 870,
    "ts": 7429770631374.185, "dur": 0.335,
    "args": {
      "External id": 5861,"Record function id": 0, "finished": true, "Ev Idx": 5860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770631376.484, "dur": 12.810,
    "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": 870, "tid": 870,
    "ts": 7429770631377.339, "dur": 10.135,
    "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": 870, "tid": 870,
    "ts": 7429770631378.055, "dur": 1.654,
    "args": {
      "External id": 5864,"Record function id": 0, "finished": true, "Ev Idx": 5863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770631378.417, "dur": 1.164,
    "args": {
      "External id": 5865,"Record function id": 0, "finished": true, "Ev Idx": 5864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770631380.002, "dur": 7.169,
    "args": {
      "External id": 5866,"Record function id": 0, "finished": true, "Ev Idx": 5865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770631387.856, "dur": 1.267,
    "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": 870, "tid": 870,
    "ts": 7429770631391.278, "dur": 1.706,
    "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": 870, "tid": 870,
    "ts": 7429770631392.150, "dur": 0.334,
    "args": {
      "External id": 5869,"Record function id": 0, "finished": true, "Ev Idx": 5868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770631394.497, "dur": 106.148,
    "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": 870, "tid": 870,
    "ts": 7429770631394.970, "dur": 1.513,
    "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": 870, "tid": 870,
    "ts": 7429770631395.705, "dur": 0.320,
    "args": {
      "External id": 5872,"Record function id": 0, "finished": true, "Ev Idx": 5871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770631396.838, "dur": 3.459,
    "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": 870, "tid": 870,
    "ts": 7429770631398.489, "dur": 1.685,
    "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": 870, "tid": 870,
    "ts": 7429770631400.781, "dur": 1.313,
    "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": 870, "tid": 870,
    "ts": 7429770631401.428, "dur": 0.293,
    "args": {
      "External id": 5876,"Record function id": 0, "finished": true, "Ev Idx": 5875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770631402.398, "dur": 2.573,
    "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": 870, "tid": 870,
    "ts": 7429770631403.298, "dur": 1.562,
    "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": 870, "tid": 870,
    "ts": 7429770631405.429, "dur": 93.032,
    "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": 870, "tid": 870,
    "ts": 7429770631476.640, "dur": 1.265,
    "args": {
      "External id": 5880,"Record function id": 0, "finished": true, "Ev Idx": 5879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770631499.048, "dur": 1.255,
    "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": 870, "tid": 870,
    "ts": 7429770631504.918, "dur": 11.296,
    "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": 870, "tid": 870,
    "ts": 7429770631519.691, "dur": 2.396,
    "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": 870, "tid": 870,
    "ts": 7429770631520.738, "dur": 0.527,
    "args": {
      "External id": 5884,"Record function id": 0, "finished": true, "Ev Idx": 5883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770631522.765, "dur": 1.709,
    "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": 870, "tid": 870,
    "ts": 7429770631523.583, "dur": 0.481,
    "args": {
      "External id": 5886,"Record function id": 0, "finished": true, "Ev Idx": 5885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770631525.085, "dur": 1.097,
    "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": 870, "tid": 870,
    "ts": 7429770631525.629, "dur": 0.226,
    "args": {
      "External id": 5888,"Record function id": 0, "finished": true, "Ev Idx": 5887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770631527.952, "dur": 7.740,
    "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": 870, "tid": 870,
    "ts": 7429770631540.940, "dur": 14.723,
    "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": 870, "tid": 870,
    "ts": 7429770631541.953, "dur": 0.370,
    "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": 870, "tid": 870,
    "ts": 7429770631543.071, "dur": 12.363,
    "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": 870, "tid": 870,
    "ts": 7429770631558.540, "dur": 0.205,
    "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": 870, "tid": 870,
    "ts": 7429770631569.746, "dur": 0.521,
    "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": 870, "tid": 870,
    "ts": 7429770631572.139, "dur": 39.989,
    "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": 870, "tid": 870,
    "ts": 7429770631572.710, "dur": 1.964,
    "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": 870, "tid": 870,
    "ts": 7429770631573.785, "dur": 0.346,
    "args": {
      "External id": 5897,"Record function id": 0, "finished": true, "Ev Idx": 5896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770631575.111, "dur": 2.727,
    "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": 870, "tid": 870,
    "ts": 7429770631575.502, "dur": 2.209,
    "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": 870, "tid": 870,
    "ts": 7429770631578.280, "dur": 1.335,
    "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": 870, "tid": 870,
    "ts": 7429770631578.944, "dur": 0.291,
    "args": {
      "External id": 5901,"Record function id": 0, "finished": true, "Ev Idx": 5900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770631579.932, "dur": 1.458,
    "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": 870, "tid": 870,
    "ts": 7429770631580.305, "dur": 1.002,
    "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": 870, "tid": 870,
    "ts": 7429770631581.808, "dur": 28.374,
    "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": 870, "tid": 870,
    "ts": 7429770631592.293, "dur": 1.232,
    "args": {
      "External id": 5905,"Record function id": 0, "finished": true, "Ev Idx": 5904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770631610.752, "dur": 1.094,
    "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": 870, "tid": 870,
    "ts": 7429770631614.102, "dur": 2.003,
    "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": 870, "tid": 870,
    "ts": 7429770631615.102, "dur": 0.356,
    "args": {
      "External id": 5908,"Record function id": 0, "finished": true, "Ev Idx": 5907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770631618.440, "dur": 12.027,
    "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": 870, "tid": 870,
    "ts": 7429770631619.010, "dur": 11.319,
    "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": 870, "tid": 870,
    "ts": 7429770631620.022, "dur": 2.249,
    "args": {
      "External id": 5911,"Record function id": 0, "finished": true, "Ev Idx": 5910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770631620.546, "dur": 1.557,
    "args": {
      "External id": 5912,"Record function id": 0, "finished": true, "Ev Idx": 5911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770631622.632, "dur": 7.265,
    "args": {
      "External id": 5913,"Record function id": 0, "finished": true, "Ev Idx": 5912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770631636.179, "dur": 1.534,
    "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": 870, "tid": 870,
    "ts": 7429770631636.479, "dur": 1.133,
    "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": 870, "tid": 870,
    "ts": 7429770631651.765, "dur": 36.217,
    "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": 870, "tid": 870,
    "ts": 7429770631652.502, "dur": 3.018,
    "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": 870, "tid": 870,
    "ts": 7429770631654.074, "dur": 0.880,
    "args": {
      "External id": 5918,"Record function id": 0, "finished": true, "Ev Idx": 5917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770631654.500, "dur": 0.332,
    "args": {
      "External id": 5919,"Record function id": 0, "finished": true, "Ev Idx": 5918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770631655.940, "dur": 31.853,
    "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": 870, "tid": 870,
    "ts": 7429770631656.480, "dur": 1.809,
    "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": 870, "tid": 870,
    "ts": 7429770631656.755, "dur": 1.416,
    "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": 870, "tid": 870,
    "ts": 7429770631658.699, "dur": 27.245,
    "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": 870, "tid": 870,
    "ts": 7429770631669.725, "dur": 1.160,
    "args": {
      "External id": 5924,"Record function id": 0, "finished": true, "Ev Idx": 5923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770631686.522, "dur": 0.978,
    "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": 870, "tid": 870,
    "ts": 7429770631693.333, "dur": 8.363,
    "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": 870, "tid": 870,
    "ts": 7429770631710.880, "dur": 0.187,
    "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": 870, "tid": 870,
    "ts": 7429770631713.294, "dur": 12.063,
    "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": 870, "tid": 870,
    "ts": 7429770631715.138, "dur": 0.420,
    "args": {
      "External id": 5929,"Record function id": 0, "finished": true, "Ev Idx": 5928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770631715.988, "dur": 0.106,
    "args": {
      "External id": 5930,"Record function id": 0, "finished": true, "Ev Idx": 5929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770631727.952, "dur": 13.906,
    "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": 870, "tid": 870,
    "ts": 7429770631745.228, "dur": 7.911,
    "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": 870, "tid": 870,
    "ts": 7429770631755.067, "dur": 10.865,
    "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": 870, "tid": 870,
    "ts": 7429770631767.631, "dur": 9.671,
    "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": 870, "tid": 870,
    "ts": 7429770631779.657, "dur": 0.135,
    "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": 870, "tid": 870,
    "ts": 7429770631780.608, "dur": 8.017,
    "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": 870, "tid": 870,
    "ts": 7429770631803.437, "dur": 35.718,
    "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": 870, "tid": 870,
    "ts": 7429770631803.959, "dur": 3.626,
    "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": 870, "tid": 870,
    "ts": 7429770631805.685, "dur": 1.237,
    "args": {
      "External id": 5939,"Record function id": 0, "finished": true, "Ev Idx": 5938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770631806.370, "dur": 0.420,
    "args": {
      "External id": 5940,"Record function id": 0, "finished": true, "Ev Idx": 5939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770631807.997, "dur": 31.022,
    "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": 870, "tid": 870,
    "ts": 7429770631808.620, "dur": 2.027,
    "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": 870, "tid": 870,
    "ts": 7429770631808.924, "dur": 1.609,
    "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": 870, "tid": 870,
    "ts": 7429770631811.032, "dur": 26.139,
    "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": 870, "tid": 870,
    "ts": 7429770631821.095, "dur": 1.143,
    "args": {
      "External id": 5945,"Record function id": 0, "finished": true, "Ev Idx": 5944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770631837.731, "dur": 1.003,
    "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": 870, "tid": 870,
    "ts": 7429770631846.937, "dur": 11.296,
    "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": 870, "tid": 870,
    "ts": 7429770631865.046, "dur": 32.975,
    "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": 870, "tid": 870,
    "ts": 7429770631865.354, "dur": 2.386,
    "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": 870, "tid": 870,
    "ts": 7429770631866.265, "dur": 0.960,
    "args": {
      "External id": 5950,"Record function id": 0, "finished": true, "Ev Idx": 5949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770631866.651, "dur": 0.460,
    "args": {
      "External id": 5951,"Record function id": 0, "finished": true, "Ev Idx": 5950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770631868.055, "dur": 29.848,
    "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": 870, "tid": 870,
    "ts": 7429770631868.535, "dur": 1.509,
    "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": 870, "tid": 870,
    "ts": 7429770631868.898, "dur": 1.047,
    "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": 870, "tid": 870,
    "ts": 7429770631870.344, "dur": 25.731,
    "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": 870, "tid": 870,
    "ts": 7429770631881.008, "dur": 1.238,
    "args": {
      "External id": 5956,"Record function id": 0, "finished": true, "Ev Idx": 5955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770631896.705, "dur": 0.954,
    "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": 870, "tid": 870,
    "ts": 7429770631899.714, "dur": 8.279,
    "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": 870, "tid": 870,
    "ts": 7429770631913.278, "dur": 101.189,
    "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": 870, "tid": 870,
    "ts": 7429770631913.638, "dur": 2.394,
    "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": 870, "tid": 870,
    "ts": 7429770631914.546, "dur": 1.019,
    "args": {
      "External id": 5961,"Record function id": 0, "finished": true, "Ev Idx": 5960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770631915.045, "dur": 0.402,
    "args": {
      "External id": 5962,"Record function id": 0, "finished": true, "Ev Idx": 5961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770631916.328, "dur": 98.009,
    "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": 870, "tid": 870,
    "ts": 7429770631916.725, "dur": 1.337,
    "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": 870, "tid": 870,
    "ts": 7429770631917.007, "dur": 0.936,
    "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": 870, "tid": 870,
    "ts": 7429770631918.403, "dur": 93.450,
    "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": 870, "tid": 870,
    "ts": 7429770631927.070, "dur": 55.167,
    "args": {
      "External id": 5967,"Record function id": 0, "finished": true, "Ev Idx": 5966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770632012.535, "dur": 1.478,
    "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": 870, "tid": 870,
    "ts": 7429770632017.842, "dur": 9.054,
    "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": 870, "tid": 870,
    "ts": 7429770632047.503, "dur": 0.247,
    "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": 870, "tid": 870,
    "ts": 7429770632049.684, "dur": 10.746,
    "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": 870, "tid": 870,
    "ts": 7429770632051.155, "dur": 0.341,
    "args": {
      "External id": 5972,"Record function id": 0, "finished": true, "Ev Idx": 5971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770632051.858, "dur": 0.073,
    "args": {
      "External id": 5973,"Record function id": 0, "finished": true, "Ev Idx": 5972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770632062.903, "dur": 11.936,
    "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": 870, "tid": 870,
    "ts": 7429770632078.420, "dur": 9.248,
    "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": 870, "tid": 870,
    "ts": 7429770632089.387, "dur": 8.732,
    "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": 870, "tid": 870,
    "ts": 7429770632099.650, "dur": 7.387,
    "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": 870, "tid": 870,
    "ts": 7429770632109.165, "dur": 0.218,
    "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": 870, "tid": 870,
    "ts": 7429770632110.145, "dur": 7.968,
    "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": 870, "tid": 870,
    "ts": 7429770632135.952, "dur": 38.262,
    "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": 870, "tid": 870,
    "ts": 7429770632136.434, "dur": 3.920,
    "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": 870, "tid": 870,
    "ts": 7429770632137.953, "dur": 1.565,
    "args": {
      "External id": 5982,"Record function id": 0, "finished": true, "Ev Idx": 5981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770632138.724, "dur": 0.595,
    "args": {
      "External id": 5983,"Record function id": 0, "finished": true, "Ev Idx": 5982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770632140.759, "dur": 33.329,
    "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": 870, "tid": 870,
    "ts": 7429770632141.264, "dur": 1.714,
    "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": 870, "tid": 870,
    "ts": 7429770632141.599, "dur": 1.239,
    "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": 870, "tid": 870,
    "ts": 7429770632143.380, "dur": 28.883,
    "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": 870, "tid": 870,
    "ts": 7429770632154.535, "dur": 1.356,
    "args": {
      "External id": 5988,"Record function id": 0, "finished": true, "Ev Idx": 5987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770632172.815, "dur": 0.998,
    "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": 870, "tid": 870,
    "ts": 7429770632178.020, "dur": 1.516,
    "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": 870, "tid": 870,
    "ts": 7429770632181.368, "dur": 2.507,
    "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": 870, "tid": 870,
    "ts": 7429770632182.533, "dur": 0.548,
    "args": {
      "External id": 5992,"Record function id": 0, "finished": true, "Ev Idx": 5991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770632190.745, "dur": 33.792,
    "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": 870, "tid": 870,
    "ts": 7429770632191.115, "dur": 2.274,
    "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": 870, "tid": 870,
    "ts": 7429770632192.074, "dur": 0.824,
    "args": {
      "External id": 5995,"Record function id": 0, "finished": true, "Ev Idx": 5994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770632192.541, "dur": 0.257,
    "args": {
      "External id": 5996,"Record function id": 0, "finished": true, "Ev Idx": 5995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770632193.712, "dur": 30.732,
    "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": 870, "tid": 870,
    "ts": 7429770632194.129, "dur": 1.425,
    "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": 870, "tid": 870,
    "ts": 7429770632194.424, "dur": 1.018,
    "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": 870, "tid": 870,
    "ts": 7429770632195.879, "dur": 26.831,
    "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": 870, "tid": 870,
    "ts": 7429770632205.520, "dur": 1.134,
    "args": {
      "External id": 6001,"Record function id": 0, "finished": true, "Ev Idx": 6000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770632223.233, "dur": 0.987,
    "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": 870, "tid": 870,
    "ts": 7429770632226.869, "dur": 1.159,
    "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": 870, "tid": 870,
    "ts": 7429770632229.353, "dur": 1.593,
    "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": 870, "tid": 870,
    "ts": 7429770632230.093, "dur": 0.340,
    "args": {
      "External id": 6005,"Record function id": 0, "finished": true, "Ev Idx": 6004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770632237.313, "dur": 31.968,
    "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": 870, "tid": 870,
    "ts": 7429770632237.613, "dur": 2.011,
    "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": 870, "tid": 870,
    "ts": 7429770632238.484, "dur": 0.734,
    "args": {
      "External id": 6008,"Record function id": 0, "finished": true, "Ev Idx": 6007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770632238.836, "dur": 0.290,
    "args": {
      "External id": 6009,"Record function id": 0, "finished": true, "Ev Idx": 6008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770632239.907, "dur": 29.295,
    "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": 870, "tid": 870,
    "ts": 7429770632240.297, "dur": 1.293,
    "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": 870, "tid": 870,
    "ts": 7429770632240.677, "dur": 0.809,
    "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": 870, "tid": 870,
    "ts": 7429770632241.893, "dur": 25.628,
    "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": 870, "tid": 870,
    "ts": 7429770632253.062, "dur": 1.092,
    "args": {
      "External id": 6014,"Record function id": 0, "finished": true, "Ev Idx": 6013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770632268.054, "dur": 0.922,
    "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": 870, "tid": 870,
    "ts": 7429770632271.574, "dur": 1.099,
    "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": 870, "tid": 870,
    "ts": 7429770632274.032, "dur": 1.532,
    "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": 870, "tid": 870,
    "ts": 7429770632274.759, "dur": 0.303,
    "args": {
      "External id": 6018,"Record function id": 0, "finished": true, "Ev Idx": 6017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770632278.315, "dur": 2.021,
    "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": 870, "tid": 870,
    "ts": 7429770632279.491, "dur": 0.287,
    "args": {
      "External id": 6020,"Record function id": 0, "finished": true, "Ev Idx": 6019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770632281.436, "dur": 1.285,
    "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": 870, "tid": 870,
    "ts": 7429770632282.105, "dur": 0.305,
    "args": {
      "External id": 6022,"Record function id": 0, "finished": true, "Ev Idx": 6021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770632283.671, "dur": 9.146,
    "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": 870, "tid": 870,
    "ts": 7429770632297.590, "dur": 4.875,
    "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": 870, "tid": 870,
    "ts": 7429770632299.918, "dur": 0.806,
    "args": {
      "External id": 6025,"Record function id": 0, "finished": true, "Ev Idx": 6024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770632304.577, "dur": 2.565,
    "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": 870, "tid": 870,
    "ts": 7429770632306.220, "dur": 0.326,
    "args": {
      "External id": 6027,"Record function id": 0, "finished": true, "Ev Idx": 6026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770632317.122, "dur": 15.023,
    "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": 870, "tid": 870,
    "ts": 7429770632337.214, "dur": 17.832,
    "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": 870, "tid": 870,
    "ts": 7429770632358.296, "dur": 10.172,
    "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": 870, "tid": 870,
    "ts": 7429770632370.131, "dur": 8.970,
    "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": 870, "tid": 870,
    "ts": 7429770632381.355, "dur": 7.983,
    "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": 870, "tid": 870,
    "ts": 7429770632392.355, "dur": 2.490,
    "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": 870, "tid": 870,
    "ts": 7429770632393.538, "dur": 0.487,
    "args": {
      "External id": 6034,"Record function id": 0, "finished": true, "Ev Idx": 6033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770632396.483, "dur": 1.676,
    "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": 870, "tid": 870,
    "ts": 7429770632397.282, "dur": 0.348,
    "args": {
      "External id": 6036,"Record function id": 0, "finished": true, "Ev Idx": 6035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770632399.225, "dur": 7.711,
    "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": 870, "tid": 870,
    "ts": 7429770632408.855, "dur": 9.019,
    "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": 870, "tid": 870,
    "ts": 7429770632420.187, "dur": 8.282,
    "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": 870, "tid": 870,
    "ts": 7429770632429.820, "dur": 7.152,
    "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": 870, "tid": 870,
    "ts": 7429770632462.150, "dur": 2.635,
    "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": 870, "tid": 870,
    "ts": 7429770632463.546, "dur": 0.412,
    "args": {
      "External id": 6042,"Record function id": 0, "finished": true, "Ev Idx": 6041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770632465.473, "dur": 1.986,
    "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": 870, "tid": 870,
    "ts": 7429770632466.468, "dur": 0.460,
    "args": {
      "External id": 6044,"Record function id": 0, "finished": true, "Ev Idx": 6043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770632468.200, "dur": 2.484,
    "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": 870, "tid": 870,
    "ts": 7429770632469.372, "dur": 0.634,
    "args": {
      "External id": 6046,"Record function id": 0, "finished": true, "Ev Idx": 6045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770632471.337, "dur": 2.126,
    "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": 870, "tid": 870,
    "ts": 7429770632472.607, "dur": 0.341,
    "args": {
      "External id": 6048,"Record function id": 0, "finished": true, "Ev Idx": 6047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770632474.105, "dur": 1.628,
    "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": 870, "tid": 870,
    "ts": 7429770632474.943, "dur": 0.310,
    "args": {
      "External id": 6050,"Record function id": 0, "finished": true, "Ev Idx": 6049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770632478.596, "dur": 2.931,
    "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": 870, "tid": 870,
    "ts": 7429770632480.426, "dur": 0.393,
    "args": {
      "External id": 6052,"Record function id": 0, "finished": true, "Ev Idx": 6051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770632483.846, "dur": 20.348,
    "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": 870, "tid": 870,
    "ts": 7429770632486.601, "dur": 14.955,
    "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": 870, "tid": 870,
    "ts": 7429770632488.274, "dur": 2.856,
    "args": {
      "External id": 6055,"Record function id": 0, "finished": true, "Ev Idx": 6054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770632489.173, "dur": 1.701,
    "args": {
      "External id": 6056,"Record function id": 0, "finished": true, "Ev Idx": 6055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770632491.826, "dur": 9.260,
    "args": {
      "External id": 6057,"Record function id": 0, "finished": true, "Ev Idx": 6056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770632502.035, "dur": 1.876,
    "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": 870, "tid": 870,
    "ts": 7429770632507.469, "dur": 2.050,
    "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": 870, "tid": 870,
    "ts": 7429770632508.560, "dur": 0.377,
    "args": {
      "External id": 6060,"Record function id": 0, "finished": true, "Ev Idx": 6059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770632510.146, "dur": 1.490,
    "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": 870, "tid": 870,
    "ts": 7429770632510.888, "dur": 0.266,
    "args": {
      "External id": 6062,"Record function id": 0, "finished": true, "Ev Idx": 6061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770632512.163, "dur": 1.531,
    "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": 870, "tid": 870,
    "ts": 7429770632512.911, "dur": 0.369,
    "args": {
      "External id": 6064,"Record function id": 0, "finished": true, "Ev Idx": 6063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770632514.270, "dur": 1.554,
    "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": 870, "tid": 870,
    "ts": 7429770632515.002, "dur": 0.309,
    "args": {
      "External id": 6066,"Record function id": 0, "finished": true, "Ev Idx": 6065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770632516.519, "dur": 1.563,
    "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": 870, "tid": 870,
    "ts": 7429770632517.323, "dur": 0.271,
    "args": {
      "External id": 6068,"Record function id": 0, "finished": true, "Ev Idx": 6067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770632519.561, "dur": 2.122,
    "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": 870, "tid": 870,
    "ts": 7429770632520.927, "dur": 0.324,
    "args": {
      "External id": 6070,"Record function id": 0, "finished": true, "Ev Idx": 6069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770632523.107, "dur": 12.720,
    "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": 870, "tid": 870,
    "ts": 7429770632523.991, "dur": 10.341,
    "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": 870, "tid": 870,
    "ts": 7429770632525.289, "dur": 1.861,
    "args": {
      "External id": 6073,"Record function id": 0, "finished": true, "Ev Idx": 6072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770632525.810, "dur": 1.228,
    "args": {
      "External id": 6074,"Record function id": 0, "finished": true, "Ev Idx": 6073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770632527.524, "dur": 6.508,
    "args": {
      "External id": 6075,"Record function id": 0, "finished": true, "Ev Idx": 6074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770632534.764, "dur": 0.914,
    "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": 870, "tid": 870,
    "ts": 7429770632538.066, "dur": 1.720,
    "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": 870, "tid": 870,
    "ts": 7429770632538.946, "dur": 0.362,
    "args": {
      "External id": 6078,"Record function id": 0, "finished": true, "Ev Idx": 6077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770632541.325, "dur": 108.821,
    "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": 870, "tid": 870,
    "ts": 7429770632542.111, "dur": 1.593,
    "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": 870, "tid": 870,
    "ts": 7429770632542.916, "dur": 0.346,
    "args": {
      "External id": 6081,"Record function id": 0, "finished": true, "Ev Idx": 6080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770632544.041, "dur": 4.522,
    "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": 870, "tid": 870,
    "ts": 7429770632545.844, "dur": 2.585,
    "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": 870, "tid": 870,
    "ts": 7429770632549.013, "dur": 2.239,
    "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": 870, "tid": 870,
    "ts": 7429770632550.533, "dur": 0.320,
    "args": {
      "External id": 6085,"Record function id": 0, "finished": true, "Ev Idx": 6084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770632551.526, "dur": 1.770,
    "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": 870, "tid": 870,
    "ts": 7429770632552.299, "dur": 0.890,
    "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": 870, "tid": 870,
    "ts": 7429770632553.937, "dur": 93.842,
    "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": 870, "tid": 870,
    "ts": 7429770632622.232, "dur": 1.488,
    "args": {
      "External id": 6089,"Record function id": 0, "finished": true, "Ev Idx": 6088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770632648.480, "dur": 1.261,
    "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": 870, "tid": 870,
    "ts": 7429770632654.618, "dur": 11.593,
    "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": 870, "tid": 870,
    "ts": 7429770632669.880, "dur": 2.417,
    "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": 870, "tid": 870,
    "ts": 7429770632670.914, "dur": 0.476,
    "args": {
      "External id": 6093,"Record function id": 0, "finished": true, "Ev Idx": 6092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770632672.996, "dur": 1.492,
    "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": 870, "tid": 870,
    "ts": 7429770632673.689, "dur": 0.373,
    "args": {
      "External id": 6095,"Record function id": 0, "finished": true, "Ev Idx": 6094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770632675.222, "dur": 0.977,
    "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": 870, "tid": 870,
    "ts": 7429770632675.679, "dur": 0.208,
    "args": {
      "External id": 6097,"Record function id": 0, "finished": true, "Ev Idx": 6096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770632678.042, "dur": 11.389,
    "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": 870, "tid": 870,
    "ts": 7429770632695.335, "dur": 14.982,
    "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": 870, "tid": 870,
    "ts": 7429770632696.581, "dur": 0.430,
    "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": 870, "tid": 870,
    "ts": 7429770632697.827, "dur": 12.245,
    "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": 870, "tid": 870,
    "ts": 7429770632713.668, "dur": 0.243,
    "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": 870, "tid": 870,
    "ts": 7429770632727.068, "dur": 0.475,
    "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": 870, "tid": 870,
    "ts": 7429770632729.622, "dur": 44.127,
    "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": 870, "tid": 870,
    "ts": 7429770632730.257, "dur": 2.494,
    "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": 870, "tid": 870,
    "ts": 7429770632731.833, "dur": 0.318,
    "args": {
      "External id": 6106,"Record function id": 0, "finished": true, "Ev Idx": 6105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770632733.197, "dur": 2.633,
    "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": 870, "tid": 870,
    "ts": 7429770632733.611, "dur": 2.089,
    "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": 870, "tid": 870,
    "ts": 7429770632736.361, "dur": 1.469,
    "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": 870, "tid": 870,
    "ts": 7429770632737.121, "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": 870, "tid": 870,
    "ts": 7429770632738.161, "dur": 1.698,
    "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": 870, "tid": 870,
    "ts": 7429770632738.505, "dur": 1.290,
    "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": 870, "tid": 870,
    "ts": 7429770632740.336, "dur": 31.420,
    "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": 870, "tid": 870,
    "ts": 7429770632752.272, "dur": 1.292,
    "args": {
      "External id": 6114,"Record function id": 0, "finished": true, "Ev Idx": 6113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770632772.349, "dur": 1.098,
    "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": 870, "tid": 870,
    "ts": 7429770632775.978, "dur": 2.084,
    "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": 870, "tid": 870,
    "ts": 7429770632777.123, "dur": 0.402,
    "args": {
      "External id": 6117,"Record function id": 0, "finished": true, "Ev Idx": 6116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770632780.525, "dur": 21.263,
    "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": 870, "tid": 870,
    "ts": 7429770632781.099, "dur": 20.544,
    "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": 870, "tid": 870,
    "ts": 7429770632782.200, "dur": 2.437,
    "args": {
      "External id": 6120,"Record function id": 0, "finished": true, "Ev Idx": 6119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770632782.854, "dur": 1.605,
    "args": {
      "External id": 6121,"Record function id": 0, "finished": true, "Ev Idx": 6120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770632785.075, "dur": 15.998,
    "args": {
      "External id": 6122,"Record function id": 0, "finished": true, "Ev Idx": 6121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770632808.168, "dur": 1.658,
    "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": 870, "tid": 870,
    "ts": 7429770632808.469, "dur": 1.258,
    "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": 870, "tid": 870,
    "ts": 7429770632827.727, "dur": 35.633,
    "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": 870, "tid": 870,
    "ts": 7429770632828.619, "dur": 3.434,
    "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": 870, "tid": 870,
    "ts": 7429770632830.407, "dur": 1.037,
    "args": {
      "External id": 6127,"Record function id": 0, "finished": true, "Ev Idx": 6126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770632830.869, "dur": 0.458,
    "args": {
      "External id": 6128,"Record function id": 0, "finished": true, "Ev Idx": 6127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770632832.496, "dur": 30.659,
    "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": 870, "tid": 870,
    "ts": 7429770632832.965, "dur": 1.650,
    "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": 870, "tid": 870,
    "ts": 7429770632833.350, "dur": 1.167,
    "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": 870, "tid": 870,
    "ts": 7429770632835.135, "dur": 25.964,
    "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": 870, "tid": 870,
    "ts": 7429770632844.862, "dur": 1.403,
    "args": {
      "External id": 6133,"Record function id": 0, "finished": true, "Ev Idx": 6132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770632861.764, "dur": 1.110,
    "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": 870, "tid": 870,
    "ts": 7429770632869.065, "dur": 8.780,
    "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": 870, "tid": 870,
    "ts": 7429770632887.807, "dur": 0.344,
    "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": 870, "tid": 870,
    "ts": 7429770632890.787, "dur": 13.062,
    "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": 870, "tid": 870,
    "ts": 7429770632893.016, "dur": 0.568,
    "args": {
      "External id": 6138,"Record function id": 0, "finished": true, "Ev Idx": 6137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770632893.980, "dur": 0.088,
    "args": {
      "External id": 6139,"Record function id": 0, "finished": true, "Ev Idx": 6138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770632906.970, "dur": 13.939,
    "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": 870, "tid": 870,
    "ts": 7429770632924.719, "dur": 9.166,
    "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": 870, "tid": 870,
    "ts": 7429770632936.083, "dur": 9.512,
    "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": 870, "tid": 870,
    "ts": 7429770632947.410, "dur": 9.395,
    "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": 870, "tid": 870,
    "ts": 7429770632959.471, "dur": 0.262,
    "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": 870, "tid": 870,
    "ts": 7429770632984.423, "dur": 7.795,
    "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": 870, "tid": 870,
    "ts": 7429770633007.778, "dur": 36.613,
    "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": 870, "tid": 870,
    "ts": 7429770633008.249, "dur": 3.012,
    "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": 870, "tid": 870,
    "ts": 7429770633009.380, "dur": 1.123,
    "args": {
      "External id": 6148,"Record function id": 0, "finished": true, "Ev Idx": 6147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770633009.907, "dur": 0.457,
    "args": {
      "External id": 6149,"Record function id": 0, "finished": true, "Ev Idx": 6148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770633011.629, "dur": 32.574,
    "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": 870, "tid": 870,
    "ts": 7429770633012.077, "dur": 1.717,
    "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": 870, "tid": 870,
    "ts": 7429770633012.432, "dur": 1.248,
    "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": 870, "tid": 870,
    "ts": 7429770633014.150, "dur": 28.146,
    "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": 870, "tid": 870,
    "ts": 7429770633024.924, "dur": 1.170,
    "args": {
      "External id": 6154,"Record function id": 0, "finished": true, "Ev Idx": 6153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770633042.944, "dur": 0.954,
    "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": 870, "tid": 870,
    "ts": 7429770633052.948, "dur": 10.795,
    "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": 870, "tid": 870,
    "ts": 7429770633071.077, "dur": 32.541,
    "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": 870, "tid": 870,
    "ts": 7429770633071.430, "dur": 2.238,
    "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": 870, "tid": 870,
    "ts": 7429770633072.276, "dur": 0.864,
    "args": {
      "External id": 6159,"Record function id": 0, "finished": true, "Ev Idx": 6158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770633072.696, "dur": 0.334,
    "args": {
      "External id": 6160,"Record function id": 0, "finished": true, "Ev Idx": 6159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770633074.009, "dur": 29.493,
    "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": 870, "tid": 870,
    "ts": 7429770633074.419, "dur": 1.422,
    "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": 870, "tid": 870,
    "ts": 7429770633074.676, "dur": 1.056,
    "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": 870, "tid": 870,
    "ts": 7429770633076.153, "dur": 25.477,
    "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": 870, "tid": 870,
    "ts": 7429770633086.799, "dur": 1.208,
    "args": {
      "External id": 6165,"Record function id": 0, "finished": true, "Ev Idx": 6164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770633102.221, "dur": 0.995,
    "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": 870, "tid": 870,
    "ts": 7429770633105.238, "dur": 8.235,
    "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": 870, "tid": 870,
    "ts": 7429770633119.373, "dur": 44.565,
    "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": 870, "tid": 870,
    "ts": 7429770633119.739, "dur": 2.234,
    "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": 870, "tid": 870,
    "ts": 7429770633120.619, "dur": 0.839,
    "args": {
      "External id": 6170,"Record function id": 0, "finished": true, "Ev Idx": 6169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770633120.987, "dur": 0.346,
    "args": {
      "External id": 6171,"Record function id": 0, "finished": true, "Ev Idx": 6170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770633122.279, "dur": 41.540,
    "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": 870, "tid": 870,
    "ts": 7429770633122.645, "dur": 1.497,
    "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": 870, "tid": 870,
    "ts": 7429770633122.916, "dur": 1.093,
    "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": 870, "tid": 870,
    "ts": 7429770633124.427, "dur": 37.610,
    "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": 870, "tid": 870,
    "ts": 7429770633133.626, "dur": 1.089,
    "args": {
      "External id": 6176,"Record function id": 0, "finished": true, "Ev Idx": 6175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770633162.636, "dur": 0.893,
    "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": 870, "tid": 870,
    "ts": 7429770633166.739, "dur": 8.915,
    "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": 870, "tid": 870,
    "ts": 7429770633195.856, "dur": 0.160,
    "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": 870, "tid": 870,
    "ts": 7429770633197.703, "dur": 9.370,
    "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": 870, "tid": 870,
    "ts": 7429770633198.905, "dur": 0.289,
    "args": {
      "External id": 6181,"Record function id": 0, "finished": true, "Ev Idx": 6180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770633199.506, "dur": 0.075,
    "args": {
      "External id": 6182,"Record function id": 0, "finished": true, "Ev Idx": 6181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770633209.610, "dur": 10.920,
    "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": 870, "tid": 870,
    "ts": 7429770633223.782, "dur": 7.356,
    "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": 870, "tid": 870,
    "ts": 7429770633232.762, "dur": 7.873,
    "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": 870, "tid": 870,
    "ts": 7429770633242.199, "dur": 7.034,
    "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": 870, "tid": 870,
    "ts": 7429770633251.600, "dur": 0.124,
    "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": 870, "tid": 870,
    "ts": 7429770633252.418, "dur": 8.083,
    "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": 870, "tid": 870,
    "ts": 7429770633278.118, "dur": 35.591,
    "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": 870, "tid": 870,
    "ts": 7429770633278.492, "dur": 3.371,
    "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": 870, "tid": 870,
    "ts": 7429770633279.767, "dur": 1.311,
    "args": {
      "External id": 6191,"Record function id": 0, "finished": true, "Ev Idx": 6190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770633280.437, "dur": 0.485,
    "args": {
      "External id": 6192,"Record function id": 0, "finished": true, "Ev Idx": 6191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770633282.238, "dur": 31.346,
    "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": 870, "tid": 870,
    "ts": 7429770633282.696, "dur": 1.889,
    "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": 870, "tid": 870,
    "ts": 7429770633283.016, "dur": 1.430,
    "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": 870, "tid": 870,
    "ts": 7429770633284.985, "dur": 26.674,
    "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": 870, "tid": 870,
    "ts": 7429770633294.993, "dur": 1.216,
    "args": {
      "External id": 6197,"Record function id": 0, "finished": true, "Ev Idx": 6196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770633312.254, "dur": 1.061,
    "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": 870, "tid": 870,
    "ts": 7429770633317.173, "dur": 1.233,
    "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": 870, "tid": 870,
    "ts": 7429770633320.350, "dur": 1.932,
    "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": 870, "tid": 870,
    "ts": 7429770633321.255, "dur": 0.397,
    "args": {
      "External id": 6201,"Record function id": 0, "finished": true, "Ev Idx": 6200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770633328.965, "dur": 32.676,
    "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": 870, "tid": 870,
    "ts": 7429770633329.289, "dur": 1.969,
    "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": 870, "tid": 870,
    "ts": 7429770633330.054, "dur": 0.744,
    "args": {
      "External id": 6204,"Record function id": 0, "finished": true, "Ev Idx": 6203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770633330.425, "dur": 0.275,
    "args": {
      "External id": 6205,"Record function id": 0, "finished": true, "Ev Idx": 6204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770633331.535, "dur": 30.028,
    "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": 870, "tid": 870,
    "ts": 7429770633332.066, "dur": 1.225,
    "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": 870, "tid": 870,
    "ts": 7429770633332.339, "dur": 0.829,
    "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": 870, "tid": 870,
    "ts": 7429770633333.597, "dur": 26.203,
    "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": 870, "tid": 870,
    "ts": 7429770633343.195, "dur": 1.104,
    "args": {
      "External id": 6210,"Record function id": 0, "finished": true, "Ev Idx": 6209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770633360.388, "dur": 0.915,
    "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": 870, "tid": 870,
    "ts": 7429770633363.994, "dur": 1.107,
    "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": 870, "tid": 870,
    "ts": 7429770633366.507, "dur": 1.557,
    "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": 870, "tid": 870,
    "ts": 7429770633367.266, "dur": 0.330,
    "args": {
      "External id": 6214,"Record function id": 0, "finished": true, "Ev Idx": 6213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770633374.433, "dur": 29.632,
    "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": 870, "tid": 870,
    "ts": 7429770633374.733, "dur": 1.825,
    "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": 870, "tid": 870,
    "ts": 7429770633375.433, "dur": 0.701,
    "args": {
      "External id": 6217,"Record function id": 0, "finished": true, "Ev Idx": 6216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770633375.778, "dur": 0.255,
    "args": {
      "External id": 6218,"Record function id": 0, "finished": true, "Ev Idx": 6217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770633376.832, "dur": 27.155,
    "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": 870, "tid": 870,
    "ts": 7429770633377.204, "dur": 1.177,
    "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": 870, "tid": 870,
    "ts": 7429770633377.449, "dur": 0.824,
    "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": 870, "tid": 870,
    "ts": 7429770633378.686, "dur": 23.656,
    "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": 870, "tid": 870,
    "ts": 7429770633388.479, "dur": 1.051,
    "args": {
      "External id": 6223,"Record function id": 0, "finished": true, "Ev Idx": 6222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770633402.840, "dur": 0.898,
    "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": 870, "tid": 870,
    "ts": 7429770633406.340, "dur": 1.104,
    "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": 870, "tid": 870,
    "ts": 7429770633408.833, "dur": 1.492,
    "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": 870, "tid": 870,
    "ts": 7429770633409.541, "dur": 0.343,
    "args": {
      "External id": 6227,"Record function id": 0, "finished": true, "Ev Idx": 6226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770633413.383, "dur": 1.904,
    "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": 870, "tid": 870,
    "ts": 7429770633414.484, "dur": 0.300,
    "args": {
      "External id": 6229,"Record function id": 0, "finished": true, "Ev Idx": 6228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770633416.445, "dur": 0.941,
    "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": 870, "tid": 870,
    "ts": 7429770633416.839, "dur": 0.257,
    "args": {
      "External id": 6231,"Record function id": 0, "finished": true, "Ev Idx": 6230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770633418.196, "dur": 9.519,
    "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": 870, "tid": 870,
    "ts": 7429770633432.329, "dur": 4.373,
    "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": 870, "tid": 870,
    "ts": 7429770633434.969, "dur": 0.609,
    "args": {
      "External id": 6234,"Record function id": 0, "finished": true, "Ev Idx": 6233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770633438.932, "dur": 1.697,
    "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": 870, "tid": 870,
    "ts": 7429770633439.725, "dur": 0.324,
    "args": {
      "External id": 6236,"Record function id": 0, "finished": true, "Ev Idx": 6235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770633442.197, "dur": 11.044,
    "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": 870, "tid": 870,
    "ts": 7429770633457.287, "dur": 17.027,
    "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": 870, "tid": 870,
    "ts": 7429770633477.246, "dur": 7.865,
    "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": 870, "tid": 870,
    "ts": 7429770633486.701, "dur": 11.837,
    "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": 870, "tid": 870,
    "ts": 7429770633500.628, "dur": 9.097,
    "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": 870, "tid": 870,
    "ts": 7429770633512.083, "dur": 2.529,
    "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": 870, "tid": 870,
    "ts": 7429770633513.499, "dur": 0.467,
    "args": {
      "External id": 6243,"Record function id": 0, "finished": true, "Ev Idx": 6242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770633516.258, "dur": 1.700,
    "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": 870, "tid": 870,
    "ts": 7429770633517.085, "dur": 0.319,
    "args": {
      "External id": 6245,"Record function id": 0, "finished": true, "Ev Idx": 6244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770633518.836, "dur": 7.561,
    "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": 870, "tid": 870,
    "ts": 7429770633528.226, "dur": 10.096,
    "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": 870, "tid": 870,
    "ts": 7429770633540.508, "dur": 8.488,
    "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": 870, "tid": 870,
    "ts": 7429770633550.423, "dur": 7.931,
    "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": 870, "tid": 870,
    "ts": 7429770633580.647, "dur": 2.426,
    "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": 870, "tid": 870,
    "ts": 7429770633581.832, "dur": 0.405,
    "args": {
      "External id": 6251,"Record function id": 0, "finished": true, "Ev Idx": 6250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770633583.765, "dur": 2.373,
    "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": 870, "tid": 870,
    "ts": 7429770633585.239, "dur": 0.408,
    "args": {
      "External id": 6253,"Record function id": 0, "finished": true, "Ev Idx": 6252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770633586.838, "dur": 2.274,
    "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": 870, "tid": 870,
    "ts": 7429770633587.947, "dur": 0.513,
    "args": {
      "External id": 6255,"Record function id": 0, "finished": true, "Ev Idx": 6254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770633589.731, "dur": 2.052,
    "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": 870, "tid": 870,
    "ts": 7429770633590.906, "dur": 0.346,
    "args": {
      "External id": 6257,"Record function id": 0, "finished": true, "Ev Idx": 6256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770633592.379, "dur": 2.401,
    "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": 870, "tid": 870,
    "ts": 7429770633593.931, "dur": 0.389,
    "args": {
      "External id": 6259,"Record function id": 0, "finished": true, "Ev Idx": 6258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770633597.205, "dur": 3.093,
    "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": 870, "tid": 870,
    "ts": 7429770633599.226, "dur": 0.381,
    "args": {
      "External id": 6261,"Record function id": 0, "finished": true, "Ev Idx": 6260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770633602.522, "dur": 18.329,
    "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": 870, "tid": 870,
    "ts": 7429770633604.943, "dur": 14.019,
    "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": 870, "tid": 870,
    "ts": 7429770633606.210, "dur": 2.415,
    "args": {
      "External id": 6264,"Record function id": 0, "finished": true, "Ev Idx": 6263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770633606.924, "dur": 1.466,
    "args": {
      "External id": 6265,"Record function id": 0, "finished": true, "Ev Idx": 6264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770633609.134, "dur": 9.346,
    "args": {
      "External id": 6266,"Record function id": 0, "finished": true, "Ev Idx": 6265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770633619.425, "dur": 1.159,
    "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": 870, "tid": 870,
    "ts": 7429770633623.881, "dur": 2.484,
    "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": 870, "tid": 870,
    "ts": 7429770633625.443, "dur": 0.392,
    "args": {
      "External id": 6269,"Record function id": 0, "finished": true, "Ev Idx": 6268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770633626.977, "dur": 2.272,
    "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": 870, "tid": 870,
    "ts": 7429770633628.518, "dur": 0.273,
    "args": {
      "External id": 6271,"Record function id": 0, "finished": true, "Ev Idx": 6270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770633629.751, "dur": 1.966,
    "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": 870, "tid": 870,
    "ts": 7429770633630.996, "dur": 0.331,
    "args": {
      "External id": 6273,"Record function id": 0, "finished": true, "Ev Idx": 6272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770633632.260, "dur": 1.517,
    "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": 870, "tid": 870,
    "ts": 7429770633632.981, "dur": 0.299,
    "args": {
      "External id": 6275,"Record function id": 0, "finished": true, "Ev Idx": 6274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770633634.329, "dur": 1.427,
    "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": 870, "tid": 870,
    "ts": 7429770633634.978, "dur": 0.319,
    "args": {
      "External id": 6277,"Record function id": 0, "finished": true, "Ev Idx": 6276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770633637.139, "dur": 1.512,
    "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": 870, "tid": 870,
    "ts": 7429770633637.871, "dur": 0.328,
    "args": {
      "External id": 6279,"Record function id": 0, "finished": true, "Ev Idx": 6278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770633640.138, "dur": 11.857,
    "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": 870, "tid": 870,
    "ts": 7429770633641.037, "dur": 9.430,
    "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": 870, "tid": 870,
    "ts": 7429770633641.657, "dur": 1.711,
    "args": {
      "External id": 6282,"Record function id": 0, "finished": true, "Ev Idx": 6281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770633641.987, "dur": 1.258,
    "args": {
      "External id": 6283,"Record function id": 0, "finished": true, "Ev Idx": 6282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770633643.709, "dur": 6.415,
    "args": {
      "External id": 6284,"Record function id": 0, "finished": true, "Ev Idx": 6283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770633650.834, "dur": 0.993,
    "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": 870, "tid": 870,
    "ts": 7429770633654.077, "dur": 1.693,
    "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": 870, "tid": 870,
    "ts": 7429770633654.931, "dur": 0.347,
    "args": {
      "External id": 6287,"Record function id": 0, "finished": true, "Ev Idx": 6286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770633657.240, "dur": 107.574,
    "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": 870, "tid": 870,
    "ts": 7429770633657.729, "dur": 1.559,
    "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": 870, "tid": 870,
    "ts": 7429770633658.543, "dur": 0.331,
    "args": {
      "External id": 6290,"Record function id": 0, "finished": true, "Ev Idx": 6289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770633659.633, "dur": 3.539,
    "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": 870, "tid": 870,
    "ts": 7429770633661.280, "dur": 1.761,
    "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": 870, "tid": 870,
    "ts": 7429770633663.614, "dur": 1.463,
    "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": 870, "tid": 870,
    "ts": 7429770633664.389, "dur": 0.296,
    "args": {
      "External id": 6294,"Record function id": 0, "finished": true, "Ev Idx": 6293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770633665.367, "dur": 2.407,
    "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": 870, "tid": 870,
    "ts": 7429770633666.121, "dur": 1.550,
    "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": 870, "tid": 870,
    "ts": 7429770633668.330, "dur": 94.178,
    "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": 870, "tid": 870,
    "ts": 7429770633741.311, "dur": 1.314,
    "args": {
      "External id": 6298,"Record function id": 0, "finished": true, "Ev Idx": 6297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770633763.117, "dur": 1.325,
    "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": 870, "tid": 870,
    "ts": 7429770633769.013, "dur": 10.460,
    "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": 870, "tid": 870,
    "ts": 7429770633782.887, "dur": 2.260,
    "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": 870, "tid": 870,
    "ts": 7429770633783.802, "dur": 0.476,
    "args": {
      "External id": 6302,"Record function id": 0, "finished": true, "Ev Idx": 6301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770633785.814, "dur": 1.518,
    "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": 870, "tid": 870,
    "ts": 7429770633786.542, "dur": 0.353,
    "args": {
      "External id": 6304,"Record function id": 0, "finished": true, "Ev Idx": 6303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770633788.060, "dur": 1.165,
    "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": 870, "tid": 870,
    "ts": 7429770633788.631, "dur": 0.226,
    "args": {
      "External id": 6306,"Record function id": 0, "finished": true, "Ev Idx": 6305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770633791.033, "dur": 7.919,
    "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": 870, "tid": 870,
    "ts": 7429770633804.465, "dur": 14.126,
    "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": 870, "tid": 870,
    "ts": 7429770633805.663, "dur": 0.465,
    "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": 870, "tid": 870,
    "ts": 7429770633806.808, "dur": 11.494,
    "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": 870, "tid": 870,
    "ts": 7429770633821.500, "dur": 0.232,
    "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": 870, "tid": 870,
    "ts": 7429770633833.039, "dur": 0.496,
    "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": 870, "tid": 870,
    "ts": 7429770633835.482, "dur": 42.011,
    "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": 870, "tid": 870,
    "ts": 7429770633836.008, "dur": 2.579,
    "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": 870, "tid": 870,
    "ts": 7429770633837.641, "dur": 0.381,
    "args": {
      "External id": 6315,"Record function id": 0, "finished": true, "Ev Idx": 6314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770633839.103, "dur": 2.821,
    "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": 870, "tid": 870,
    "ts": 7429770633839.518, "dur": 2.277,
    "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": 870, "tid": 870,
    "ts": 7429770633842.403, "dur": 1.284,
    "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": 870, "tid": 870,
    "ts": 7429770633843.025, "dur": 0.257,
    "args": {
      "External id": 6319,"Record function id": 0, "finished": true, "Ev Idx": 6318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770633844.027, "dur": 1.263,
    "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": 870, "tid": 870,
    "ts": 7429770633844.297, "dur": 0.918,
    "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": 870, "tid": 870,
    "ts": 7429770633845.714, "dur": 29.780,
    "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": 870, "tid": 870,
    "ts": 7429770633856.840, "dur": 1.712,
    "args": {
      "External id": 6323,"Record function id": 0, "finished": true, "Ev Idx": 6322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770633876.106, "dur": 1.081,
    "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": 870, "tid": 870,
    "ts": 7429770633879.683, "dur": 1.915,
    "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": 870, "tid": 870,
    "ts": 7429770633880.685, "dur": 0.344,
    "args": {
      "External id": 6326,"Record function id": 0, "finished": true, "Ev Idx": 6325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770633884.169, "dur": 11.719,
    "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": 870, "tid": 870,
    "ts": 7429770633884.669, "dur": 11.073,
    "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": 870, "tid": 870,
    "ts": 7429770633885.812, "dur": 2.171,
    "args": {
      "External id": 6329,"Record function id": 0, "finished": true, "Ev Idx": 6328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770633886.285, "dur": 1.564,
    "args": {
      "External id": 6330,"Record function id": 0, "finished": true, "Ev Idx": 6329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770633888.443, "dur": 6.760,
    "args": {
      "External id": 6331,"Record function id": 0, "finished": true, "Ev Idx": 6330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770633902.000, "dur": 1.584,
    "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": 870, "tid": 870,
    "ts": 7429770633902.300, "dur": 1.195,
    "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": 870, "tid": 870,
    "ts": 7429770633918.870, "dur": 36.705,
    "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": 870, "tid": 870,
    "ts": 7429770633919.459, "dur": 2.847,
    "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": 870, "tid": 870,
    "ts": 7429770633920.878, "dur": 0.835,
    "args": {
      "External id": 6336,"Record function id": 0, "finished": true, "Ev Idx": 6335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770633921.255, "dur": 0.348,
    "args": {
      "External id": 6337,"Record function id": 0, "finished": true, "Ev Idx": 6336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770633922.692, "dur": 32.681,
    "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": 870, "tid": 870,
    "ts": 7429770633923.157, "dur": 1.487,
    "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": 870, "tid": 870,
    "ts": 7429770633923.430, "dur": 1.112,
    "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": 870, "tid": 870,
    "ts": 7429770633925.181, "dur": 28.278,
    "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": 870, "tid": 870,
    "ts": 7429770633937.889, "dur": 1.165,
    "args": {
      "External id": 6342,"Record function id": 0, "finished": true, "Ev Idx": 6341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770633954.150, "dur": 0.962,
    "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": 870, "tid": 870,
    "ts": 7429770633960.698, "dur": 8.400,
    "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": 870, "tid": 870,
    "ts": 7429770633978.251, "dur": 0.169,
    "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": 870, "tid": 870,
    "ts": 7429770633980.610, "dur": 12.509,
    "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": 870, "tid": 870,
    "ts": 7429770633982.364, "dur": 0.384,
    "args": {
      "External id": 6347,"Record function id": 0, "finished": true, "Ev Idx": 6346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770633983.271, "dur": 0.068,
    "args": {
      "External id": 6348,"Record function id": 0, "finished": true, "Ev Idx": 6347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770633995.864, "dur": 13.169,
    "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": 870, "tid": 870,
    "ts": 7429770634012.312, "dur": 9.262,
    "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": 870, "tid": 870,
    "ts": 7429770634023.507, "dur": 10.278,
    "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": 870, "tid": 870,
    "ts": 7429770634035.415, "dur": 8.947,
    "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": 870, "tid": 870,
    "ts": 7429770634046.766, "dur": 0.092,
    "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": 870, "tid": 870,
    "ts": 7429770634047.774, "dur": 8.597,
    "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": 870, "tid": 870,
    "ts": 7429770634089.193, "dur": 42.829,
    "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": 870, "tid": 870,
    "ts": 7429770634089.800, "dur": 4.493,
    "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": 870, "tid": 870,
    "ts": 7429770634091.504, "dur": 1.898,
    "args": {
      "External id": 6357,"Record function id": 0, "finished": true, "Ev Idx": 6356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770634092.397, "dur": 0.721,
    "args": {
      "External id": 6358,"Record function id": 0, "finished": true, "Ev Idx": 6357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770634094.775, "dur": 37.058,
    "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": 870, "tid": 870,
    "ts": 7429770634095.445, "dur": 2.014,
    "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": 870, "tid": 870,
    "ts": 7429770634095.784, "dur": 1.551,
    "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": 870, "tid": 870,
    "ts": 7429770634097.872, "dur": 31.650,
    "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": 870, "tid": 870,
    "ts": 7429770634110.310, "dur": 1.499,
    "args": {
      "External id": 6363,"Record function id": 0, "finished": true, "Ev Idx": 6362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770634130.230, "dur": 1.209,
    "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": 870, "tid": 870,
    "ts": 7429770634141.466, "dur": 11.151,
    "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": 870, "tid": 870,
    "ts": 7429770634160.946, "dur": 32.636,
    "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": 870, "tid": 870,
    "ts": 7429770634161.303, "dur": 2.731,
    "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": 870, "tid": 870,
    "ts": 7429770634162.445, "dur": 1.061,
    "args": {
      "External id": 6368,"Record function id": 0, "finished": true, "Ev Idx": 6367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770634162.906, "dur": 0.478,
    "args": {
      "External id": 6369,"Record function id": 0, "finished": true, "Ev Idx": 6368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770634164.466, "dur": 29.016,
    "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": 870, "tid": 870,
    "ts": 7429770634164.909, "dur": 2.157,
    "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": 870, "tid": 870,
    "ts": 7429770634165.201, "dur": 1.766,
    "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": 870, "tid": 870,
    "ts": 7429770634167.399, "dur": 24.326,
    "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": 870, "tid": 870,
    "ts": 7429770634176.163, "dur": 1.307,
    "args": {
      "External id": 6374,"Record function id": 0, "finished": true, "Ev Idx": 6373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770634192.262, "dur": 0.923,
    "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": 870, "tid": 870,
    "ts": 7429770634195.340, "dur": 8.823,
    "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": 870, "tid": 870,
    "ts": 7429770634210.254, "dur": 44.927,
    "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": 870, "tid": 870,
    "ts": 7429770634210.590, "dur": 2.302,
    "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": 870, "tid": 870,
    "ts": 7429770634211.573, "dur": 0.874,
    "args": {
      "External id": 6379,"Record function id": 0, "finished": true, "Ev Idx": 6378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770634211.981, "dur": 0.339,
    "args": {
      "External id": 6380,"Record function id": 0, "finished": true, "Ev Idx": 6379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770634213.197, "dur": 41.851,
    "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": 870, "tid": 870,
    "ts": 7429770634213.591, "dur": 1.409,
    "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": 870, "tid": 870,
    "ts": 7429770634213.959, "dur": 0.936,
    "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": 870, "tid": 870,
    "ts": 7429770634215.324, "dur": 37.919,
    "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": 870, "tid": 870,
    "ts": 7429770634224.323, "dur": 1.063,
    "args": {
      "External id": 6385,"Record function id": 0, "finished": true, "Ev Idx": 6384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770634253.832, "dur": 0.956,
    "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": 870, "tid": 870,
    "ts": 7429770634258.011, "dur": 9.742,
    "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": 870, "tid": 870,
    "ts": 7429770634286.449, "dur": 0.176,
    "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": 870, "tid": 870,
    "ts": 7429770634288.531, "dur": 10.514,
    "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": 870, "tid": 870,
    "ts": 7429770634290.096, "dur": 0.515,
    "args": {
      "External id": 6390,"Record function id": 0, "finished": true, "Ev Idx": 6389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770634290.927, "dur": 0.091,
    "args": {
      "External id": 6391,"Record function id": 0, "finished": true, "Ev Idx": 6390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770634301.598, "dur": 12.437,
    "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": 870, "tid": 870,
    "ts": 7429770634317.529, "dur": 7.875,
    "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": 870, "tid": 870,
    "ts": 7429770634327.166, "dur": 8.495,
    "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": 870, "tid": 870,
    "ts": 7429770634337.325, "dur": 7.937,
    "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": 870, "tid": 870,
    "ts": 7429770634347.628, "dur": 0.087,
    "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": 870, "tid": 870,
    "ts": 7429770634348.473, "dur": 8.363,
    "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": 870, "tid": 870,
    "ts": 7429770634375.106, "dur": 35.771,
    "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": 870, "tid": 870,
    "ts": 7429770634375.632, "dur": 3.004,
    "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": 870, "tid": 870,
    "ts": 7429770634376.604, "dur": 1.105,
    "args": {
      "External id": 6400,"Record function id": 0, "finished": true, "Ev Idx": 6399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770634377.089, "dur": 0.470,
    "args": {
      "External id": 6401,"Record function id": 0, "finished": true, "Ev Idx": 6400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770634378.993, "dur": 31.765,
    "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": 870, "tid": 870,
    "ts": 7429770634379.520, "dur": 2.596,
    "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": 870, "tid": 870,
    "ts": 7429770634379.807, "dur": 2.181,
    "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": 870, "tid": 870,
    "ts": 7429770634382.485, "dur": 26.413,
    "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": 870, "tid": 870,
    "ts": 7429770634392.377, "dur": 1.275,
    "args": {
      "External id": 6406,"Record function id": 0, "finished": true, "Ev Idx": 6405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770634409.453, "dur": 0.941,
    "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": 870, "tid": 870,
    "ts": 7429770634414.340, "dur": 1.308,
    "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": 870, "tid": 870,
    "ts": 7429770634417.314, "dur": 2.060,
    "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": 870, "tid": 870,
    "ts": 7429770634418.306, "dur": 0.453,
    "args": {
      "External id": 6410,"Record function id": 0, "finished": true, "Ev Idx": 6409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770634425.852, "dur": 31.015,
    "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": 870, "tid": 870,
    "ts": 7429770634426.217, "dur": 2.113,
    "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": 870, "tid": 870,
    "ts": 7429770634427.103, "dur": 0.755,
    "args": {
      "External id": 6413,"Record function id": 0, "finished": true, "Ev Idx": 6412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770634427.490, "dur": 0.274,
    "args": {
      "External id": 6414,"Record function id": 0, "finished": true, "Ev Idx": 6413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770634428.654, "dur": 28.115,
    "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": 870, "tid": 870,
    "ts": 7429770634429.035, "dur": 1.491,
    "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": 870, "tid": 870,
    "ts": 7429770634429.385, "dur": 1.013,
    "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": 870, "tid": 870,
    "ts": 7429770634430.906, "dur": 24.167,
    "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": 870, "tid": 870,
    "ts": 7429770634439.480, "dur": 1.030,
    "args": {
      "External id": 6419,"Record function id": 0, "finished": true, "Ev Idx": 6418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770634455.638, "dur": 0.887,
    "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": 870, "tid": 870,
    "ts": 7429770634459.072, "dur": 1.081,
    "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": 870, "tid": 870,
    "ts": 7429770634461.493, "dur": 1.609,
    "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": 870, "tid": 870,
    "ts": 7429770634462.277, "dur": 0.340,
    "args": {
      "External id": 6423,"Record function id": 0, "finished": true, "Ev Idx": 6422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770634469.045, "dur": 29.818,
    "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": 870, "tid": 870,
    "ts": 7429770634469.350, "dur": 1.799,
    "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": 870, "tid": 870,
    "ts": 7429770634470.061, "dur": 0.708,
    "args": {
      "External id": 6426,"Record function id": 0, "finished": true, "Ev Idx": 6425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770634470.427, "dur": 0.240,
    "args": {
      "External id": 6427,"Record function id": 0, "finished": true, "Ev Idx": 6426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770634471.440, "dur": 27.345,
    "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": 870, "tid": 870,
    "ts": 7429770634471.840, "dur": 1.194,
    "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": 870, "tid": 870,
    "ts": 7429770634472.084, "dur": 0.842,
    "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": 870, "tid": 870,
    "ts": 7429770634473.341, "dur": 23.838,
    "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": 870, "tid": 870,
    "ts": 7429770634482.955, "dur": 1.023,
    "args": {
      "External id": 6432,"Record function id": 0, "finished": true, "Ev Idx": 6431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770634497.692, "dur": 0.872,
    "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": 870, "tid": 870,
    "ts": 7429770634501.309, "dur": 1.131,
    "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": 870, "tid": 870,
    "ts": 7429770634503.800, "dur": 1.514,
    "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": 870, "tid": 870,
    "ts": 7429770634504.535, "dur": 0.315,
    "args": {
      "External id": 6436,"Record function id": 0, "finished": true, "Ev Idx": 6435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770634508.209, "dur": 2.310,
    "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": 870, "tid": 870,
    "ts": 7429770634509.587, "dur": 0.443,
    "args": {
      "External id": 6438,"Record function id": 0, "finished": true, "Ev Idx": 6437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770634511.731, "dur": 1.220,
    "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": 870, "tid": 870,
    "ts": 7429770634512.160, "dur": 0.474,
    "args": {
      "External id": 6440,"Record function id": 0, "finished": true, "Ev Idx": 6439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770634513.795, "dur": 9.056,
    "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": 870, "tid": 870,
    "ts": 7429770634527.689, "dur": 3.710,
    "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": 870, "tid": 870,
    "ts": 7429770634529.783, "dur": 0.490,
    "args": {
      "External id": 6443,"Record function id": 0, "finished": true, "Ev Idx": 6442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770634533.481, "dur": 3.060,
    "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": 870, "tid": 870,
    "ts": 7429770634534.702, "dur": 0.842,
    "args": {
      "External id": 6445,"Record function id": 0, "finished": true, "Ev Idx": 6444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770634538.097, "dur": 10.118,
    "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": 870, "tid": 870,
    "ts": 7429770634552.071, "dur": 18.326,
    "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": 870, "tid": 870,
    "ts": 7429770634573.513, "dur": 7.404,
    "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": 870, "tid": 870,
    "ts": 7429770634582.528, "dur": 9.294,
    "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": 870, "tid": 870,
    "ts": 7429770634593.821, "dur": 7.096,
    "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": 870, "tid": 870,
    "ts": 7429770634603.160, "dur": 2.115,
    "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": 870, "tid": 870,
    "ts": 7429770634604.224, "dur": 0.392,
    "args": {
      "External id": 6452,"Record function id": 0, "finished": true, "Ev Idx": 6451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770634606.859, "dur": 2.335,
    "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": 870, "tid": 870,
    "ts": 7429770634608.318, "dur": 0.339,
    "args": {
      "External id": 6454,"Record function id": 0, "finished": true, "Ev Idx": 6453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770634610.128, "dur": 6.895,
    "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": 870, "tid": 870,
    "ts": 7429770634618.463, "dur": 9.456,
    "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": 870, "tid": 870,
    "ts": 7429770634630.228, "dur": 7.197,
    "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": 870, "tid": 870,
    "ts": 7429770634638.662, "dur": 6.886,
    "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": 870, "tid": 870,
    "ts": 7429770634668.038, "dur": 3.073,
    "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": 870, "tid": 870,
    "ts": 7429770634669.914, "dur": 0.458,
    "args": {
      "External id": 6460,"Record function id": 0, "finished": true, "Ev Idx": 6459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770634671.796, "dur": 2.299,
    "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": 870, "tid": 870,
    "ts": 7429770634673.133, "dur": 0.396,
    "args": {
      "External id": 6462,"Record function id": 0, "finished": true, "Ev Idx": 6461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770634674.845, "dur": 2.313,
    "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": 870, "tid": 870,
    "ts": 7429770634675.891, "dur": 0.713,
    "args": {
      "External id": 6464,"Record function id": 0, "finished": true, "Ev Idx": 6463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770634677.807, "dur": 1.882,
    "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": 870, "tid": 870,
    "ts": 7429770634678.688, "dur": 0.481,
    "args": {
      "External id": 6466,"Record function id": 0, "finished": true, "Ev Idx": 6465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770634680.327, "dur": 1.750,
    "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": 870, "tid": 870,
    "ts": 7429770634681.224, "dur": 0.386,
    "args": {
      "External id": 6468,"Record function id": 0, "finished": true, "Ev Idx": 6467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770634684.717, "dur": 2.685,
    "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": 870, "tid": 870,
    "ts": 7429770634686.398, "dur": 0.362,
    "args": {
      "External id": 6470,"Record function id": 0, "finished": true, "Ev Idx": 6469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770634689.553, "dur": 20.146,
    "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": 870, "tid": 870,
    "ts": 7429770634691.945, "dur": 15.881,
    "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": 870, "tid": 870,
    "ts": 7429770634693.921, "dur": 2.634,
    "args": {
      "External id": 6473,"Record function id": 0, "finished": true, "Ev Idx": 6472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770634694.634, "dur": 1.683,
    "args": {
      "External id": 6474,"Record function id": 0, "finished": true, "Ev Idx": 6473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770634697.374, "dur": 10.038,
    "args": {
      "External id": 6475,"Record function id": 0, "finished": true, "Ev Idx": 6474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770634708.351, "dur": 1.116,
    "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": 870, "tid": 870,
    "ts": 7429770634712.773, "dur": 1.751,
    "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": 870, "tid": 870,
    "ts": 7429770634713.615, "dur": 0.366,
    "args": {
      "External id": 6478,"Record function id": 0, "finished": true, "Ev Idx": 6477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770634715.195, "dur": 1.505,
    "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": 870, "tid": 870,
    "ts": 7429770634715.967, "dur": 0.262,
    "args": {
      "External id": 6480,"Record function id": 0, "finished": true, "Ev Idx": 6479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770634717.174, "dur": 1.407,
    "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": 870, "tid": 870,
    "ts": 7429770634717.808, "dur": 0.332,
    "args": {
      "External id": 6482,"Record function id": 0, "finished": true, "Ev Idx": 6481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770634719.140, "dur": 2.238,
    "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": 870, "tid": 870,
    "ts": 7429770634720.533, "dur": 0.310,
    "args": {
      "External id": 6484,"Record function id": 0, "finished": true, "Ev Idx": 6483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770634721.951, "dur": 2.200,
    "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": 870, "tid": 870,
    "ts": 7429770634723.393, "dur": 0.288,
    "args": {
      "External id": 6486,"Record function id": 0, "finished": true, "Ev Idx": 6485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770634725.667, "dur": 1.751,
    "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": 870, "tid": 870,
    "ts": 7429770634726.640, "dur": 0.349,
    "args": {
      "External id": 6488,"Record function id": 0, "finished": true, "Ev Idx": 6487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770634728.848, "dur": 10.889,
    "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": 870, "tid": 870,
    "ts": 7429770634729.723, "dur": 8.233,
    "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": 870, "tid": 870,
    "ts": 7429770634730.358, "dur": 1.710,
    "args": {
      "External id": 6491,"Record function id": 0, "finished": true, "Ev Idx": 6490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770634730.664, "dur": 1.272,
    "args": {
      "External id": 6492,"Record function id": 0, "finished": true, "Ev Idx": 6491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770634732.406, "dur": 5.244,
    "args": {
      "External id": 6493,"Record function id": 0, "finished": true, "Ev Idx": 6492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770634738.328, "dur": 1.270,
    "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": 870, "tid": 870,
    "ts": 7429770634741.727, "dur": 2.233,
    "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": 870, "tid": 870,
    "ts": 7429770634743.151, "dur": 0.343,
    "args": {
      "External id": 6496,"Record function id": 0, "finished": true, "Ev Idx": 6495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770634745.513, "dur": 110.005,
    "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": 870, "tid": 870,
    "ts": 7429770634746.000, "dur": 2.070,
    "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": 870, "tid": 870,
    "ts": 7429770634747.331, "dur": 0.314,
    "args": {
      "External id": 6499,"Record function id": 0, "finished": true, "Ev Idx": 6498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770634748.417, "dur": 3.675,
    "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": 870, "tid": 870,
    "ts": 7429770634750.083, "dur": 1.883,
    "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": 870, "tid": 870,
    "ts": 7429770634752.528, "dur": 1.554,
    "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": 870, "tid": 870,
    "ts": 7429770634753.416, "dur": 0.290,
    "args": {
      "External id": 6503,"Record function id": 0, "finished": true, "Ev Idx": 6502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770634754.458, "dur": 1.727,
    "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": 870, "tid": 870,
    "ts": 7429770634755.229, "dur": 0.849,
    "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": 870, "tid": 870,
    "ts": 7429770634756.773, "dur": 96.376,
    "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": 870, "tid": 870,
    "ts": 7429770634831.186, "dur": 1.489,
    "args": {
      "External id": 6507,"Record function id": 0, "finished": true, "Ev Idx": 6506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770634853.808, "dur": 1.348,
    "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": 870, "tid": 870,
    "ts": 7429770634859.865, "dur": 10.564,
    "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": 870, "tid": 870,
    "ts": 7429770634873.743, "dur": 2.487,
    "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": 870, "tid": 870,
    "ts": 7429770634874.651, "dur": 0.625,
    "args": {
      "External id": 6511,"Record function id": 0, "finished": true, "Ev Idx": 6510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770634876.971, "dur": 1.480,
    "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": 870, "tid": 870,
    "ts": 7429770634877.604, "dur": 0.345,
    "args": {
      "External id": 6513,"Record function id": 0, "finished": true, "Ev Idx": 6512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770634879.044, "dur": 1.338,
    "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": 870, "tid": 870,
    "ts": 7429770634879.775, "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": 870, "tid": 870,
    "ts": 7429770634882.222, "dur": 7.722,
    "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": 870, "tid": 870,
    "ts": 7429770634895.585, "dur": 14.837,
    "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": 870, "tid": 870,
    "ts": 7429770634896.799, "dur": 0.494,
    "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": 870, "tid": 870,
    "ts": 7429770634898.082, "dur": 12.071,
    "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": 870, "tid": 870,
    "ts": 7429770634913.607, "dur": 0.353,
    "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": 870, "tid": 870,
    "ts": 7429770634925.698, "dur": 0.483,
    "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": 870, "tid": 870,
    "ts": 7429770634928.234, "dur": 42.693,
    "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": 870, "tid": 870,
    "ts": 7429770634929.003, "dur": 2.175,
    "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": 870, "tid": 870,
    "ts": 7429770634930.162, "dur": 0.403,
    "args": {
      "External id": 6524,"Record function id": 0, "finished": true, "Ev Idx": 6523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770634931.550, "dur": 2.635,
    "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": 870, "tid": 870,
    "ts": 7429770634932.120, "dur": 1.943,
    "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": 870, "tid": 870,
    "ts": 7429770634934.627, "dur": 2.450,
    "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": 870, "tid": 870,
    "ts": 7429770634936.412, "dur": 0.273,
    "args": {
      "External id": 6528,"Record function id": 0, "finished": true, "Ev Idx": 6527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770634937.407, "dur": 1.566,
    "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": 870, "tid": 870,
    "ts": 7429770634937.743, "dur": 1.145,
    "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": 870, "tid": 870,
    "ts": 7429770634939.582, "dur": 29.388,
    "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": 870, "tid": 870,
    "ts": 7429770634950.071, "dur": 1.347,
    "args": {
      "External id": 6532,"Record function id": 0, "finished": true, "Ev Idx": 6531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770634969.577, "dur": 1.035,
    "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": 870, "tid": 870,
    "ts": 7429770634973.115, "dur": 2.192,
    "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": 870, "tid": 870,
    "ts": 7429770634974.330, "dur": 0.393,
    "args": {
      "External id": 6535,"Record function id": 0, "finished": true, "Ev Idx": 6534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770634977.800, "dur": 12.872,
    "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": 870, "tid": 870,
    "ts": 7429770634978.684, "dur": 11.852,
    "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": 870, "tid": 870,
    "ts": 7429770634979.787, "dur": 2.106,
    "args": {
      "External id": 6538,"Record function id": 0, "finished": true, "Ev Idx": 6537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770634980.266, "dur": 1.456,
    "args": {
      "External id": 6539,"Record function id": 0, "finished": true, "Ev Idx": 6538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770634982.406, "dur": 7.664,
    "args": {
      "External id": 6540,"Record function id": 0, "finished": true, "Ev Idx": 6539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770634997.015, "dur": 1.708,
    "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": 870, "tid": 870,
    "ts": 7429770634997.391, "dur": 1.211,
    "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": 870, "tid": 870,
    "ts": 7429770635014.745, "dur": 39.996,
    "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": 870, "tid": 870,
    "ts": 7429770635015.443, "dur": 3.109,
    "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": 870, "tid": 870,
    "ts": 7429770635016.914, "dur": 1.012,
    "args": {
      "External id": 6545,"Record function id": 0, "finished": true, "Ev Idx": 6544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770635017.344, "dur": 0.393,
    "args": {
      "External id": 6546,"Record function id": 0, "finished": true, "Ev Idx": 6545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770635019.013, "dur": 35.551,
    "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": 870, "tid": 870,
    "ts": 7429770635019.455, "dur": 1.763,
    "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": 870, "tid": 870,
    "ts": 7429770635019.769, "dur": 1.358,
    "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": 870, "tid": 870,
    "ts": 7429770635021.684, "dur": 30.723,
    "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": 870, "tid": 870,
    "ts": 7429770635033.632, "dur": 1.226,
    "args": {
      "External id": 6551,"Record function id": 0, "finished": true, "Ev Idx": 6550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770635053.133, "dur": 1.060,
    "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": 870, "tid": 870,
    "ts": 7429770635060.093, "dur": 8.847,
    "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": 870, "tid": 870,
    "ts": 7429770635078.668, "dur": 0.197,
    "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": 870, "tid": 870,
    "ts": 7429770635081.300, "dur": 12.594,
    "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": 870, "tid": 870,
    "ts": 7429770635083.233, "dur": 0.370,
    "args": {
      "External id": 6556,"Record function id": 0, "finished": true, "Ev Idx": 6555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770635084.046, "dur": 0.117,
    "args": {
      "External id": 6557,"Record function id": 0, "finished": true, "Ev Idx": 6556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770635096.614, "dur": 13.235,
    "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": 870, "tid": 870,
    "ts": 7429770635113.276, "dur": 9.944,
    "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": 870, "tid": 870,
    "ts": 7429770635125.244, "dur": 9.301,
    "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": 870, "tid": 870,
    "ts": 7429770635136.305, "dur": 10.001,
    "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": 870, "tid": 870,
    "ts": 7429770635148.942, "dur": 0.223,
    "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": 870, "tid": 870,
    "ts": 7429770635150.010, "dur": 8.496,
    "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": 870, "tid": 870,
    "ts": 7429770635173.148, "dur": 137.243,
    "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": 870, "tid": 870,
    "ts": 7429770635173.600, "dur": 3.430,
    "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": 870, "tid": 870,
    "ts": 7429770635175.036, "dur": 1.376,
    "args": {
      "External id": 6566,"Record function id": 0, "finished": true, "Ev Idx": 6565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770635175.662, "dur": 0.596,
    "args": {
      "External id": 6567,"Record function id": 0, "finished": true, "Ev Idx": 6566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770635177.417, "dur": 132.776,
    "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": 870, "tid": 870,
    "ts": 7429770635178.073, "dur": 1.541,
    "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": 870, "tid": 870,
    "ts": 7429770635178.385, "dur": 1.135,
    "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": 870, "tid": 870,
    "ts": 7429770635179.981, "dur": 127.228,
    "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": 870, "tid": 870,
    "ts": 7429770635191.604, "dur": 95.302,
    "args": {
      "External id": 6572,"Record function id": 0, "finished": true, "Ev Idx": 6571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770635308.289, "dur": 1.461,
    "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": 870, "tid": 870,
    "ts": 7429770635320.667, "dur": 11.144,
    "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": 870, "tid": 870,
    "ts": 7429770635339.550, "dur": 36.986,
    "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": 870, "tid": 870,
    "ts": 7429770635340.039, "dur": 3.004,
    "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": 870, "tid": 870,
    "ts": 7429770635341.176, "dur": 1.299,
    "args": {
      "External id": 6577,"Record function id": 0, "finished": true, "Ev Idx": 6576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770635341.793, "dur": 0.487,
    "args": {
      "External id": 6578,"Record function id": 0, "finished": true, "Ev Idx": 6577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770635343.438, "dur": 32.969,
    "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": 870, "tid": 870,
    "ts": 7429770635343.895, "dur": 1.596,
    "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": 870, "tid": 870,
    "ts": 7429770635344.230, "dur": 1.148,
    "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": 870, "tid": 870,
    "ts": 7429770635345.975, "dur": 28.467,
    "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": 870, "tid": 870,
    "ts": 7429770635356.983, "dur": 1.426,
    "args": {
      "External id": 6583,"Record function id": 0, "finished": true, "Ev Idx": 6582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770635375.071, "dur": 1.067,
    "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": 870, "tid": 870,
    "ts": 7429770635378.569, "dur": 8.842,
    "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": 870, "tid": 870,
    "ts": 7429770635393.037, "dur": 45.304,
    "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": 870, "tid": 870,
    "ts": 7429770635393.441, "dur": 2.372,
    "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": 870, "tid": 870,
    "ts": 7429770635394.467, "dur": 0.903,
    "args": {
      "External id": 6588,"Record function id": 0, "finished": true, "Ev Idx": 6587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770635394.893, "dur": 0.347,
    "args": {
      "External id": 6589,"Record function id": 0, "finished": true, "Ev Idx": 6588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770635396.097, "dur": 42.146,
    "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": 870, "tid": 870,
    "ts": 7429770635396.498, "dur": 1.269,
    "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": 870, "tid": 870,
    "ts": 7429770635396.781, "dur": 0.887,
    "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": 870, "tid": 870,
    "ts": 7429770635398.071, "dur": 38.376,
    "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": 870, "tid": 870,
    "ts": 7429770635407.452, "dur": 1.130,
    "args": {
      "External id": 6594,"Record function id": 0, "finished": true, "Ev Idx": 6593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770635437.041, "dur": 0.932,
    "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": 870, "tid": 870,
    "ts": 7429770635441.238, "dur": 7.702,
    "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": 870, "tid": 870,
    "ts": 7429770635468.958, "dur": 0.202,
    "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": 870, "tid": 870,
    "ts": 7429770635471.329, "dur": 10.087,
    "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": 870, "tid": 870,
    "ts": 7429770635472.949, "dur": 0.350,
    "args": {
      "External id": 6599,"Record function id": 0, "finished": true, "Ev Idx": 6598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770635473.636, "dur": 0.097,
    "args": {
      "External id": 6600,"Record function id": 0, "finished": true, "Ev Idx": 6599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770635484.036, "dur": 13.321,
    "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": 870, "tid": 870,
    "ts": 7429770635500.793, "dur": 7.923,
    "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": 870, "tid": 870,
    "ts": 7429770635510.561, "dur": 8.361,
    "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": 870, "tid": 870,
    "ts": 7429770635520.523, "dur": 8.072,
    "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": 870, "tid": 870,
    "ts": 7429770635531.074, "dur": 0.202,
    "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": 870, "tid": 870,
    "ts": 7429770635532.108, "dur": 7.979,
    "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": 870, "tid": 870,
    "ts": 7429770635558.424, "dur": 37.065,
    "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": 870, "tid": 870,
    "ts": 7429770635558.967, "dur": 3.694,
    "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": 870, "tid": 870,
    "ts": 7429770635560.795, "dur": 1.012,
    "args": {
      "External id": 6609,"Record function id": 0, "finished": true, "Ev Idx": 6608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770635561.268, "dur": 0.405,
    "args": {
      "External id": 6610,"Record function id": 0, "finished": true, "Ev Idx": 6609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770635563.078, "dur": 32.282,
    "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": 870, "tid": 870,
    "ts": 7429770635563.755, "dur": 1.593,
    "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": 870, "tid": 870,
    "ts": 7429770635564.056, "dur": 1.179,
    "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": 870, "tid": 870,
    "ts": 7429770635565.723, "dur": 27.776,
    "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": 870, "tid": 870,
    "ts": 7429770635575.353, "dur": 1.202,
    "args": {
      "External id": 6615,"Record function id": 0, "finished": true, "Ev Idx": 6614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770635594.054, "dur": 0.987,
    "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": 870, "tid": 870,
    "ts": 7429770635598.897, "dur": 1.450,
    "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": 870, "tid": 870,
    "ts": 7429770635602.222, "dur": 2.356,
    "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": 870, "tid": 870,
    "ts": 7429770635603.309, "dur": 0.517,
    "args": {
      "External id": 6619,"Record function id": 0, "finished": true, "Ev Idx": 6618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770635611.287, "dur": 31.725,
    "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": 870, "tid": 870,
    "ts": 7429770635611.605, "dur": 2.152,
    "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": 870, "tid": 870,
    "ts": 7429770635612.409, "dur": 0.904,
    "args": {
      "External id": 6622,"Record function id": 0, "finished": true, "Ev Idx": 6621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770635612.811, "dur": 0.385,
    "args": {
      "External id": 6623,"Record function id": 0, "finished": true, "Ev Idx": 6622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770635614.069, "dur": 28.844,
    "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": 870, "tid": 870,
    "ts": 7429770635614.498, "dur": 1.761,
    "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": 870, "tid": 870,
    "ts": 7429770635614.752, "dur": 1.401,
    "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": 870, "tid": 870,
    "ts": 7429770635616.622, "dur": 24.565,
    "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": 870, "tid": 870,
    "ts": 7429770635624.857, "dur": 1.073,
    "args": {
      "External id": 6628,"Record function id": 0, "finished": true, "Ev Idx": 6627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770635641.743, "dur": 0.921,
    "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": 870, "tid": 870,
    "ts": 7429770635645.340, "dur": 1.144,
    "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": 870, "tid": 870,
    "ts": 7429770635647.839, "dur": 1.621,
    "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": 870, "tid": 870,
    "ts": 7429770635648.616, "dur": 0.363,
    "args": {
      "External id": 6632,"Record function id": 0, "finished": true, "Ev Idx": 6631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770635656.483, "dur": 30.663,
    "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": 870, "tid": 870,
    "ts": 7429770635656.760, "dur": 1.944,
    "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": 870, "tid": 870,
    "ts": 7429770635657.471, "dur": 0.849,
    "args": {
      "External id": 6635,"Record function id": 0, "finished": true, "Ev Idx": 6634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770635657.955, "dur": 0.243,
    "args": {
      "External id": 6636,"Record function id": 0, "finished": true, "Ev Idx": 6635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770635659.037, "dur": 28.011,
    "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": 870, "tid": 870,
    "ts": 7429770635659.501, "dur": 1.337,
    "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": 870, "tid": 870,
    "ts": 7429770635659.849, "dur": 0.877,
    "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": 870, "tid": 870,
    "ts": 7429770635661.228, "dur": 24.228,
    "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": 870, "tid": 870,
    "ts": 7429770635670.578, "dur": 1.065,
    "args": {
      "External id": 6641,"Record function id": 0, "finished": true, "Ev Idx": 6640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770635685.977, "dur": 0.845,
    "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": 870, "tid": 870,
    "ts": 7429770635689.427, "dur": 1.156,
    "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": 870, "tid": 870,
    "ts": 7429770635691.916, "dur": 1.758,
    "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": 870, "tid": 870,
    "ts": 7429770635692.662, "dur": 0.439,
    "args": {
      "External id": 6645,"Record function id": 0, "finished": true, "Ev Idx": 6644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770635696.807, "dur": 2.333,
    "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": 870, "tid": 870,
    "ts": 7429770635698.268, "dur": 0.347,
    "args": {
      "External id": 6647,"Record function id": 0, "finished": true, "Ev Idx": 6646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770635700.342, "dur": 1.483,
    "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": 870, "tid": 870,
    "ts": 7429770635700.851, "dur": 0.346,
    "args": {
      "External id": 6649,"Record function id": 0, "finished": true, "Ev Idx": 6648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 870, "tid": 870,
    "ts": 7429770635702.718, "dur": 10.589,
    "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": 870, "tid": 870,
    "ts": 7429770635718.815, "dur": 4.196,
    "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": 870, "tid": 870,
    "ts": 7429770635721.075, "dur": 0.556,
    "args": {
      "External id": 6652,"Record function id": 0, "finished": true, "Ev Idx": 6651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770635725.506, "dur": 3.011,
    "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": 870, "tid": 870,
    "ts": 7429770635726.985, "dur": 0.683,
    "args": {
      "External id": 6654,"Record function id": 0, "finished": true, "Ev Idx": 6653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770635730.476, "dur": 10.343,
    "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": 870, "tid": 870,
    "ts": 7429770635745.537, "dur": 19.610,
    "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": 870, "tid": 870,
    "ts": 7429770635792.845, "dur": 9.842,
    "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": 870, "tid": 870,
    "ts": 7429770635804.477, "dur": 8.871,
    "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": 870, "tid": 870,
    "ts": 7429770635815.431, "dur": 7.667,
    "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": 870, "tid": 870,
    "ts": 7429770635825.598, "dur": 2.297,
    "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": 870, "tid": 870,
    "ts": 7429770635826.727, "dur": 0.410,
    "args": {
      "External id": 6661,"Record function id": 0, "finished": true, "Ev Idx": 6660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770635829.514, "dur": 1.705,
    "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": 870, "tid": 870,
    "ts": 7429770635830.377, "dur": 0.293,
    "args": {
      "External id": 6663,"Record function id": 0, "finished": true, "Ev Idx": 6662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::neg", "pid": 870, "tid": 870,
    "ts": 7429770635832.217, "dur": 7.185,
    "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": 870, "tid": 870,
    "ts": 7429770635841.384, "dur": 9.087,
    "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": 870, "tid": 870,
    "ts": 7429770635852.704, "dur": 7.453,
    "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": 870, "tid": 870,
    "ts": 7429770635861.445, "dur": 6.751,
    "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": 870, "tid": 870,
    "ts": 7429770635892.811, "dur": 2.590,
    "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": 870, "tid": 870,
    "ts": 7429770635894.135, "dur": 0.442,
    "args": {
      "External id": 6669,"Record function id": 0, "finished": true, "Ev Idx": 6668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770635896.050, "dur": 1.680,
    "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": 870, "tid": 870,
    "ts": 7429770635896.864, "dur": 0.336,
    "args": {
      "External id": 6671,"Record function id": 0, "finished": true, "Ev Idx": 6670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770635898.437, "dur": 3.058,
    "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": 870, "tid": 870,
    "ts": 7429770635900.059, "dur": 0.768,
    "args": {
      "External id": 6673,"Record function id": 0, "finished": true, "Ev Idx": 6672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770635902.202, "dur": 2.607,
    "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": 870, "tid": 870,
    "ts": 7429770635903.924, "dur": 0.362,
    "args": {
      "External id": 6675,"Record function id": 0, "finished": true, "Ev Idx": 6674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770635905.422, "dur": 2.084,
    "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": 870, "tid": 870,
    "ts": 7429770635906.743, "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": 870, "tid": 870,
    "ts": 7429770635910.352, "dur": 3.154,
    "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": 870, "tid": 870,
    "ts": 7429770635912.377, "dur": 0.405,
    "args": {
      "External id": 6679,"Record function id": 0, "finished": true, "Ev Idx": 6678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770635915.970, "dur": 19.684,
    "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": 870, "tid": 870,
    "ts": 7429770635918.550, "dur": 15.126,
    "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": 870, "tid": 870,
    "ts": 7429770635919.948, "dur": 2.887,
    "args": {
      "External id": 6682,"Record function id": 0, "finished": true, "Ev Idx": 6681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770635920.794, "dur": 1.777,
    "args": {
      "External id": 6683,"Record function id": 0, "finished": true, "Ev Idx": 6682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770635923.588, "dur": 9.497,
    "args": {
      "External id": 6684,"Record function id": 0, "finished": true, "Ev Idx": 6683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770635934.161, "dur": 1.238,
    "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": 870, "tid": 870,
    "ts": 7429770635938.815, "dur": 2.456,
    "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": 870, "tid": 870,
    "ts": 7429770635940.317, "dur": 0.374,
    "args": {
      "External id": 6687,"Record function id": 0, "finished": true, "Ev Idx": 6686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770635941.907, "dur": 1.898,
    "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": 870, "tid": 870,
    "ts": 7429770635943.041, "dur": 0.286,
    "args": {
      "External id": 6689,"Record function id": 0, "finished": true, "Ev Idx": 6688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 870, "tid": 870,
    "ts": 7429770635944.326, "dur": 1.445,
    "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": 870, "tid": 870,
    "ts": 7429770635944.978, "dur": 0.356,
    "args": {
      "External id": 6691,"Record function id": 0, "finished": true, "Ev Idx": 6690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770635946.363, "dur": 1.562,
    "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": 870, "tid": 870,
    "ts": 7429770635947.119, "dur": 0.311,
    "args": {
      "External id": 6693,"Record function id": 0, "finished": true, "Ev Idx": 6692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770635948.502, "dur": 1.440,
    "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": 870, "tid": 870,
    "ts": 7429770635949.181, "dur": 0.305,
    "args": {
      "External id": 6695,"Record function id": 0, "finished": true, "Ev Idx": 6694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 870, "tid": 870,
    "ts": 7429770635951.419, "dur": 1.712,
    "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": 870, "tid": 870,
    "ts": 7429770635952.258, "dur": 0.428,
    "args": {
      "External id": 6697,"Record function id": 0, "finished": true, "Ev Idx": 6696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770635954.564, "dur": 12.422,
    "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": 870, "tid": 870,
    "ts": 7429770635955.439, "dur": 10.034,
    "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": 870, "tid": 870,
    "ts": 7429770635956.819, "dur": 1.712,
    "args": {
      "External id": 6700,"Record function id": 0, "finished": true, "Ev Idx": 6699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770635957.147, "dur": 1.254,
    "args": {
      "External id": 6701,"Record function id": 0, "finished": true, "Ev Idx": 6700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770635958.850, "dur": 6.318,
    "args": {
      "External id": 6702,"Record function id": 0, "finished": true, "Ev Idx": 6701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770635965.853, "dur": 0.981,
    "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": 870, "tid": 870,
    "ts": 7429770635968.977, "dur": 1.906,
    "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": 870, "tid": 870,
    "ts": 7429770635969.913, "dur": 0.405,
    "args": {
      "External id": 6705,"Record function id": 0, "finished": true, "Ev Idx": 6704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770635972.362, "dur": 104.135,
    "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": 870, "tid": 870,
    "ts": 7429770635973.029, "dur": 1.572,
    "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": 870, "tid": 870,
    "ts": 7429770635973.799, "dur": 0.342,
    "args": {
      "External id": 6708,"Record function id": 0, "finished": true, "Ev Idx": 6707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770635974.948, "dur": 3.985,
    "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": 870, "tid": 870,
    "ts": 7429770635976.841, "dur": 1.953,
    "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": 870, "tid": 870,
    "ts": 7429770635979.400, "dur": 1.821,
    "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": 870, "tid": 870,
    "ts": 7429770635980.556, "dur": 0.294,
    "args": {
      "External id": 6712,"Record function id": 0, "finished": true, "Ev Idx": 6711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770635981.503, "dur": 2.337,
    "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": 870, "tid": 870,
    "ts": 7429770635982.316, "dur": 1.422,
    "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": 870, "tid": 870,
    "ts": 7429770635984.529, "dur": 89.705,
    "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": 870, "tid": 870,
    "ts": 7429770636052.558, "dur": 1.411,
    "args": {
      "External id": 6716,"Record function id": 0, "finished": true, "Ev Idx": 6715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770636074.913, "dur": 1.258,
    "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": 870, "tid": 870,
    "ts": 7429770636081.022, "dur": 10.254,
    "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": 870, "tid": 870,
    "ts": 7429770636094.648, "dur": 2.143,
    "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": 870, "tid": 870,
    "ts": 7429770636095.528, "dur": 0.448,
    "args": {
      "External id": 6720,"Record function id": 0, "finished": true, "Ev Idx": 6719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770636097.461, "dur": 1.375,
    "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": 870, "tid": 870,
    "ts": 7429770636098.113, "dur": 0.280,
    "args": {
      "External id": 6722,"Record function id": 0, "finished": true, "Ev Idx": 6721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770636099.520, "dur": 0.978,
    "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": 870, "tid": 870,
    "ts": 7429770636099.978, "dur": 0.201,
    "args": {
      "External id": 6724,"Record function id": 0, "finished": true, "Ev Idx": 6723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 870, "tid": 870,
    "ts": 7429770636102.145, "dur": 7.829,
    "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": 870, "tid": 870,
    "ts": 7429770636115.495, "dur": 14.502,
    "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": 870, "tid": 870,
    "ts": 7429770636116.489, "dur": 0.361,
    "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": 870, "tid": 870,
    "ts": 7429770636117.562, "dur": 12.180,
    "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": 870, "tid": 870,
    "ts": 7429770636133.075, "dur": 0.204,
    "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": 870, "tid": 870,
    "ts": 7429770636144.439, "dur": 0.522,
    "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": 870, "tid": 870,
    "ts": 7429770636146.880, "dur": 42.273,
    "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": 870, "tid": 870,
    "ts": 7429770636147.569, "dur": 2.233,
    "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": 870, "tid": 870,
    "ts": 7429770636148.903, "dur": 0.316,
    "args": {
      "External id": 6733,"Record function id": 0, "finished": true, "Ev Idx": 6732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770636150.344, "dur": 3.139,
    "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": 870, "tid": 870,
    "ts": 7429770636150.731, "dur": 2.631,
    "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": 870, "tid": 870,
    "ts": 7429770636153.946, "dur": 1.616,
    "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": 870, "tid": 870,
    "ts": 7429770636154.907, "dur": 0.276,
    "args": {
      "External id": 6737,"Record function id": 0, "finished": true, "Ev Idx": 6736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770636155.896, "dur": 1.932,
    "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": 870, "tid": 870,
    "ts": 7429770636156.212, "dur": 1.533,
    "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": 870, "tid": 870,
    "ts": 7429770636158.442, "dur": 28.816,
    "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": 870, "tid": 870,
    "ts": 7429770636168.108, "dur": 1.244,
    "args": {
      "External id": 6741,"Record function id": 0, "finished": true, "Ev Idx": 6740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770636187.885, "dur": 0.978,
    "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": 870, "tid": 870,
    "ts": 7429770636191.212, "dur": 1.985,
    "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": 870, "tid": 870,
    "ts": 7429770636192.288, "dur": 0.325,
    "args": {
      "External id": 6744,"Record function id": 0, "finished": true, "Ev Idx": 6743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 870, "tid": 870,
    "ts": 7429770636195.871, "dur": 11.871,
    "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": 870, "tid": 870,
    "ts": 7429770636196.467, "dur": 11.128,
    "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": 870, "tid": 870,
    "ts": 7429770636197.422, "dur": 1.922,
    "args": {
      "External id": 6747,"Record function id": 0, "finished": true, "Ev Idx": 6746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 870, "tid": 870,
    "ts": 7429770636197.862, "dur": 1.317,
    "args": {
      "External id": 6748,"Record function id": 0, "finished": true, "Ev Idx": 6747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 870, "tid": 870,
    "ts": 7429770636199.766, "dur": 7.311,
    "args": {
      "External id": 6749,"Record function id": 0, "finished": true, "Ev Idx": 6748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 870, "tid": 870,
    "ts": 7429770636213.777, "dur": 1.510,
    "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": 870, "tid": 870,
    "ts": 7429770636214.052, "dur": 1.137,
    "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": 870, "tid": 870,
    "ts": 7429770636230.777, "dur": 35.141,
    "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": 870, "tid": 870,
    "ts": 7429770636231.446, "dur": 3.586,
    "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": 870, "tid": 870,
    "ts": 7429770636233.412, "dur": 1.052,
    "args": {
      "External id": 6754,"Record function id": 0, "finished": true, "Ev Idx": 6753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770636233.975, "dur": 0.358,
    "args": {
      "External id": 6755,"Record function id": 0, "finished": true, "Ev Idx": 6754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770636235.444, "dur": 30.294,
    "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": 870, "tid": 870,
    "ts": 7429770636235.933, "dur": 1.690,
    "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": 870, "tid": 870,
    "ts": 7429770636236.166, "dur": 1.366,
    "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": 870, "tid": 870,
    "ts": 7429770636238.141, "dur": 25.674,
    "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": 870, "tid": 870,
    "ts": 7429770636248.324, "dur": 1.221,
    "args": {
      "External id": 6760,"Record function id": 0, "finished": true, "Ev Idx": 6759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770636264.414, "dur": 1.024,
    "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": 870, "tid": 870,
    "ts": 7429770636271.634, "dur": 9.455,
    "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": 870, "tid": 870,
    "ts": 7429770636290.209, "dur": 0.192,
    "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": 870, "tid": 870,
    "ts": 7429770636292.722, "dur": 11.806,
    "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": 870, "tid": 870,
    "ts": 7429770636294.612, "dur": 0.373,
    "args": {
      "External id": 6765,"Record function id": 0, "finished": true, "Ev Idx": 6764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770636295.491, "dur": 0.074,
    "args": {
      "External id": 6766,"Record function id": 0, "finished": true, "Ev Idx": 6765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770636307.202, "dur": 13.465,
    "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": 870, "tid": 870,
    "ts": 7429770636324.157, "dur": 9.996,
    "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": 870, "tid": 870,
    "ts": 7429770636336.092, "dur": 9.836,
    "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": 870, "tid": 870,
    "ts": 7429770636347.676, "dur": 9.948,
    "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": 870, "tid": 870,
    "ts": 7429770636360.124, "dur": 0.079,
    "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": 870, "tid": 870,
    "ts": 7429770636361.045, "dur": 7.993,
    "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": 870, "tid": 870,
    "ts": 7429770636383.154, "dur": 35.973,
    "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": 870, "tid": 870,
    "ts": 7429770636383.631, "dur": 3.375,
    "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": 870, "tid": 870,
    "ts": 7429770636385.222, "dur": 1.168,
    "args": {
      "External id": 6775,"Record function id": 0, "finished": true, "Ev Idx": 6774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770636385.769, "dur": 0.451,
    "args": {
      "External id": 6776,"Record function id": 0, "finished": true, "Ev Idx": 6775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770636387.351, "dur": 31.633,
    "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": 870, "tid": 870,
    "ts": 7429770636387.887, "dur": 1.548,
    "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": 870, "tid": 870,
    "ts": 7429770636388.198, "dur": 1.143,
    "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": 870, "tid": 870,
    "ts": 7429770636389.777, "dur": 27.349,
    "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": 870, "tid": 870,
    "ts": 7429770636401.013, "dur": 1.219,
    "args": {
      "External id": 6781,"Record function id": 0, "finished": true, "Ev Idx": 6780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770636417.743, "dur": 0.958,
    "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": 870, "tid": 870,
    "ts": 7429770636426.773, "dur": 10.372,
    "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": 870, "tid": 870,
    "ts": 7429770636444.650, "dur": 35.206,
    "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": 870, "tid": 870,
    "ts": 7429770636445.016, "dur": 2.336,
    "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": 870, "tid": 870,
    "ts": 7429770636445.943, "dur": 0.915,
    "args": {
      "External id": 6786,"Record function id": 0, "finished": true, "Ev Idx": 6785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770636446.327, "dur": 0.401,
    "args": {
      "External id": 6787,"Record function id": 0, "finished": true, "Ev Idx": 6786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770636447.720, "dur": 32.036,
    "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": 870, "tid": 870,
    "ts": 7429770636448.235, "dur": 1.448,
    "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": 870, "tid": 870,
    "ts": 7429770636448.548, "dur": 1.025,
    "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": 870, "tid": 870,
    "ts": 7429770636450.000, "dur": 27.889,
    "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": 870, "tid": 870,
    "ts": 7429770636462.236, "dur": 1.139,
    "args": {
      "External id": 6792,"Record function id": 0, "finished": true, "Ev Idx": 6791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770636478.477, "dur": 1.034,
    "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": 870, "tid": 870,
    "ts": 7429770636481.512, "dur": 7.935,
    "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": 870, "tid": 870,
    "ts": 7429770636495.256, "dur": 47.250,
    "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": 870, "tid": 870,
    "ts": 7429770636495.735, "dur": 2.255,
    "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": 870, "tid": 870,
    "ts": 7429770636496.530, "dur": 0.957,
    "args": {
      "External id": 6797,"Record function id": 0, "finished": true, "Ev Idx": 6796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770636496.933, "dur": 0.423,
    "args": {
      "External id": 6798,"Record function id": 0, "finished": true, "Ev Idx": 6797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770636498.268, "dur": 44.136,
    "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": 870, "tid": 870,
    "ts": 7429770636498.695, "dur": 1.875,
    "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": 870, "tid": 870,
    "ts": 7429770636498.959, "dur": 1.507,
    "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": 870, "tid": 870,
    "ts": 7429770636500.875, "dur": 39.798,
    "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": 870, "tid": 870,
    "ts": 7429770636509.472, "dur": 1.153,
    "args": {
      "External id": 6803,"Record function id": 0, "finished": true, "Ev Idx": 6802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770636541.249, "dur": 0.897,
    "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": 870, "tid": 870,
    "ts": 7429770636545.086, "dur": 7.538,
    "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": 870, "tid": 870,
    "ts": 7429770636568.905, "dur": 0.158,
    "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": 870, "tid": 870,
    "ts": 7429770636571.086, "dur": 8.520,
    "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": 870, "tid": 870,
    "ts": 7429770636572.241, "dur": 0.273,
    "args": {
      "External id": 6808,"Record function id": 0, "finished": true, "Ev Idx": 6807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 870, "tid": 870,
    "ts": 7429770636572.831, "dur": 0.070,
    "args": {
      "External id": 6809,"Record function id": 0, "finished": true, "Ev Idx": 6808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mean", "pid": 870, "tid": 870,
    "ts": 7429770636581.847, "dur": 11.048,
    "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": 870, "tid": 870,
    "ts": 7429770636595.893, "dur": 7.564,
    "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": 870, "tid": 870,
    "ts": 7429770636605.073, "dur": 7.302,
    "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": 870, "tid": 870,
    "ts": 7429770636614.208, "dur": 7.205,
    "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": 870, "tid": 870,
    "ts": 7429770636623.825, "dur": 0.124,
    "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": 870, "tid": 870,
    "ts": 7429770636624.753, "dur": 7.538,
    "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": 870, "tid": 870,
    "ts": 7429770636723.479, "dur": 4.881,
    "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": 870, "tid": 870,
    "ts": 7429770636726.063, "dur": 0.755,
    "args": {
      "External id": 6817,"Record function id": 0, "finished": true, "Ev Idx": 6816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770636729.555, "dur": 2.283,
    "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": 870, "tid": 870,
    "ts": 7429770636730.505, "dur": 0.610,
    "args": {
      "External id": 6819,"Record function id": 0, "finished": true, "Ev Idx": 6818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 870, "tid": 870,
    "ts": 7429770636732.626, "dur": 1.688,
    "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": 870, "tid": 870,
    "ts": 7429770636733.397, "dur": 0.347,
    "args": {
      "External id": 6821,"Record function id": 0, "finished": true, "Ev Idx": 6820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 870, "tid": 870,
    "ts": 7429770636745.125, "dur": 106.478,
    "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": 870, "tid": 870,
    "ts": 7429770636745.634, "dur": 2.949,
    "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": 870, "tid": 870,
    "ts": 7429770636746.724, "dur": 1.204,
    "args": {
      "External id": 6824,"Record function id": 0, "finished": true, "Ev Idx": 6823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 870, "tid": 870,
    "ts": 7429770636747.313, "dur": 0.407,
    "args": {
      "External id": 6825,"Record function id": 0, "finished": true, "Ev Idx": 6824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 870, "tid": 870,
    "ts": 7429770636749.016, "dur": 102.343,
    "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": 870, "tid": 870,
    "ts": 7429770636749.597, "dur": 2.021,
    "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": 870, "tid": 870,
    "ts": 7429770636749.941, "dur": 1.546,
    "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": 870, "tid": 870,
    "ts": 7429770636752.036, "dur": 96.227,
    "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": 870, "tid": 870,
    "ts": 7429770636763.281, "dur": 60.558,
    "args": {
      "External id": 6830,"Record function id": 0, "finished": true, "Ev Idx": 6829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 870, "tid": 870,
    "ts": 7429770636849.104, "dur": 1.651,
    "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": 7429770598847.363, "dur": 77364.858,
    "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": 7429770598849.524, "dur": 77362.697,
    "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": 7429770598851.726, "dur": 77360.495,
    "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": 7429770598853.047, "dur": 77359.174,
    "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": 7429770598855.959, "dur": 77356.262,
    "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": 870, "tid": 870,
    "ts": 7429770598886.621, "dur": 77325.600,
    "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": 870, "tid": 870,
    "ts": 7429770598887.601, "dur": 137.185,
    "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": 870, "tid": 870,
    "ts": 7429770598888.623, "dur": 135.903,
    "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": 870, "tid": 870,
    "ts": 7429770598890.738, "dur": 132.606,
    "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": 870, "tid": 870,
    "ts": 7429770598892.261, "dur": 130.770,
    "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": 870, "tid": 870,
    "ts": 7429770598893.118, "dur": 31.146,
    "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": 870, "tid": 870,
    "ts": 7429770598920.437, "dur": 1.385,
    "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": 870, "tid": 870,
    "ts": 7429770598930.185, "dur": 11.453,
    "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 0x7f984a8770f0>", "pid": 870, "tid": 870,
    "ts": 7429770598938.119, "dur": 3.075,
    "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 0x7f984a877150>", "pid": 870, "tid": 870,
    "ts": 7429770598944.654, "dur": 1.282,
    "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": 870, "tid": 870,
    "ts": 7429770598951.138, "dur": 20.291,
    "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": 870, "tid": 870,
    "ts": 7429770598958.625, "dur": 2.886,
    "args": {
      "Python parent id": 16, "Python id": 17, "finished": true, "Ev Idx": 6847
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770598960.401, "dur": 1.022,
    "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": 870, "tid": 870,
    "ts": 7429770598964.236, "dur": 7.077,
    "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": 870, "tid": 870,
    "ts": 7429770598968.810, "dur": 2.188,
    "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": 870, "tid": 870,
    "ts": 7429770598970.382, "dur": 0.509,
    "args": {
      "Python parent id": 20, "Python id": 21, "finished": true, "Ev Idx": 6851
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770598973.276, "dur": 1.835,
    "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": 870, "tid": 870,
    "ts": 7429770598989.250, "dur": 28.854,
    "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": 870, "tid": 870,
    "ts": 7429770598996.400, "dur": 21.508,
    "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": 870, "tid": 870,
    "ts": 7429770599006.786, "dur": 6.371,
    "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": 870, "tid": 870,
    "ts": 7429770599010.539, "dur": 0.800,
    "args": {
      "Python parent id": 25, "Python id": 26, "finished": true, "Ev Idx": 6856
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770599016.967, "dur": 0.720,
    "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 0x7f8f8b4c7fc0>", "pid": 870, "tid": 870,
    "ts": 7429770599021.781, "dur": 0.445,
    "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": 870, "tid": 870,
    "ts": 7429770599028.328, "dur": 1.384,
    "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 0x7f8f8b77b180>", "pid": 870, "tid": 870,
    "ts": 7429770599029.362, "dur": 0.259,
    "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": 870, "tid": 870,
    "ts": 7429770599033.287, "dur": 1.350,
    "args": {
      "Python parent id": 6, "Python id": 31, "finished": true, "Ev Idx": 6861
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770599034.292, "dur": 0.126,
    "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": 870, "tid": 870,
    "ts": 7429770599035.505, "dur": 0.251,
    "args": {
      "Python parent id": 6, "Python id": 33, "finished": true, "Ev Idx": 6863
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770599035.670, "dur": 0.041,
    "args": {
      "Python parent id": 33, "Python id": 34, "finished": true, "Ev Idx": 6864
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaForCausalLM_0", "pid": 870, "tid": 870,
    "ts": 7429770599050.928, "dur": 37836.354,
    "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": 870, "tid": 870,
    "ts": 7429770599057.376, "dur": 37828.608,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770599059.630, "dur": 1.743,
    "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": 870, "tid": 870,
    "ts": 7429770599066.306, "dur": 37818.872,
    "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 0x7f8f8b4c4d80>", "pid": 870, "tid": 870,
    "ts": 7429770599067.995, "dur": 0.548,
    "args": {
      "Python parent id": 38, "Python id": 39, "finished": true, "Ev Idx": 6869
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770599069.134, "dur": 0.228,
    "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": 870, "tid": 870,
    "ts": 7429770599071.504, "dur": 6.347,
    "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": 870, "tid": 870,
    "ts": 7429770599075.441, "dur": 2.151,
    "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": 870, "tid": 870,
    "ts": 7429770599075.793, "dur": 0.810,
    "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": 870, "tid": 870,
    "ts": 7429770599077.027, "dur": 0.456,
    "args": {
      "Python parent id": 42, "Python id": 44, "finished": true, "Ev Idx": 6874
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770599078.711, "dur": 6.182,
    "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": 870, "tid": 870,
    "ts": 7429770599080.552, "dur": 3.586,
    "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": 870, "tid": 870,
    "ts": 7429770599090.411, "dur": 37793.637,
    "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": 870, "tid": 870,
    "ts": 7429770599098.276, "dur": 37784.336,
    "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": 870, "tid": 870,
    "ts": 7429770599098.866, "dur": 0.800,
    "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": 870, "tid": 870,
    "ts": 7429770599100.268, "dur": 0.463,
    "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": 870, "tid": 870,
    "ts": 7429770599101.485, "dur": 0.709,
    "args": {
      "Python parent id": 48, "Python id": 51, "finished": true, "Ev Idx": 6881
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaModel_0", "pid": 870, "tid": 870,
    "ts": 7429770599107.693, "dur": 37607.542,
    "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": 870, "tid": 870,
    "ts": 7429770599110.266, "dur": 37604.085,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770599110.549, "dur": 0.228,
    "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": 870, "tid": 870,
    "ts": 7429770599113.253, "dur": 37600.398,
    "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 0x7f8f8b4c4d80>", "pid": 870, "tid": 870,
    "ts": 7429770599113.477, "dur": 0.079,
    "args": {
      "Python parent id": 55, "Python id": 56, "finished": true, "Ev Idx": 6886
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770599113.769, "dur": 0.691,
    "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": 870, "tid": 870,
    "ts": 7429770599114.777, "dur": 3.307,
    "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": 870, "tid": 870,
    "ts": 7429770599116.330, "dur": 1.666,
    "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": 870, "tid": 870,
    "ts": 7429770599116.769, "dur": 0.489,
    "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": 870, "tid": 870,
    "ts": 7429770599117.506, "dur": 0.398,
    "args": {
      "Python parent id": 59, "Python id": 61, "finished": true, "Ev Idx": 6891
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770599118.322, "dur": 1.982,
    "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": 870, "tid": 870,
    "ts": 7429770599118.911, "dur": 1.279,
    "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": 870, "tid": 870,
    "ts": 7429770599125.093, "dur": 37584.366,
    "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": 870, "tid": 870,
    "ts": 7429770599130.341, "dur": 0.577,
    "args": {
      "Python parent id": 64, "Python id": 65, "finished": true, "Ev Idx": 6895
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770599132.704, "dur": 0.177,
    "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": 870, "tid": 870,
    "ts": 7429770599133.672, "dur": 0.494,
    "args": {
      "Python parent id": 64, "Python id": 67, "finished": true, "Ev Idx": 6897
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Embedding_0", "pid": 870, "tid": 870,
    "ts": 7429770599146.752, "dur": 196.922,
    "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": 870, "tid": 870,
    "ts": 7429770599148.198, "dur": 194.961,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770599148.438, "dur": 0.191,
    "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": 870, "tid": 870,
    "ts": 7429770599152.811, "dur": 189.864,
    "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": 870, "tid": 870,
    "ts": 7429770599153.878, "dur": 0.374,
    "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": 870, "tid": 870,
    "ts": 7429770599157.980, "dur": 184.454,
    "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": 870, "tid": 870,
    "ts": 7429770599159.501, "dur": 0.205,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770599162.990, "dur": 179.222,
    "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": 870, "tid": 870,
    "ts": 7429770599347.136, "dur": 3.145,
    "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": 870, "tid": 870,
    "ts": 7429770599349.229, "dur": 0.643,
    "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": 870, "tid": 870,
    "ts": 7429770599353.080, "dur": 2.560,
    "args": {
      "Python parent id": 64, "Python id": 78, "finished": true, "Ev Idx": 6908
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770599355.132, "dur": 0.238,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770599360.287, "dur": 28.920,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770599391.728, "dur": 21.098,
    "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": 870, "tid": 870,
    "ts": 7429770599417.408, "dur": 331.981,
    "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": 870, "tid": 870,
    "ts": 7429770599417.856, "dur": 6.477,
    "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": 870, "tid": 870,
    "ts": 7429770599420.444, "dur": 3.768,
    "args": {
      "Python parent id": 83, "Python id": 84, "finished": true, "Ev Idx": 6914
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770599420.914, "dur": 1.646,
    "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": 870, "tid": 870,
    "ts": 7429770599421.478, "dur": 0.923,
    "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": 870, "tid": 870,
    "ts": 7429770599422.920, "dur": 0.430,
    "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": 870, "tid": 870,
    "ts": 7429770599423.725, "dur": 0.408,
    "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": 870, "tid": 870,
    "ts": 7429770599425.215, "dur": 2.720,
    "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": 870, "tid": 870,
    "ts": 7429770599425.810, "dur": 2.050,
    "args": {
      "Python parent id": 89, "Python id": 90, "finished": true, "Ev Idx": 6920
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770599425.955, "dur": 0.647,
    "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": 870, "tid": 870,
    "ts": 7429770599426.132, "dur": 0.397,
    "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": 870, "tid": 870,
    "ts": 7429770599426.812, "dur": 0.372,
    "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": 870, "tid": 870,
    "ts": 7429770599427.403, "dur": 0.402,
    "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": 870, "tid": 870,
    "ts": 7429770599428.828, "dur": 0.367,
    "args": {
      "Python parent id": 82, "Python id": 95, "finished": true, "Ev Idx": 6925
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770599428.990, "dur": 0.083,
    "args": {
      "Python parent id": 95, "Python id": 96, "finished": true, "Ev Idx": 6926
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770599429.667, "dur": 0.299,
    "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": 870, "tid": 870,
    "ts": 7429770599430.403, "dur": 2.590,
    "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": 870, "tid": 870,
    "ts": 7429770599431.059, "dur": 1.858,
    "args": {
      "Python parent id": 98, "Python id": 99, "finished": true, "Ev Idx": 6929
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770599431.144, "dur": 0.602,
    "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": 870, "tid": 870,
    "ts": 7429770599431.260, "dur": 0.393,
    "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": 870, "tid": 870,
    "ts": 7429770599431.895, "dur": 0.404,
    "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": 870, "tid": 870,
    "ts": 7429770599432.489, "dur": 0.368,
    "args": {
      "Python parent id": 99, "Python id": 103, "finished": true, "Ev Idx": 6933
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770599435.495, "dur": 0.082,
    "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": 870, "tid": 870,
    "ts": 7429770599440.050, "dur": 302.881,
    "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 0x7f8f8b5c3bb0>", "pid": 870, "tid": 870,
    "ts": 7429770599442.024, "dur": 0.346,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770599447.006, "dur": 24.762,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770599473.666, "dur": 14.742,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770599491.679, "dur": 11.050,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770599504.712, "dur": 4.843,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770599571.686, "dur": 11.279,
    "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 0x7f8f8b5c3f70>", "pid": 870, "tid": 870,
    "ts": 7429770599585.061, "dur": 32.340,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770599637.766, "dur": 3.516,
    "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 0x7f8f8b5c3110>", "pid": 870, "tid": 870,
    "ts": 7429770599687.858, "dur": 38.443,
    "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": 870, "tid": 870,
    "ts": 7429770599744.988, "dur": 4.270,
    "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": 870, "tid": 870,
    "ts": 7429770599746.462, "dur": 2.713,
    "args": {
      "Python parent id": 115, "Python id": 116, "finished": true, "Ev Idx": 6946
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770599746.666, "dur": 1.072,
    "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": 870, "tid": 870,
    "ts": 7429770599746.878, "dur": 0.756,
    "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": 870, "tid": 870,
    "ts": 7429770599747.943, "dur": 0.473,
    "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": 870, "tid": 870,
    "ts": 7429770599748.682, "dur": 0.429,
    "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": 870, "tid": 870,
    "ts": 7429770599750.987, "dur": 0.869,
    "args": {
      "Python parent id": 64, "Python id": 121, "finished": true, "Ev Idx": 6951
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRotaryEmbedding_0", "pid": 870, "tid": 870,
    "ts": 7429770599757.854, "dur": 322.342,
    "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": 870, "tid": 870,
    "ts": 7429770599759.459, "dur": 320.176,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770599760.005, "dur": 0.851,
    "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": 870, "tid": 870,
    "ts": 7429770599764.403, "dur": 314.919,
    "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": 870, "tid": 870,
    "ts": 7429770599765.651, "dur": 9.141,
    "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": 870, "tid": 870,
    "ts": 7429770599767.577, "dur": 2.223,
    "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 0x55a819ef3300>", "pid": 870, "tid": 870,
    "ts": 7429770599769.420, "dur": 0.314,
    "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": 870, "tid": 870,
    "ts": 7429770599771.568, "dur": 2.979,
    "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": 870, "tid": 870,
    "ts": 7429770599773.619, "dur": 0.075,
    "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": 870, "tid": 870,
    "ts": 7429770599776.566, "dur": 8.763,
    "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": 870, "tid": 870,
    "ts": 7429770599779.963, "dur": 0.378,
    "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": 870, "tid": 870,
    "ts": 7429770599782.332, "dur": 2.677,
    "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": 870, "tid": 870,
    "ts": 7429770599782.635, "dur": 0.087,
    "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": 870, "tid": 870,
    "ts": 7429770599784.028, "dur": 0.862,
    "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": 870, "tid": 870,
    "ts": 7429770599787.173, "dur": 287.781,
    "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": 870, "tid": 870,
    "ts": 7429770599791.079, "dur": 280.657,
    "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": 870, "tid": 870,
    "ts": 7429770599791.947, "dur": 0.684,
    "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 0x7f8f8b5c3020>", "pid": 870, "tid": 870,
    "ts": 7429770599802.018, "dur": 2.371,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770599805.398, "dur": 3.420,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770599809.727, "dur": 1.318,
    "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 0x7f8f8b5c3020>", "pid": 870, "tid": 870,
    "ts": 7429770599818.004, "dur": 16.329,
    "args": {
      "Python parent id": 137, "Python id": 142, "finished": true, "Ev Idx": 6972
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770599839.960, "dur": 0.189,
    "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": 870, "tid": 870,
    "ts": 7429770599857.532, "dur": 13.562,
    "args": {
      "Python parent id": 137, "Python id": 144, "finished": true, "Ev Idx": 6974
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770599857.776, "dur": 0.118,
    "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": 870, "tid": 870,
    "ts": 7429770599859.310, "dur": 1.673,
    "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": 870, "tid": 870,
    "ts": 7429770599861.726, "dur": 0.090,
    "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": 870, "tid": 870,
    "ts": 7429770599863.525, "dur": 1.359,
    "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": 870, "tid": 870,
    "ts": 7429770599864.485, "dur": 0.347,
    "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": 870, "tid": 870,
    "ts": 7429770599866.773, "dur": 0.500,
    "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": 870, "tid": 870,
    "ts": 7429770599867.812, "dur": 0.253,
    "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": 870, "tid": 870,
    "ts": 7429770599869.949, "dur": 0.216,
    "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": 870, "tid": 870,
    "ts": 7429770599873.995, "dur": 13.772,
    "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": 870, "tid": 870,
    "ts": 7429770599876.687, "dur": 0.046,
    "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": 870, "tid": 870,
    "ts": 7429770599877.384, "dur": 0.416,
    "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": 870, "tid": 870,
    "ts": 7429770599879.119, "dur": 0.687,
    "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": 870, "tid": 870,
    "ts": 7429770599880.207, "dur": 0.192,
    "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": 870, "tid": 870,
    "ts": 7429770599881.945, "dur": 0.693,
    "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": 870, "tid": 870,
    "ts": 7429770599883.594, "dur": 0.611,
    "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": 870, "tid": 870,
    "ts": 7429770599884.860, "dur": 0.216,
    "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": 870, "tid": 870,
    "ts": 7429770599886.117, "dur": 0.224,
    "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": 870, "tid": 870,
    "ts": 7429770599887.274, "dur": 0.359,
    "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 0x7f8f8b5c3020>", "pid": 870, "tid": 870,
    "ts": 7429770599888.348, "dur": 1.101,
    "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 0x7f8f8b5c3020>", "pid": 870, "tid": 870,
    "ts": 7429770599889.888, "dur": 0.617,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770599970.404, "dur": 6.675,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770599979.134, "dur": 22.106,
    "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 0x7f8f8b4d1860>", "pid": 870, "tid": 870,
    "ts": 7429770600003.347, "dur": 13.491,
    "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 0x7f8f8b4d1860>", "pid": 870, "tid": 870,
    "ts": 7429770600036.322, "dur": 10.566,
    "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": 870, "tid": 870,
    "ts": 7429770600057.735, "dur": 8.746,
    "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": 870, "tid": 870,
    "ts": 7429770600058.510, "dur": 0.053,
    "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": 870, "tid": 870,
    "ts": 7429770600059.382, "dur": 0.313,
    "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": 870, "tid": 870,
    "ts": 7429770600060.615, "dur": 0.534,
    "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": 870, "tid": 870,
    "ts": 7429770600061.712, "dur": 0.624,
    "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": 870, "tid": 870,
    "ts": 7429770600062.801, "dur": 0.366,
    "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": 870, "tid": 870,
    "ts": 7429770600063.608, "dur": 0.153,
    "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": 870, "tid": 870,
    "ts": 7429770600066.216, "dur": 0.186,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770600068.289, "dur": 1.792,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770600070.742, "dur": 0.873,
    "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": 870, "tid": 870,
    "ts": 7429770600076.409, "dur": 2.580,
    "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": 870, "tid": 870,
    "ts": 7429770600077.511, "dur": 1.270,
    "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": 870, "tid": 870,
    "ts": 7429770600077.712, "dur": 0.167,
    "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": 870, "tid": 870,
    "ts": 7429770600078.223, "dur": 0.463,
    "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": 870, "tid": 870,
    "ts": 7429770600082.158, "dur": 0.777,
    "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": 870, "tid": 870,
    "ts": 7429770600083.882, "dur": 1.243,
    "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": 870, "tid": 870,
    "ts": 7429770600087.502, "dur": 131.053,
    "args": {
      "Python parent id": 64, "Python id": 185, "finished": true, "Ev Idx": 7015
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600087.722, "dur": 0.104,
    "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 0x7f8f8b4ab740>", "pid": 870, "tid": 870,
    "ts": 7429770600090.869, "dur": 0.283,
    "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": 870, "tid": 870,
    "ts": 7429770600095.286, "dur": 122.978,
    "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": 870, "tid": 870,
    "ts": 7429770600100.350, "dur": 11.573,
    "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 0x7f99b7992c10>", "pid": 870, "tid": 870,
    "ts": 7429770600101.954, "dur": 1.385,
    "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": 870, "tid": 870,
    "ts": 7429770600113.009, "dur": 105.037,
    "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": 870, "tid": 870,
    "ts": 7429770600114.629, "dur": 103.194,
    "args": {
      "Python parent id": 191, "Python id": 192, "finished": true, "Ev Idx": 7022
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600115.196, "dur": 5.016,
    "args": {
      "Python parent id": 192, "Python id": 193, "finished": true, "Ev Idx": 7023
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen abc>(117): __instancecheck__", "pid": 870, "tid": 870,
    "ts": 7429770600116.882, "dur": 3.146,
    "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": 870, "tid": 870,
    "ts": 7429770600118.335, "dur": 1.598,
    "args": {
      "Python parent id": 194, "Python id": 195, "finished": true, "Ev Idx": 7025
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770600120.923, "dur": 1.534,
    "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": 870, "tid": 870,
    "ts": 7429770600121.970, "dur": 0.380,
    "args": {
      "Python parent id": 196, "Python id": 197, "finished": true, "Ev Idx": 7027
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770600122.148, "dur": 0.156,
    "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": 870, "tid": 870,
    "ts": 7429770600127.474, "dur": 9.067,
    "args": {
      "Python parent id": 192, "Python id": 199, "finished": true, "Ev Idx": 7029
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600127.740, "dur": 0.174,
    "args": {
      "Python parent id": 199, "Python id": 200, "finished": true, "Ev Idx": 7030
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600128.418, "dur": 0.045,
    "args": {
      "Python parent id": 199, "Python id": 201, "finished": true, "Ev Idx": 7031
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770600128.872, "dur": 4.389,
    "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": 870, "tid": 870,
    "ts": 7429770600129.648, "dur": 2.952,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600135.594, "dur": 0.281,
    "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": 870, "tid": 870,
    "ts": 7429770600137.692, "dur": 2.950,
    "args": {
      "Python parent id": 192, "Python id": 205, "finished": true, "Ev Idx": 7035
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600137.868, "dur": 0.107,
    "args": {
      "Python parent id": 205, "Python id": 206, "finished": true, "Ev Idx": 7036
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600138.155, "dur": 0.048,
    "args": {
      "Python parent id": 205, "Python id": 207, "finished": true, "Ev Idx": 7037
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770600138.398, "dur": 1.558,
    "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": 870, "tid": 870,
    "ts": 7429770600139.016, "dur": 0.784,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600140.264, "dur": 0.117,
    "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": 870, "tid": 870,
    "ts": 7429770600141.231, "dur": 1.985,
    "args": {
      "Python parent id": 192, "Python id": 211, "finished": true, "Ev Idx": 7041
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600141.302, "dur": 0.049,
    "args": {
      "Python parent id": 211, "Python id": 212, "finished": true, "Ev Idx": 7042
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600141.452, "dur": 0.063,
    "args": {
      "Python parent id": 211, "Python id": 213, "finished": true, "Ev Idx": 7043
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770600141.680, "dur": 1.048,
    "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": 870, "tid": 870,
    "ts": 7429770600142.061, "dur": 0.557,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600142.999, "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": 870, "tid": 870,
    "ts": 7429770600143.660, "dur": 2.645,
    "args": {
      "Python parent id": 192, "Python id": 217, "finished": true, "Ev Idx": 7047
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600143.734, "dur": 0.041,
    "args": {
      "Python parent id": 217, "Python id": 218, "finished": true, "Ev Idx": 7048
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600143.873, "dur": 0.038,
    "args": {
      "Python parent id": 217, "Python id": 219, "finished": true, "Ev Idx": 7049
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770600144.043, "dur": 1.643,
    "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": 870, "tid": 870,
    "ts": 7429770600144.907, "dur": 0.546,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600145.980, "dur": 0.147,
    "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": 870, "tid": 870,
    "ts": 7429770600146.767, "dur": 2.033,
    "args": {
      "Python parent id": 192, "Python id": 223, "finished": true, "Ev Idx": 7053
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600146.837, "dur": 0.042,
    "args": {
      "Python parent id": 223, "Python id": 224, "finished": true, "Ev Idx": 7054
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600147.000, "dur": 0.039,
    "args": {
      "Python parent id": 223, "Python id": 225, "finished": true, "Ev Idx": 7055
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770600147.168, "dur": 1.080,
    "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": 870, "tid": 870,
    "ts": 7429770600147.629, "dur": 0.511,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600148.486, "dur": 0.084,
    "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": 870, "tid": 870,
    "ts": 7429770600149.211, "dur": 2.805,
    "args": {
      "Python parent id": 192, "Python id": 229, "finished": true, "Ev Idx": 7059
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600149.284, "dur": 0.044,
    "args": {
      "Python parent id": 229, "Python id": 230, "finished": true, "Ev Idx": 7060
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600149.539, "dur": 0.036,
    "args": {
      "Python parent id": 229, "Python id": 231, "finished": true, "Ev Idx": 7061
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770600149.700, "dur": 1.467,
    "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": 870, "tid": 870,
    "ts": 7429770600150.529, "dur": 0.536,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600151.407, "dur": 0.086,
    "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": 870, "tid": 870,
    "ts": 7429770600152.453, "dur": 1.972,
    "args": {
      "Python parent id": 192, "Python id": 235, "finished": true, "Ev Idx": 7065
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600152.523, "dur": 0.043,
    "args": {
      "Python parent id": 235, "Python id": 236, "finished": true, "Ev Idx": 7066
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600152.663, "dur": 0.037,
    "args": {
      "Python parent id": 235, "Python id": 237, "finished": true, "Ev Idx": 7067
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770600152.827, "dur": 1.020,
    "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": 870, "tid": 870,
    "ts": 7429770600153.201, "dur": 0.523,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600154.114, "dur": 0.094,
    "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": 870, "tid": 870,
    "ts": 7429770600154.838, "dur": 2.012,
    "args": {
      "Python parent id": 192, "Python id": 241, "finished": true, "Ev Idx": 7071
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600154.909, "dur": 0.046,
    "args": {
      "Python parent id": 241, "Python id": 242, "finished": true, "Ev Idx": 7072
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600155.058, "dur": 0.041,
    "args": {
      "Python parent id": 241, "Python id": 243, "finished": true, "Ev Idx": 7073
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770600155.251, "dur": 1.085,
    "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": 870, "tid": 870,
    "ts": 7429770600155.584, "dur": 0.534,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600156.582, "dur": 0.066,
    "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": 870, "tid": 870,
    "ts": 7429770600157.266, "dur": 1.998,
    "args": {
      "Python parent id": 192, "Python id": 247, "finished": true, "Ev Idx": 7077
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600157.341, "dur": 0.046,
    "args": {
      "Python parent id": 247, "Python id": 248, "finished": true, "Ev Idx": 7078
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600157.487, "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": 870, "tid": 870,
    "ts": 7429770600157.654, "dur": 0.970,
    "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": 870, "tid": 870,
    "ts": 7429770600158.003, "dur": 0.514,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600158.888, "dur": 0.079,
    "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": 870, "tid": 870,
    "ts": 7429770600159.663, "dur": 1.972,
    "args": {
      "Python parent id": 192, "Python id": 253, "finished": true, "Ev Idx": 7083
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600159.735, "dur": 0.044,
    "args": {
      "Python parent id": 253, "Python id": 254, "finished": true, "Ev Idx": 7084
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600159.880, "dur": 0.035,
    "args": {
      "Python parent id": 253, "Python id": 255, "finished": true, "Ev Idx": 7085
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770600160.068, "dur": 1.004,
    "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": 870, "tid": 870,
    "ts": 7429770600160.428, "dur": 0.540,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600161.363, "dur": 0.082,
    "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": 870, "tid": 870,
    "ts": 7429770600162.099, "dur": 2.821,
    "args": {
      "Python parent id": 192, "Python id": 259, "finished": true, "Ev Idx": 7089
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600162.168, "dur": 0.045,
    "args": {
      "Python parent id": 259, "Python id": 260, "finished": true, "Ev Idx": 7090
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600162.315, "dur": 0.045,
    "args": {
      "Python parent id": 259, "Python id": 261, "finished": true, "Ev Idx": 7091
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770600162.496, "dur": 1.141,
    "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": 870, "tid": 870,
    "ts": 7429770600162.930, "dur": 0.609,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600164.001, "dur": 0.235,
    "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": 870, "tid": 870,
    "ts": 7429770600165.334, "dur": 2.339,
    "args": {
      "Python parent id": 192, "Python id": 265, "finished": true, "Ev Idx": 7095
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600165.406, "dur": 0.048,
    "args": {
      "Python parent id": 265, "Python id": 266, "finished": true, "Ev Idx": 7096
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600165.577, "dur": 0.038,
    "args": {
      "Python parent id": 265, "Python id": 267, "finished": true, "Ev Idx": 7097
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770600165.753, "dur": 1.335,
    "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": 870, "tid": 870,
    "ts": 7429770600166.438, "dur": 0.544,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600167.370, "dur": 0.121,
    "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": 870, "tid": 870,
    "ts": 7429770600168.101, "dur": 2.045,
    "args": {
      "Python parent id": 192, "Python id": 271, "finished": true, "Ev Idx": 7101
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600168.175, "dur": 0.038,
    "args": {
      "Python parent id": 271, "Python id": 272, "finished": true, "Ev Idx": 7102
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600168.317, "dur": 0.035,
    "args": {
      "Python parent id": 271, "Python id": 273, "finished": true, "Ev Idx": 7103
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770600168.484, "dur": 0.971,
    "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": 870, "tid": 870,
    "ts": 7429770600168.858, "dur": 0.490,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600169.793, "dur": 0.068,
    "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": 870, "tid": 870,
    "ts": 7429770600170.547, "dur": 1.837,
    "args": {
      "Python parent id": 192, "Python id": 277, "finished": true, "Ev Idx": 7107
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600170.620, "dur": 0.039,
    "args": {
      "Python parent id": 277, "Python id": 278, "finished": true, "Ev Idx": 7108
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600170.757, "dur": 0.033,
    "args": {
      "Python parent id": 277, "Python id": 279, "finished": true, "Ev Idx": 7109
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770600170.921, "dur": 0.870,
    "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": 870, "tid": 870,
    "ts": 7429770600171.179, "dur": 0.510,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600172.093, "dur": 0.095,
    "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": 870, "tid": 870,
    "ts": 7429770600172.808, "dur": 2.213,
    "args": {
      "Python parent id": 192, "Python id": 283, "finished": true, "Ev Idx": 7113
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600172.905, "dur": 0.041,
    "args": {
      "Python parent id": 283, "Python id": 284, "finished": true, "Ev Idx": 7114
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600173.043, "dur": 0.032,
    "args": {
      "Python parent id": 283, "Python id": 285, "finished": true, "Ev Idx": 7115
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770600173.208, "dur": 1.069,
    "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": 870, "tid": 870,
    "ts": 7429770600173.692, "dur": 0.483,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600174.574, "dur": 0.087,
    "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": 870, "tid": 870,
    "ts": 7429770600175.418, "dur": 2.048,
    "args": {
      "Python parent id": 192, "Python id": 289, "finished": true, "Ev Idx": 7119
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600175.491, "dur": 0.040,
    "args": {
      "Python parent id": 289, "Python id": 290, "finished": true, "Ev Idx": 7120
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600175.629, "dur": 0.036,
    "args": {
      "Python parent id": 289, "Python id": 291, "finished": true, "Ev Idx": 7121
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770600175.795, "dur": 1.205,
    "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": 870, "tid": 870,
    "ts": 7429770600176.278, "dur": 0.488,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600177.262, "dur": 0.070,
    "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": 870, "tid": 870,
    "ts": 7429770600177.922, "dur": 2.073,
    "args": {
      "Python parent id": 192, "Python id": 295, "finished": true, "Ev Idx": 7125
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600177.995, "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": 870, "tid": 870,
    "ts": 7429770600178.138, "dur": 0.034,
    "args": {
      "Python parent id": 295, "Python id": 297, "finished": true, "Ev Idx": 7127
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770600178.307, "dur": 0.911,
    "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": 870, "tid": 870,
    "ts": 7429770600178.656, "dur": 0.461,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600179.485, "dur": 0.209,
    "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": 870, "tid": 870,
    "ts": 7429770600180.382, "dur": 2.679,
    "args": {
      "Python parent id": 192, "Python id": 301, "finished": true, "Ev Idx": 7131
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600180.452, "dur": 0.039,
    "args": {
      "Python parent id": 301, "Python id": 302, "finished": true, "Ev Idx": 7132
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600180.743, "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": 870, "tid": 870,
    "ts": 7429770600180.908, "dur": 1.154,
    "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": 870, "tid": 870,
    "ts": 7429770600181.437, "dur": 0.524,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600182.311, "dur": 0.219,
    "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": 870, "tid": 870,
    "ts": 7429770600183.460, "dur": 2.197,
    "args": {
      "Python parent id": 192, "Python id": 307, "finished": true, "Ev Idx": 7137
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600183.532, "dur": 0.045,
    "args": {
      "Python parent id": 307, "Python id": 308, "finished": true, "Ev Idx": 7138
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600183.675, "dur": 0.035,
    "args": {
      "Python parent id": 307, "Python id": 309, "finished": true, "Ev Idx": 7139
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770600183.840, "dur": 1.267,
    "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": 870, "tid": 870,
    "ts": 7429770600184.511, "dur": 0.495,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600185.394, "dur": 0.102,
    "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": 870, "tid": 870,
    "ts": 7429770600186.066, "dur": 2.551,
    "args": {
      "Python parent id": 192, "Python id": 313, "finished": true, "Ev Idx": 7143
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600186.139, "dur": 0.038,
    "args": {
      "Python parent id": 313, "Python id": 314, "finished": true, "Ev Idx": 7144
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600186.398, "dur": 0.035,
    "args": {
      "Python parent id": 313, "Python id": 315, "finished": true, "Ev Idx": 7145
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770600186.559, "dur": 1.441,
    "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": 870, "tid": 870,
    "ts": 7429770600187.247, "dur": 0.492,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600188.255, "dur": 0.211,
    "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": 870, "tid": 870,
    "ts": 7429770600189.046, "dur": 2.800,
    "args": {
      "Python parent id": 192, "Python id": 319, "finished": true, "Ev Idx": 7149
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600189.118, "dur": 0.044,
    "args": {
      "Python parent id": 319, "Python id": 320, "finished": true, "Ev Idx": 7150
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600189.262, "dur": 0.035,
    "args": {
      "Python parent id": 319, "Python id": 321, "finished": true, "Ev Idx": 7151
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770600189.457, "dur": 1.764,
    "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": 870, "tid": 870,
    "ts": 7429770600190.429, "dur": 0.633,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600191.518, "dur": 0.124,
    "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": 870, "tid": 870,
    "ts": 7429770600192.428, "dur": 2.549,
    "args": {
      "Python parent id": 192, "Python id": 325, "finished": true, "Ev Idx": 7155
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600192.500, "dur": 0.054,
    "args": {
      "Python parent id": 325, "Python id": 326, "finished": true, "Ev Idx": 7156
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600192.734, "dur": 0.035,
    "args": {
      "Python parent id": 325, "Python id": 327, "finished": true, "Ev Idx": 7157
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770600192.899, "dur": 1.018,
    "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": 870, "tid": 870,
    "ts": 7429770600193.291, "dur": 0.504,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600194.183, "dur": 0.278,
    "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": 870, "tid": 870,
    "ts": 7429770600195.365, "dur": 1.842,
    "args": {
      "Python parent id": 192, "Python id": 331, "finished": true, "Ev Idx": 7161
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600195.435, "dur": 0.042,
    "args": {
      "Python parent id": 331, "Python id": 332, "finished": true, "Ev Idx": 7162
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600195.575, "dur": 0.035,
    "args": {
      "Python parent id": 331, "Python id": 333, "finished": true, "Ev Idx": 7163
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770600195.743, "dur": 0.940,
    "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": 870, "tid": 870,
    "ts": 7429770600196.038, "dur": 0.524,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600196.952, "dur": 0.118,
    "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": 870, "tid": 870,
    "ts": 7429770600197.592, "dur": 1.944,
    "args": {
      "Python parent id": 192, "Python id": 337, "finished": true, "Ev Idx": 7167
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600197.661, "dur": 0.043,
    "args": {
      "Python parent id": 337, "Python id": 338, "finished": true, "Ev Idx": 7168
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600197.803, "dur": 0.035,
    "args": {
      "Python parent id": 337, "Python id": 339, "finished": true, "Ev Idx": 7169
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770600197.970, "dur": 1.038,
    "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": 870, "tid": 870,
    "ts": 7429770600198.396, "dur": 0.484,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600199.306, "dur": 0.109,
    "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": 870, "tid": 870,
    "ts": 7429770600199.924, "dur": 2.110,
    "args": {
      "Python parent id": 192, "Python id": 343, "finished": true, "Ev Idx": 7173
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600199.997, "dur": 0.039,
    "args": {
      "Python parent id": 343, "Python id": 344, "finished": true, "Ev Idx": 7174
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600200.134, "dur": 0.033,
    "args": {
      "Python parent id": 343, "Python id": 345, "finished": true, "Ev Idx": 7175
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770600200.296, "dur": 1.197,
    "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": 870, "tid": 870,
    "ts": 7429770600200.810, "dur": 0.575,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600201.738, "dur": 0.110,
    "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": 870, "tid": 870,
    "ts": 7429770600202.460, "dur": 1.857,
    "args": {
      "Python parent id": 192, "Python id": 349, "finished": true, "Ev Idx": 7179
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600202.531, "dur": 0.047,
    "args": {
      "Python parent id": 349, "Python id": 350, "finished": true, "Ev Idx": 7180
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600202.681, "dur": 0.040,
    "args": {
      "Python parent id": 349, "Python id": 351, "finished": true, "Ev Idx": 7181
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770600202.862, "dur": 0.925,
    "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": 870, "tid": 870,
    "ts": 7429770600203.211, "dur": 0.474,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600204.072, "dur": 0.088,
    "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": 870, "tid": 870,
    "ts": 7429770600204.696, "dur": 1.925,
    "args": {
      "Python parent id": 192, "Python id": 355, "finished": true, "Ev Idx": 7185
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600204.768, "dur": 0.044,
    "args": {
      "Python parent id": 355, "Python id": 356, "finished": true, "Ev Idx": 7186
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600204.908, "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": 870, "tid": 870,
    "ts": 7429770600205.108, "dur": 1.041,
    "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": 870, "tid": 870,
    "ts": 7429770600205.565, "dur": 0.477,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600206.415, "dur": 0.068,
    "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": 870, "tid": 870,
    "ts": 7429770600206.989, "dur": 1.704,
    "args": {
      "Python parent id": 192, "Python id": 361, "finished": true, "Ev Idx": 7191
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600207.058, "dur": 0.044,
    "args": {
      "Python parent id": 361, "Python id": 362, "finished": true, "Ev Idx": 7192
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600207.197, "dur": 0.034,
    "args": {
      "Python parent id": 361, "Python id": 363, "finished": true, "Ev Idx": 7193
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770600207.359, "dur": 0.844,
    "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": 870, "tid": 870,
    "ts": 7429770600207.635, "dur": 0.464,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600208.441, "dur": 0.102,
    "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": 870, "tid": 870,
    "ts": 7429770600209.093, "dur": 1.835,
    "args": {
      "Python parent id": 192, "Python id": 367, "finished": true, "Ev Idx": 7197
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600209.163, "dur": 0.037,
    "args": {
      "Python parent id": 367, "Python id": 368, "finished": true, "Ev Idx": 7198
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600209.301, "dur": 0.035,
    "args": {
      "Python parent id": 367, "Python id": 369, "finished": true, "Ev Idx": 7199
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770600209.466, "dur": 1.001,
    "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": 870, "tid": 870,
    "ts": 7429770600209.821, "dur": 0.542,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600210.721, "dur": 0.089,
    "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": 870, "tid": 870,
    "ts": 7429770600211.363, "dur": 1.812,
    "args": {
      "Python parent id": 192, "Python id": 373, "finished": true, "Ev Idx": 7203
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600211.434, "dur": 0.043,
    "args": {
      "Python parent id": 373, "Python id": 374, "finished": true, "Ev Idx": 7204
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600211.583, "dur": 0.038,
    "args": {
      "Python parent id": 373, "Python id": 375, "finished": true, "Ev Idx": 7205
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770600211.780, "dur": 0.888,
    "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": 870, "tid": 870,
    "ts": 7429770600212.040, "dur": 0.519,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600212.917, "dur": 0.102,
    "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": 870, "tid": 870,
    "ts": 7429770600213.562, "dur": 1.680,
    "args": {
      "Python parent id": 192, "Python id": 379, "finished": true, "Ev Idx": 7209
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600213.631, "dur": 0.038,
    "args": {
      "Python parent id": 379, "Python id": 380, "finished": true, "Ev Idx": 7210
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600213.767, "dur": 0.037,
    "args": {
      "Python parent id": 379, "Python id": 381, "finished": true, "Ev Idx": 7211
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770600213.934, "dur": 0.839,
    "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": 870, "tid": 870,
    "ts": 7429770600214.206, "dur": 0.461,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600215.023, "dur": 0.090,
    "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": 870, "tid": 870,
    "ts": 7429770600215.621, "dur": 1.837,
    "args": {
      "Python parent id": 192, "Python id": 385, "finished": true, "Ev Idx": 7215
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770600215.723, "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": 870, "tid": 870,
    "ts": 7429770600215.866, "dur": 0.034,
    "args": {
      "Python parent id": 385, "Python id": 387, "finished": true, "Ev Idx": 7217
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770600216.029, "dur": 0.923,
    "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": 870, "tid": 870,
    "ts": 7429770600216.355, "dur": 0.487,
    "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 0x7f984abd0d40>", "pid": 870, "tid": 870,
    "ts": 7429770600217.250, "dur": 0.074,
    "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": 870, "tid": 870,
    "ts": 7429770600219.962, "dur": 1.924,
    "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 0x7f8f8b4ab740>", "pid": 870, "tid": 870,
    "ts": 7429770600220.338, "dur": 0.135,
    "args": {
      "Python parent id": 391, "Python id": 392, "finished": true, "Ev Idx": 7222
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function iter>", "pid": 870, "tid": 870,
    "ts": 7429770600221.548, "dur": 0.224,
    "args": {
      "Python parent id": 391, "Python id": 393, "finished": true, "Ev Idx": 7223
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_0", "pid": 870, "tid": 870,
    "ts": 7429770600231.458, "dur": 1234.253,
    "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": 870, "tid": 870,
    "ts": 7429770600233.636, "dur": 1231.088,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770600233.986, "dur": 0.650,
    "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": 870, "tid": 870,
    "ts": 7429770600238.415, "dur": 1225.493,
    "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": 870, "tid": 870,
    "ts": 7429770600239.110, "dur": 0.779,
    "args": {
      "Python parent id": 397, "Python id": 398, "finished": true, "Ev Idx": 7228
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_0", "pid": 870, "tid": 870,
    "ts": 7429770600246.048, "dur": 131.522,
    "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": 870, "tid": 870,
    "ts": 7429770600247.692, "dur": 129.443,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770600247.864, "dur": 0.214,
    "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": 870, "tid": 870,
    "ts": 7429770600252.329, "dur": 123.556,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770600253.519, "dur": 2.383,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770600257.334, "dur": 25.929,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770600285.137, "dur": 29.916,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770600334.047, "dur": 16.775,
    "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": 870, "tid": 870,
    "ts": 7429770600363.496, "dur": 0.303,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770600364.436, "dur": 1.415,
    "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": 870, "tid": 870,
    "ts": 7429770600378.550, "dur": 0.411,
    "args": {
      "Python parent id": 397, "Python id": 409, "finished": true, "Ev Idx": 7239
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_0", "pid": 870, "tid": 870,
    "ts": 7429770600383.969, "dur": 775.400,
    "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": 870, "tid": 870,
    "ts": 7429770600385.912, "dur": 772.708,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770600386.196, "dur": 0.373,
    "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": 870, "tid": 870,
    "ts": 7429770600391.671, "dur": 765.988,
    "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": 870, "tid": 870,
    "ts": 7429770600395.314, "dur": 0.468,
    "args": {
      "Python parent id": 413, "Python id": 414, "finished": true, "Ev Idx": 7244
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_0", "pid": 870, "tid": 870,
    "ts": 7429770600400.246, "dur": 88.832,
    "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": 870, "tid": 870,
    "ts": 7429770600401.971, "dur": 86.725,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770600402.121, "dur": 0.196,
    "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": 870, "tid": 870,
    "ts": 7429770600404.228, "dur": 84.252,
    "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": 870, "tid": 870,
    "ts": 7429770600405.331, "dur": 0.201,
    "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": 870, "tid": 870,
    "ts": 7429770600405.959, "dur": 0.172,
    "args": {
      "Python parent id": 418, "Python id": 420, "finished": true, "Ev Idx": 7250
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770600407.306, "dur": 81.074,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770600491.144, "dur": 4.190,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770600495.886, "dur": 4.484,
    "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": 870, "tid": 870,
    "ts": 7429770600501.465, "dur": 0.451,
    "args": {
      "Python parent id": 413, "Python id": 424, "finished": true, "Ev Idx": 7254
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_1", "pid": 870, "tid": 870,
    "ts": 7429770600506.408, "dur": 44.634,
    "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": 870, "tid": 870,
    "ts": 7429770600507.160, "dur": 43.611,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770600507.452, "dur": 0.298,
    "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": 870, "tid": 870,
    "ts": 7429770600508.729, "dur": 41.906,
    "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": 870, "tid": 870,
    "ts": 7429770600509.008, "dur": 0.170,
    "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": 870, "tid": 870,
    "ts": 7429770600509.475, "dur": 0.135,
    "args": {
      "Python parent id": 428, "Python id": 430, "finished": true, "Ev Idx": 7260
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770600509.771, "dur": 40.795,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770600551.721, "dur": 2.723,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770600555.192, "dur": 3.082,
    "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": 870, "tid": 870,
    "ts": 7429770600559.348, "dur": 0.385,
    "args": {
      "Python parent id": 413, "Python id": 434, "finished": true, "Ev Idx": 7264
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_2", "pid": 870, "tid": 870,
    "ts": 7429770600562.072, "dur": 37.813,
    "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": 870, "tid": 870,
    "ts": 7429770600562.617, "dur": 37.041,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770600562.852, "dur": 0.235,
    "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": 870, "tid": 870,
    "ts": 7429770600564.748, "dur": 34.766,
    "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": 870, "tid": 870,
    "ts": 7429770600565.024, "dur": 0.160,
    "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": 870, "tid": 870,
    "ts": 7429770600565.416, "dur": 0.161,
    "args": {
      "Python parent id": 438, "Python id": 440, "finished": true, "Ev Idx": 7270
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770600565.710, "dur": 33.747,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770600600.566, "dur": 2.463,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770600603.525, "dur": 2.955,
    "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": 870, "tid": 870,
    "ts": 7429770600609.019, "dur": 155.418,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770600609.319, "dur": 3.900,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770600613.848, "dur": 2.034,
    "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": 870, "tid": 870,
    "ts": 7429770600630.442, "dur": 52.663,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770600659.591, "dur": 22.292,
    "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": 870, "tid": 870,
    "ts": 7429770600716.841, "dur": 30.121,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770600733.662, "dur": 12.338,
    "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": 870, "tid": 870,
    "ts": 7429770600771.086, "dur": 7.417,
    "args": {
      "Python parent id": 413, "Python id": 451, "finished": true, "Ev Idx": 7281
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770600772.734, "dur": 0.152,
    "args": {
      "Python parent id": 451, "Python id": 452, "finished": true, "Ev Idx": 7282
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770600773.280, "dur": 0.060,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770600777.157, "dur": 0.471,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770600778.162, "dur": 0.090,
    "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": 870, "tid": 870,
    "ts": 7429770600779.892, "dur": 6.297,
    "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": 870, "tid": 870,
    "ts": 7429770600782.704, "dur": 3.364,
    "args": {
      "Python parent id": 456, "Python id": 457, "finished": true, "Ev Idx": 7287
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770600783.032, "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": 870, "tid": 870,
    "ts": 7429770600783.556, "dur": 0.728,
    "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": 870, "tid": 870,
    "ts": 7429770600784.802, "dur": 0.428,
    "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": 870, "tid": 870,
    "ts": 7429770600785.589, "dur": 0.404,
    "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": 870, "tid": 870,
    "ts": 7429770600791.582, "dur": 299.160,
    "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": 870, "tid": 870,
    "ts": 7429770600793.238, "dur": 53.058,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770600814.263, "dur": 5.787,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770600821.086, "dur": 25.078,
    "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": 870, "tid": 870,
    "ts": 7429770600847.361, "dur": 29.775,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770600861.656, "dur": 3.035,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770600865.189, "dur": 11.892,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770600878.399, "dur": 3.326,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770600883.514, "dur": 60.019,
    "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": 870, "tid": 870,
    "ts": 7429770600979.540, "dur": 21.866,
    "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": 870, "tid": 870,
    "ts": 7429770600980.730, "dur": 0.155,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770600983.326, "dur": 17.910,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770601002.418, "dur": 1.857,
    "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": 870, "tid": 870,
    "ts": 7429770601007.319, "dur": 14.126,
    "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": 870, "tid": 870,
    "ts": 7429770601007.728, "dur": 0.109,
    "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": 870, "tid": 870,
    "ts": 7429770601014.785, "dur": 1.182,
    "args": {
      "Python parent id": 475, "Python id": 477, "finished": true, "Ev Idx": 7307
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770601015.411, "dur": 0.463,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770601017.434, "dur": 3.916,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770601022.516, "dur": 46.117,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770601069.490, "dur": 3.365,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770601074.482, "dur": 15.217,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770601094.259, "dur": 3.390,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770601098.635, "dur": 0.268,
    "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": 870, "tid": 870,
    "ts": 7429770601100.503, "dur": 1.342,
    "args": {
      "Python parent id": 413, "Python id": 485, "finished": true, "Ev Idx": 7315
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_3", "pid": 870, "tid": 870,
    "ts": 7429770601107.325, "dur": 49.603,
    "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": 870, "tid": 870,
    "ts": 7429770601108.662, "dur": 47.774,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770601109.183, "dur": 0.927,
    "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": 870, "tid": 870,
    "ts": 7429770601111.714, "dur": 44.416,
    "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": 870, "tid": 870,
    "ts": 7429770601112.233, "dur": 0.371,
    "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": 870, "tid": 870,
    "ts": 7429770601112.938, "dur": 0.176,
    "args": {
      "Python parent id": 489, "Python id": 491, "finished": true, "Ev Idx": 7321
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770601113.447, "dur": 42.538,
    "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": 870, "tid": 870,
    "ts": 7429770601171.554, "dur": 0.830,
    "args": {
      "Python parent id": 397, "Python id": 493, "finished": true, "Ev Idx": 7323
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_1", "pid": 870, "tid": 870,
    "ts": 7429770601175.506, "dur": 84.150,
    "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": 870, "tid": 870,
    "ts": 7429770601176.184, "dur": 83.134,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770601176.422, "dur": 0.416,
    "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": 870, "tid": 870,
    "ts": 7429770601178.006, "dur": 80.190,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770601178.767, "dur": 1.791,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770601180.911, "dur": 14.591,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770601195.928, "dur": 15.716,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770601224.404, "dur": 11.860,
    "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": 870, "tid": 870,
    "ts": 7429770601246.495, "dur": 0.400,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770601247.336, "dur": 1.340,
    "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": 870, "tid": 870,
    "ts": 7429770601260.824, "dur": 0.421,
    "args": {
      "Python parent id": 397, "Python id": 504, "finished": true, "Ev Idx": 7334
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_0", "pid": 870, "tid": 870,
    "ts": 7429770601265.543, "dur": 188.487,
    "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": 870, "tid": 870,
    "ts": 7429770601266.097, "dur": 187.736,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770601266.320, "dur": 0.256,
    "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": 870, "tid": 870,
    "ts": 7429770601268.928, "dur": 184.783,
    "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": 870, "tid": 870,
    "ts": 7429770601269.547, "dur": 0.520,
    "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": 870, "tid": 870,
    "ts": 7429770601270.421, "dur": 0.609,
    "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": 870, "tid": 870,
    "ts": 7429770601271.532, "dur": 0.282,
    "args": {
      "Python parent id": 508, "Python id": 511, "finished": true, "Ev Idx": 7341
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_4", "pid": 870, "tid": 870,
    "ts": 7429770601274.596, "dur": 47.303,
    "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": 870, "tid": 870,
    "ts": 7429770601275.183, "dur": 46.451,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770601275.312, "dur": 0.343,
    "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": 870, "tid": 870,
    "ts": 7429770601276.574, "dur": 44.949,
    "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": 870, "tid": 870,
    "ts": 7429770601276.895, "dur": 0.137,
    "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": 870, "tid": 870,
    "ts": 7429770601277.403, "dur": 0.145,
    "args": {
      "Python parent id": 515, "Python id": 517, "finished": true, "Ev Idx": 7347
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770601277.725, "dur": 43.743,
    "args": {
      "Python parent id": 515, "Python id": 518, "finished": true, "Ev Idx": 7348
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_0", "pid": 870, "tid": 870,
    "ts": 7429770601325.301, "dur": 21.384,
    "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": 870, "tid": 870,
    "ts": 7429770601326.127, "dur": 20.365,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770601326.345, "dur": 0.236,
    "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": 870, "tid": 870,
    "ts": 7429770601328.619, "dur": 17.773,
    "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": 870, "tid": 870,
    "ts": 7429770601330.609, "dur": 15.639,
    "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": 870, "tid": 870,
    "ts": 7429770601330.822, "dur": 1.103,
    "args": {
      "Python parent id": 523, "Python id": 524, "finished": true, "Ev Idx": 7354
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770601334.092, "dur": 12.066,
    "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": 870, "tid": 870,
    "ts": 7429770601347.400, "dur": 0.386,
    "args": {
      "Python parent id": 508, "Python id": 526, "finished": true, "Ev Idx": 7356
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_5", "pid": 870, "tid": 870,
    "ts": 7429770601349.952, "dur": 37.954,
    "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": 870, "tid": 870,
    "ts": 7429770601350.337, "dur": 37.378,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770601350.501, "dur": 0.184,
    "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": 870, "tid": 870,
    "ts": 7429770601351.867, "dur": 35.742,
    "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": 870, "tid": 870,
    "ts": 7429770601352.128, "dur": 0.143,
    "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": 870, "tid": 870,
    "ts": 7429770601352.519, "dur": 0.115,
    "args": {
      "Python parent id": 530, "Python id": 532, "finished": true, "Ev Idx": 7362
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770601352.809, "dur": 34.754,
    "args": {
      "Python parent id": 530, "Python id": 533, "finished": true, "Ev Idx": 7363
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_6", "pid": 870, "tid": 870,
    "ts": 7429770601403.046, "dur": 50.371,
    "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": 870, "tid": 870,
    "ts": 7429770601403.778, "dur": 49.422,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770601404.018, "dur": 0.240,
    "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": 870, "tid": 870,
    "ts": 7429770601405.012, "dur": 48.094,
    "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": 870, "tid": 870,
    "ts": 7429770601405.285, "dur": 0.139,
    "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": 870, "tid": 870,
    "ts": 7429770601405.668, "dur": 0.168,
    "args": {
      "Python parent id": 537, "Python id": 539, "finished": true, "Ev Idx": 7369
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770601406.047, "dur": 47.009,
    "args": {
      "Python parent id": 537, "Python id": 540, "finished": true, "Ev Idx": 7370
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_1", "pid": 870, "tid": 870,
    "ts": 7429770601472.933, "dur": 1115.118,
    "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": 870, "tid": 870,
    "ts": 7429770601474.303, "dur": 1112.934,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770601474.540, "dur": 0.271,
    "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": 870, "tid": 870,
    "ts": 7429770601476.179, "dur": 1109.939,
    "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": 870, "tid": 870,
    "ts": 7429770601476.571, "dur": 0.583,
    "args": {
      "Python parent id": 544, "Python id": 545, "finished": true, "Ev Idx": 7375
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_2", "pid": 870, "tid": 870,
    "ts": 7429770601479.895, "dur": 81.810,
    "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": 870, "tid": 870,
    "ts": 7429770601480.405, "dur": 80.901,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770601480.567, "dur": 0.165,
    "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": 870, "tid": 870,
    "ts": 7429770601481.560, "dur": 78.806,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770601482.051, "dur": 1.732,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770601484.040, "dur": 11.265,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770601495.590, "dur": 14.379,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770601527.464, "dur": 9.785,
    "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": 870, "tid": 870,
    "ts": 7429770601549.439, "dur": 0.213,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770601550.109, "dur": 1.127,
    "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": 870, "tid": 870,
    "ts": 7429770601562.351, "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": 870, "tid": 870,
    "ts": 7429770601565.441, "dur": 677.384,
    "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": 870, "tid": 870,
    "ts": 7429770601566.696, "dur": 675.609,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770601566.911, "dur": 0.348,
    "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": 870, "tid": 870,
    "ts": 7429770601569.056, "dur": 672.369,
    "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": 870, "tid": 870,
    "ts": 7429770601571.996, "dur": 0.386,
    "args": {
      "Python parent id": 560, "Python id": 561, "finished": true, "Ev Idx": 7391
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_7", "pid": 870, "tid": 870,
    "ts": 7429770601575.436, "dur": 43.934,
    "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": 870, "tid": 870,
    "ts": 7429770601575.947, "dur": 43.210,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770601576.092, "dur": 0.170,
    "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": 870, "tid": 870,
    "ts": 7429770601577.223, "dur": 41.788,
    "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": 870, "tid": 870,
    "ts": 7429770601577.476, "dur": 0.143,
    "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": 870, "tid": 870,
    "ts": 7429770601577.869, "dur": 0.135,
    "args": {
      "Python parent id": 565, "Python id": 567, "finished": true, "Ev Idx": 7397
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770601578.133, "dur": 40.816,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770601620.023, "dur": 3.397,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770601623.812, "dur": 3.747,
    "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": 870, "tid": 870,
    "ts": 7429770601628.304, "dur": 0.360,
    "args": {
      "Python parent id": 560, "Python id": 571, "finished": true, "Ev Idx": 7401
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_8", "pid": 870, "tid": 870,
    "ts": 7429770601631.802, "dur": 37.645,
    "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": 870, "tid": 870,
    "ts": 7429770601632.249, "dur": 36.975,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770601632.439, "dur": 0.256,
    "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": 870, "tid": 870,
    "ts": 7429770601633.481, "dur": 35.646,
    "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": 870, "tid": 870,
    "ts": 7429770601633.721, "dur": 0.120,
    "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": 870, "tid": 870,
    "ts": 7429770601634.031, "dur": 0.116,
    "args": {
      "Python parent id": 575, "Python id": 577, "finished": true, "Ev Idx": 7407
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770601634.396, "dur": 34.639,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770601669.851, "dur": 2.738,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770601672.988, "dur": 2.760,
    "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": 870, "tid": 870,
    "ts": 7429770601676.589, "dur": 0.307,
    "args": {
      "Python parent id": 560, "Python id": 581, "finished": true, "Ev Idx": 7411
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_9", "pid": 870, "tid": 870,
    "ts": 7429770601679.437, "dur": 37.138,
    "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": 870, "tid": 870,
    "ts": 7429770601680.042, "dur": 36.222,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770601680.210, "dur": 0.234,
    "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": 870, "tid": 870,
    "ts": 7429770601681.072, "dur": 35.074,
    "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": 870, "tid": 870,
    "ts": 7429770601681.303, "dur": 0.123,
    "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": 870, "tid": 870,
    "ts": 7429770601681.622, "dur": 0.131,
    "args": {
      "Python parent id": 585, "Python id": 587, "finished": true, "Ev Idx": 7417
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770601681.859, "dur": 34.253,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770601717.229, "dur": 2.507,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770601720.134, "dur": 2.646,
    "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": 870, "tid": 870,
    "ts": 7429770601724.065, "dur": 143.528,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770601724.317, "dur": 3.629,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770601728.225, "dur": 1.757,
    "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": 870, "tid": 870,
    "ts": 7429770601740.839, "dur": 46.690,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770601766.642, "dur": 20.007,
    "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": 870, "tid": 870,
    "ts": 7429770601819.751, "dur": 30.352,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770601837.001, "dur": 12.128,
    "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": 870, "tid": 870,
    "ts": 7429770601871.760, "dur": 3.586,
    "args": {
      "Python parent id": 560, "Python id": 598, "finished": true, "Ev Idx": 7428
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770601872.449, "dur": 0.258,
    "args": {
      "Python parent id": 598, "Python id": 599, "finished": true, "Ev Idx": 7429
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770601873.043, "dur": 0.151,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770601874.441, "dur": 0.182,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770601875.022, "dur": 0.076,
    "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": 870, "tid": 870,
    "ts": 7429770601876.620, "dur": 6.181,
    "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": 870, "tid": 870,
    "ts": 7429770601879.461, "dur": 3.181,
    "args": {
      "Python parent id": 603, "Python id": 604, "finished": true, "Ev Idx": 7434
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770601879.749, "dur": 1.362,
    "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": 870, "tid": 870,
    "ts": 7429770601880.239, "dur": 0.714,
    "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": 870, "tid": 870,
    "ts": 7429770601881.311, "dur": 0.517,
    "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": 870, "tid": 870,
    "ts": 7429770601882.154, "dur": 0.415,
    "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": 870, "tid": 870,
    "ts": 7429770601885.453, "dur": 293.086,
    "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": 870, "tid": 870,
    "ts": 7429770601886.066, "dur": 45.291,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770601905.830, "dur": 4.754,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770601911.374, "dur": 19.865,
    "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": 870, "tid": 870,
    "ts": 7429770601932.238, "dur": 31.855,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770601946.908, "dur": 2.634,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770601950.001, "dur": 14.030,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770601964.994, "dur": 2.821,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770601968.088, "dur": 47.550,
    "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": 870, "tid": 870,
    "ts": 7429770602048.411, "dur": 19.631,
    "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": 870, "tid": 870,
    "ts": 7429770602048.530, "dur": 0.168,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770602049.256, "dur": 18.621,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770602068.841, "dur": 21.226,
    "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": 870, "tid": 870,
    "ts": 7429770602092.719, "dur": 11.482,
    "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": 870, "tid": 870,
    "ts": 7429770602092.911, "dur": 0.140,
    "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": 870, "tid": 870,
    "ts": 7429770602098.951, "dur": 1.109,
    "args": {
      "Python parent id": 622, "Python id": 624, "finished": true, "Ev Idx": 7454
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770602099.517, "dur": 0.466,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770602100.738, "dur": 3.395,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770602104.852, "dur": 51.983,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770602157.597, "dur": 3.390,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770602161.750, "dur": 15.730,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770602181.672, "dur": 4.074,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770602186.529, "dur": 0.239,
    "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": 870, "tid": 870,
    "ts": 7429770602187.839, "dur": 1.138,
    "args": {
      "Python parent id": 560, "Python id": 632, "finished": true, "Ev Idx": 7462
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_10", "pid": 870, "tid": 870,
    "ts": 7429770602194.436, "dur": 46.314,
    "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": 870, "tid": 870,
    "ts": 7429770602195.938, "dur": 44.400,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770602196.393, "dur": 0.974,
    "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": 870, "tid": 870,
    "ts": 7429770602198.729, "dur": 41.388,
    "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": 870, "tid": 870,
    "ts": 7429770602199.059, "dur": 0.194,
    "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": 870, "tid": 870,
    "ts": 7429770602199.553, "dur": 0.150,
    "args": {
      "Python parent id": 636, "Python id": 638, "finished": true, "Ev Idx": 7468
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770602199.855, "dur": 40.174,
    "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": 870, "tid": 870,
    "ts": 7429770602255.417, "dur": 0.678,
    "args": {
      "Python parent id": 544, "Python id": 640, "finished": true, "Ev Idx": 7470
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_3", "pid": 870, "tid": 870,
    "ts": 7429770602259.009, "dur": 100.867,
    "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": 870, "tid": 870,
    "ts": 7429770602259.605, "dur": 99.975,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770602259.825, "dur": 0.439,
    "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": 870, "tid": 870,
    "ts": 7429770602261.653, "dur": 97.061,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770602262.221, "dur": 1.532,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770602264.028, "dur": 15.420,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770602279.762, "dur": 16.963,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770602319.957, "dur": 11.419,
    "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": 870, "tid": 870,
    "ts": 7429770602346.215, "dur": 0.403,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770602347.125, "dur": 1.329,
    "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": 870, "tid": 870,
    "ts": 7429770602360.745, "dur": 0.452,
    "args": {
      "Python parent id": 544, "Python id": 651, "finished": true, "Ev Idx": 7481
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_1", "pid": 870, "tid": 870,
    "ts": 7429770602363.325, "dur": 212.502,
    "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": 870, "tid": 870,
    "ts": 7429770602364.255, "dur": 211.415,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770602364.512, "dur": 0.465,
    "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": 870, "tid": 870,
    "ts": 7429770602366.356, "dur": 209.196,
    "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": 870, "tid": 870,
    "ts": 7429770602366.848, "dur": 0.679,
    "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": 870, "tid": 870,
    "ts": 7429770602367.981, "dur": 0.297,
    "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": 870, "tid": 870,
    "ts": 7429770602368.517, "dur": 0.207,
    "args": {
      "Python parent id": 655, "Python id": 658, "finished": true, "Ev Idx": 7488
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_11", "pid": 870, "tid": 870,
    "ts": 7429770602371.076, "dur": 47.137,
    "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": 870, "tid": 870,
    "ts": 7429770602371.703, "dur": 46.249,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770602371.837, "dur": 0.160,
    "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": 870, "tid": 870,
    "ts": 7429770602372.927, "dur": 44.896,
    "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": 870, "tid": 870,
    "ts": 7429770602373.200, "dur": 0.131,
    "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": 870, "tid": 870,
    "ts": 7429770602373.559, "dur": 0.147,
    "args": {
      "Python parent id": 662, "Python id": 664, "finished": true, "Ev Idx": 7494
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770602373.821, "dur": 43.939,
    "args": {
      "Python parent id": 662, "Python id": 665, "finished": true, "Ev Idx": 7495
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_1", "pid": 870, "tid": 870,
    "ts": 7429770602420.767, "dur": 49.419,
    "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": 870, "tid": 870,
    "ts": 7429770602421.547, "dur": 48.469,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770602421.775, "dur": 0.230,
    "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": 870, "tid": 870,
    "ts": 7429770602423.101, "dur": 46.808,
    "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": 870, "tid": 870,
    "ts": 7429770602424.020, "dur": 45.717,
    "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": 870, "tid": 870,
    "ts": 7429770602424.256, "dur": 0.198,
    "args": {
      "Python parent id": 670, "Python id": 671, "finished": true, "Ev Idx": 7501
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770602425.294, "dur": 44.330,
    "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": 870, "tid": 870,
    "ts": 7429770602470.807, "dur": 0.339,
    "args": {
      "Python parent id": 655, "Python id": 673, "finished": true, "Ev Idx": 7503
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_12", "pid": 870, "tid": 870,
    "ts": 7429770602473.544, "dur": 35.884,
    "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": 870, "tid": 870,
    "ts": 7429770602474.138, "dur": 35.071,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770602474.342, "dur": 0.218,
    "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": 870, "tid": 870,
    "ts": 7429770602475.380, "dur": 33.735,
    "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": 870, "tid": 870,
    "ts": 7429770602475.639, "dur": 0.151,
    "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": 870, "tid": 870,
    "ts": 7429770602475.982, "dur": 0.116,
    "args": {
      "Python parent id": 677, "Python id": 679, "finished": true, "Ev Idx": 7509
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770602476.204, "dur": 32.861,
    "args": {
      "Python parent id": 677, "Python id": 680, "finished": true, "Ev Idx": 7510
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_13", "pid": 870, "tid": 870,
    "ts": 7429770602521.514, "dur": 53.717,
    "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": 870, "tid": 870,
    "ts": 7429770602522.269, "dur": 52.708,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770602522.478, "dur": 0.236,
    "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": 870, "tid": 870,
    "ts": 7429770602523.765, "dur": 51.093,
    "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": 870, "tid": 870,
    "ts": 7429770602524.018, "dur": 0.121,
    "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": 870, "tid": 870,
    "ts": 7429770602524.348, "dur": 0.134,
    "args": {
      "Python parent id": 684, "Python id": 686, "finished": true, "Ev Idx": 7516
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770602524.604, "dur": 50.186,
    "args": {
      "Python parent id": 684, "Python id": 687, "finished": true, "Ev Idx": 7517
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_2", "pid": 870, "tid": 870,
    "ts": 7429770602595.597, "dur": 1027.940,
    "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": 870, "tid": 870,
    "ts": 7429770602596.866, "dur": 1026.083,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770602597.126, "dur": 0.314,
    "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": 870, "tid": 870,
    "ts": 7429770602598.832, "dur": 1023.039,
    "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": 870, "tid": 870,
    "ts": 7429770602599.239, "dur": 0.688,
    "args": {
      "Python parent id": 691, "Python id": 692, "finished": true, "Ev Idx": 7522
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_4", "pid": 870, "tid": 870,
    "ts": 7429770602602.142, "dur": 79.266,
    "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": 870, "tid": 870,
    "ts": 7429770602602.795, "dur": 78.342,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770602602.928, "dur": 0.178,
    "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": 870, "tid": 870,
    "ts": 7429770602603.865, "dur": 76.471,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770602604.343, "dur": 1.410,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770602605.947, "dur": 12.671,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770602618.953, "dur": 16.144,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770602648.974, "dur": 10.558,
    "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": 870, "tid": 870,
    "ts": 7429770602668.760, "dur": 0.164,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770602669.449, "dur": 1.063,
    "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": 870, "tid": 870,
    "ts": 7429770602682.024, "dur": 0.394,
    "args": {
      "Python parent id": 691, "Python id": 703, "finished": true, "Ev Idx": 7533
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_2", "pid": 870, "tid": 870,
    "ts": 7429770602690.684, "dur": 640.653,
    "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": 870, "tid": 870,
    "ts": 7429770602692.010, "dur": 638.787,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770602692.237, "dur": 0.280,
    "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": 870, "tid": 870,
    "ts": 7429770602693.934, "dur": 636.087,
    "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": 870, "tid": 870,
    "ts": 7429770602697.249, "dur": 0.414,
    "args": {
      "Python parent id": 707, "Python id": 708, "finished": true, "Ev Idx": 7538
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_14", "pid": 870, "tid": 870,
    "ts": 7429770602700.115, "dur": 41.415,
    "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": 870, "tid": 870,
    "ts": 7429770602701.398, "dur": 39.832,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770602701.533, "dur": 0.174,
    "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": 870, "tid": 870,
    "ts": 7429770602702.721, "dur": 38.253,
    "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": 870, "tid": 870,
    "ts": 7429770602703.008, "dur": 0.192,
    "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": 870, "tid": 870,
    "ts": 7429770602703.426, "dur": 0.163,
    "args": {
      "Python parent id": 712, "Python id": 714, "finished": true, "Ev Idx": 7544
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770602703.692, "dur": 37.215,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770602742.100, "dur": 3.207,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770602745.700, "dur": 3.377,
    "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": 870, "tid": 870,
    "ts": 7429770602749.999, "dur": 0.345,
    "args": {
      "Python parent id": 707, "Python id": 718, "finished": true, "Ev Idx": 7548
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_15", "pid": 870, "tid": 870,
    "ts": 7429770602752.583, "dur": 35.783,
    "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": 870, "tid": 870,
    "ts": 7429770602753.112, "dur": 34.946,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770602753.299, "dur": 0.291,
    "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": 870, "tid": 870,
    "ts": 7429770602754.986, "dur": 32.963,
    "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": 870, "tid": 870,
    "ts": 7429770602755.240, "dur": 0.119,
    "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": 870, "tid": 870,
    "ts": 7429770602755.555, "dur": 0.115,
    "args": {
      "Python parent id": 722, "Python id": 724, "finished": true, "Ev Idx": 7554
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770602755.773, "dur": 32.114,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770602788.795, "dur": 2.383,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770602791.586, "dur": 2.655,
    "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": 870, "tid": 870,
    "ts": 7429770602795.368, "dur": 0.285,
    "args": {
      "Python parent id": 707, "Python id": 728, "finished": true, "Ev Idx": 7558
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_16", "pid": 870, "tid": 870,
    "ts": 7429770602797.541, "dur": 35.402,
    "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": 870, "tid": 870,
    "ts": 7429770602797.959, "dur": 34.757,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770602798.185, "dur": 0.218,
    "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": 870, "tid": 870,
    "ts": 7429770602799.106, "dur": 33.385,
    "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": 870, "tid": 870,
    "ts": 7429770602799.305, "dur": 0.154,
    "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": 870, "tid": 870,
    "ts": 7429770602799.753, "dur": 0.124,
    "args": {
      "Python parent id": 732, "Python id": 734, "finished": true, "Ev Idx": 7564
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770602799.982, "dur": 32.450,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770602833.558, "dur": 2.240,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770602836.258, "dur": 2.691,
    "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": 870, "tid": 870,
    "ts": 7429770602840.155, "dur": 146.424,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770602840.358, "dur": 3.802,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770602844.451, "dur": 1.736,
    "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": 870, "tid": 870,
    "ts": 7429770602857.384, "dur": 49.974,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770602885.910, "dur": 20.322,
    "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": 870, "tid": 870,
    "ts": 7429770602939.464, "dur": 28.930,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770602956.385, "dur": 11.141,
    "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": 870, "tid": 870,
    "ts": 7429770602991.045, "dur": 3.720,
    "args": {
      "Python parent id": 707, "Python id": 745, "finished": true, "Ev Idx": 7575
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770602991.729, "dur": 0.183,
    "args": {
      "Python parent id": 745, "Python id": 746, "finished": true, "Ev Idx": 7576
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770602992.209, "dur": 0.156,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770602993.716, "dur": 0.229,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770602994.374, "dur": 0.116,
    "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": 870, "tid": 870,
    "ts": 7429770602996.124, "dur": 6.323,
    "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": 870, "tid": 870,
    "ts": 7429770602999.179, "dur": 3.155,
    "args": {
      "Python parent id": 750, "Python id": 751, "finished": true, "Ev Idx": 7581
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770602999.449, "dur": 1.305,
    "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": 870, "tid": 870,
    "ts": 7429770602999.917, "dur": 0.665,
    "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": 870, "tid": 870,
    "ts": 7429770603000.983, "dur": 0.436,
    "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": 870, "tid": 870,
    "ts": 7429770603001.727, "dur": 0.532,
    "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": 870, "tid": 870,
    "ts": 7429770603004.961, "dur": 265.077,
    "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": 870, "tid": 870,
    "ts": 7429770603005.390, "dur": 47.445,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770603024.050, "dur": 4.755,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770603029.577, "dur": 23.135,
    "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": 870, "tid": 870,
    "ts": 7429770603053.690, "dur": 30.066,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770603067.598, "dur": 2.941,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770603071.006, "dur": 12.685,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770603084.564, "dur": 2.691,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770603087.543, "dur": 48.579,
    "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": 870, "tid": 870,
    "ts": 7429770603170.815, "dur": 19.860,
    "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": 870, "tid": 870,
    "ts": 7429770603170.999, "dur": 0.172,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770603171.819, "dur": 18.676,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770603191.452, "dur": 1.427,
    "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": 870, "tid": 870,
    "ts": 7429770603194.612, "dur": 9.884,
    "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": 870, "tid": 870,
    "ts": 7429770603194.755, "dur": 0.092,
    "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": 870, "tid": 870,
    "ts": 7429770603199.971, "dur": 1.062,
    "args": {
      "Python parent id": 769, "Python id": 771, "finished": true, "Ev Idx": 7601
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770603200.472, "dur": 0.470,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770603201.665, "dur": 2.769,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770603205.208, "dur": 45.185,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770603251.001, "dur": 3.292,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770603254.786, "dur": 14.269,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770603273.043, "dur": 3.197,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770603277.194, "dur": 0.213,
    "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": 870, "tid": 870,
    "ts": 7429770603278.406, "dur": 1.146,
    "args": {
      "Python parent id": 707, "Python id": 779, "finished": true, "Ev Idx": 7609
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_17", "pid": 870, "tid": 870,
    "ts": 7429770603284.691, "dur": 44.593,
    "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": 870, "tid": 870,
    "ts": 7429770603286.469, "dur": 42.353,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770603286.893, "dur": 0.994,
    "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": 870, "tid": 870,
    "ts": 7429770603289.154, "dur": 39.375,
    "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": 870, "tid": 870,
    "ts": 7429770603289.568, "dur": 0.321,
    "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": 870, "tid": 870,
    "ts": 7429770603290.156, "dur": 0.205,
    "args": {
      "Python parent id": 783, "Python id": 785, "finished": true, "Ev Idx": 7615
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770603290.592, "dur": 37.834,
    "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": 870, "tid": 870,
    "ts": 7429770603343.401, "dur": 0.529,
    "args": {
      "Python parent id": 691, "Python id": 787, "finished": true, "Ev Idx": 7617
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_5", "pid": 870, "tid": 870,
    "ts": 7429770603346.829, "dur": 84.734,
    "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": 870, "tid": 870,
    "ts": 7429770603347.390, "dur": 83.883,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770603347.602, "dur": 0.359,
    "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": 870, "tid": 870,
    "ts": 7429770603349.185, "dur": 81.263,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770603349.897, "dur": 1.592,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770603351.765, "dur": 14.570,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770603366.702, "dur": 14.564,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770603394.839, "dur": 12.899,
    "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": 870, "tid": 870,
    "ts": 7429770603418.319, "dur": 0.152,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770603418.912, "dur": 1.390,
    "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": 870, "tid": 870,
    "ts": 7429770603432.440, "dur": 0.400,
    "args": {
      "Python parent id": 691, "Python id": 798, "finished": true, "Ev Idx": 7628
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_2", "pid": 870, "tid": 870,
    "ts": 7429770603434.919, "dur": 177.457,
    "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": 870, "tid": 870,
    "ts": 7429770603435.811, "dur": 176.375,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770603436.008, "dur": 0.334,
    "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": 870, "tid": 870,
    "ts": 7429770603437.558, "dur": 174.537,
    "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": 870, "tid": 870,
    "ts": 7429770603438.019, "dur": 0.568,
    "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": 870, "tid": 870,
    "ts": 7429770603438.963, "dur": 0.339,
    "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": 870, "tid": 870,
    "ts": 7429770603439.524, "dur": 0.228,
    "args": {
      "Python parent id": 802, "Python id": 805, "finished": true, "Ev Idx": 7635
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_18", "pid": 870, "tid": 870,
    "ts": 7429770603442.542, "dur": 42.691,
    "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": 870, "tid": 870,
    "ts": 7429770603443.503, "dur": 41.535,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770603443.799, "dur": 0.185,
    "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": 870, "tid": 870,
    "ts": 7429770603444.829, "dur": 40.025,
    "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": 870, "tid": 870,
    "ts": 7429770603445.137, "dur": 0.140,
    "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": 870, "tid": 870,
    "ts": 7429770603445.524, "dur": 0.193,
    "args": {
      "Python parent id": 809, "Python id": 811, "finished": true, "Ev Idx": 7641
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770603445.848, "dur": 38.900,
    "args": {
      "Python parent id": 809, "Python id": 812, "finished": true, "Ev Idx": 7642
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_2", "pid": 870, "tid": 870,
    "ts": 7429770603487.797, "dur": 16.000,
    "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": 870, "tid": 870,
    "ts": 7429770603488.653, "dur": 14.962,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770603488.849, "dur": 0.214,
    "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": 870, "tid": 870,
    "ts": 7429770603490.221, "dur": 13.292,
    "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": 870, "tid": 870,
    "ts": 7429770603491.035, "dur": 12.337,
    "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": 870, "tid": 870,
    "ts": 7429770603491.172, "dur": 0.231,
    "args": {
      "Python parent id": 817, "Python id": 818, "finished": true, "Ev Idx": 7648
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770603492.096, "dur": 11.189,
    "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": 870, "tid": 870,
    "ts": 7429770603504.478, "dur": 0.378,
    "args": {
      "Python parent id": 802, "Python id": 820, "finished": true, "Ev Idx": 7650
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_19", "pid": 870, "tid": 870,
    "ts": 7429770603507.041, "dur": 40.537,
    "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": 870, "tid": 870,
    "ts": 7429770603507.860, "dur": 39.470,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770603508.013, "dur": 0.184,
    "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": 870, "tid": 870,
    "ts": 7429770603509.243, "dur": 37.994,
    "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": 870, "tid": 870,
    "ts": 7429770603509.562, "dur": 0.170,
    "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": 870, "tid": 870,
    "ts": 7429770603510.063, "dur": 0.140,
    "args": {
      "Python parent id": 824, "Python id": 826, "finished": true, "Ev Idx": 7656
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770603510.304, "dur": 36.881,
    "args": {
      "Python parent id": 824, "Python id": 827, "finished": true, "Ev Idx": 7657
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_20", "pid": 870, "tid": 870,
    "ts": 7429770603561.418, "dur": 50.369,
    "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": 870, "tid": 870,
    "ts": 7429770603561.985, "dur": 49.401,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770603562.239, "dur": 0.257,
    "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": 870, "tid": 870,
    "ts": 7429770603563.117, "dur": 48.128,
    "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": 870, "tid": 870,
    "ts": 7429770603563.494, "dur": 0.152,
    "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": 870, "tid": 870,
    "ts": 7429770603563.844, "dur": 0.122,
    "args": {
      "Python parent id": 831, "Python id": 833, "finished": true, "Ev Idx": 7663
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770603564.084, "dur": 47.128,
    "args": {
      "Python parent id": 831, "Python id": 834, "finished": true, "Ev Idx": 7664
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_3", "pid": 870, "tid": 870,
    "ts": 7429770603629.842, "dur": 982.628,
    "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": 870, "tid": 870,
    "ts": 7429770603631.211, "dur": 980.596,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770603631.434, "dur": 0.263,
    "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": 870, "tid": 870,
    "ts": 7429770603633.163, "dur": 977.915,
    "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": 870, "tid": 870,
    "ts": 7429770603633.488, "dur": 0.522,
    "args": {
      "Python parent id": 838, "Python id": 839, "finished": true, "Ev Idx": 7669
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_6", "pid": 870, "tid": 870,
    "ts": 7429770603640.260, "dur": 71.120,
    "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": 870, "tid": 870,
    "ts": 7429770603640.899, "dur": 70.208,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770603641.066, "dur": 0.162,
    "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": 870, "tid": 870,
    "ts": 7429770603642.143, "dur": 68.169,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770603642.596, "dur": 1.527,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770603644.386, "dur": 10.848,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770603655.672, "dur": 12.419,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770603679.489, "dur": 8.537,
    "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": 870, "tid": 870,
    "ts": 7429770603699.904, "dur": 0.190,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770603700.704, "dur": 1.160,
    "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": 870, "tid": 870,
    "ts": 7429770603711.968, "dur": 0.356,
    "args": {
      "Python parent id": 838, "Python id": 850, "finished": true, "Ev Idx": 7680
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_3", "pid": 870, "tid": 870,
    "ts": 7429770603715.363, "dur": 624.463,
    "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": 870, "tid": 870,
    "ts": 7429770603716.487, "dur": 622.745,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770603716.778, "dur": 0.271,
    "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": 870, "tid": 870,
    "ts": 7429770603718.488, "dur": 619.945,
    "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": 870, "tid": 870,
    "ts": 7429770603721.181, "dur": 0.364,
    "args": {
      "Python parent id": 854, "Python id": 855, "finished": true, "Ev Idx": 7685
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_21", "pid": 870, "tid": 870,
    "ts": 7429770603724.320, "dur": 43.727,
    "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": 870, "tid": 870,
    "ts": 7429770603725.215, "dur": 42.557,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770603725.370, "dur": 0.148,
    "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": 870, "tid": 870,
    "ts": 7429770603726.370, "dur": 41.211,
    "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": 870, "tid": 870,
    "ts": 7429770603726.649, "dur": 0.145,
    "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": 870, "tid": 870,
    "ts": 7429770603727.023, "dur": 0.159,
    "args": {
      "Python parent id": 859, "Python id": 861, "finished": true, "Ev Idx": 7691
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770603727.315, "dur": 40.216,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770603768.671, "dur": 3.094,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770603772.199, "dur": 3.052,
    "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": 870, "tid": 870,
    "ts": 7429770603775.955, "dur": 0.300,
    "args": {
      "Python parent id": 854, "Python id": 865, "finished": true, "Ev Idx": 7695
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_22", "pid": 870, "tid": 870,
    "ts": 7429770603778.452, "dur": 39.263,
    "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": 870, "tid": 870,
    "ts": 7429770603779.026, "dur": 38.400,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770603779.278, "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": 870, "tid": 870,
    "ts": 7429770603780.398, "dur": 36.945,
    "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": 870, "tid": 870,
    "ts": 7429770603780.607, "dur": 0.125,
    "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": 870, "tid": 870,
    "ts": 7429770603780.921, "dur": 0.307,
    "args": {
      "Python parent id": 869, "Python id": 871, "finished": true, "Ev Idx": 7701
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770603781.347, "dur": 35.949,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770603818.215, "dur": 2.368,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770603820.960, "dur": 2.359,
    "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": 870, "tid": 870,
    "ts": 7429770603824.194, "dur": 0.364,
    "args": {
      "Python parent id": 854, "Python id": 875, "finished": true, "Ev Idx": 7705
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_23", "pid": 870, "tid": 870,
    "ts": 7429770603826.606, "dur": 38.811,
    "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": 870, "tid": 870,
    "ts": 7429770603827.200, "dur": 37.994,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770603827.377, "dur": 0.198,
    "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": 870, "tid": 870,
    "ts": 7429770603828.299, "dur": 36.704,
    "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": 870, "tid": 870,
    "ts": 7429770603828.490, "dur": 0.120,
    "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": 870, "tid": 870,
    "ts": 7429770603828.827, "dur": 0.133,
    "args": {
      "Python parent id": 879, "Python id": 881, "finished": true, "Ev Idx": 7711
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770603829.065, "dur": 35.883,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770603866.002, "dur": 2.251,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770603868.753, "dur": 2.449,
    "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": 870, "tid": 870,
    "ts": 7429770603872.428, "dur": 146.874,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770603872.640, "dur": 3.219,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770603876.142, "dur": 1.665,
    "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": 870, "tid": 870,
    "ts": 7429770603889.696, "dur": 45.210,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770603915.039, "dur": 18.793,
    "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": 870, "tid": 870,
    "ts": 7429770603966.613, "dur": 31.776,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770603983.130, "dur": 14.572,
    "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": 870, "tid": 870,
    "ts": 7429770604023.339, "dur": 3.326,
    "args": {
      "Python parent id": 854, "Python id": 892, "finished": true, "Ev Idx": 7722
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770604024.066, "dur": 0.130,
    "args": {
      "Python parent id": 892, "Python id": 893, "finished": true, "Ev Idx": 7723
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770604024.515, "dur": 0.079,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770604025.874, "dur": 0.180,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770604026.397, "dur": 0.073,
    "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": 870, "tid": 870,
    "ts": 7429770604028.040, "dur": 6.087,
    "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": 870, "tid": 870,
    "ts": 7429770604030.869, "dur": 3.138,
    "args": {
      "Python parent id": 897, "Python id": 898, "finished": true, "Ev Idx": 7728
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770604031.164, "dur": 1.305,
    "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": 870, "tid": 870,
    "ts": 7429770604031.593, "dur": 0.657,
    "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": 870, "tid": 870,
    "ts": 7429770604032.770, "dur": 0.464,
    "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": 870, "tid": 870,
    "ts": 7429770604033.523, "dur": 0.406,
    "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": 870, "tid": 870,
    "ts": 7429770604036.643, "dur": 243.903,
    "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": 870, "tid": 870,
    "ts": 7429770604037.051, "dur": 44.473,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770604053.577, "dur": 4.851,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770604059.297, "dur": 22.132,
    "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": 870, "tid": 870,
    "ts": 7429770604082.399, "dur": 31.279,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770604095.551, "dur": 3.969,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770604100.051, "dur": 13.562,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770604114.555, "dur": 2.707,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770604117.589, "dur": 43.485,
    "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": 870, "tid": 870,
    "ts": 7429770604192.670, "dur": 15.430,
    "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": 870, "tid": 870,
    "ts": 7429770604192.839, "dur": 0.158,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770604193.541, "dur": 14.406,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770604208.927, "dur": 1.483,
    "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": 870, "tid": 870,
    "ts": 7429770604212.235, "dur": 8.611,
    "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": 870, "tid": 870,
    "ts": 7429770604212.471, "dur": 0.083,
    "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": 870, "tid": 870,
    "ts": 7429770604217.119, "dur": 0.919,
    "args": {
      "Python parent id": 916, "Python id": 918, "finished": true, "Ev Idx": 7748
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770604217.658, "dur": 0.312,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770604218.577, "dur": 2.200,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770604221.415, "dur": 41.222,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770604263.252, "dur": 2.871,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770604266.574, "dur": 13.000,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770604283.175, "dur": 3.087,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770604286.962, "dur": 0.191,
    "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": 870, "tid": 870,
    "ts": 7429770604288.132, "dur": 0.858,
    "args": {
      "Python parent id": 854, "Python id": 926, "finished": true, "Ev Idx": 7756
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_24", "pid": 870, "tid": 870,
    "ts": 7429770604293.822, "dur": 43.916,
    "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": 870, "tid": 870,
    "ts": 7429770604295.343, "dur": 42.039,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770604295.796, "dur": 0.877,
    "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": 870, "tid": 870,
    "ts": 7429770604297.639, "dur": 39.537,
    "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": 870, "tid": 870,
    "ts": 7429770604297.956, "dur": 0.311,
    "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": 870, "tid": 870,
    "ts": 7429770604298.571, "dur": 0.161,
    "args": {
      "Python parent id": 930, "Python id": 932, "finished": true, "Ev Idx": 7762
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770604298.890, "dur": 38.182,
    "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": 870, "tid": 870,
    "ts": 7429770604351.511, "dur": 0.506,
    "args": {
      "Python parent id": 838, "Python id": 934, "finished": true, "Ev Idx": 7764
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_7", "pid": 870, "tid": 870,
    "ts": 7429770604354.470, "dur": 79.278,
    "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": 870, "tid": 870,
    "ts": 7429770604355.080, "dur": 78.363,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770604355.331, "dur": 0.375,
    "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": 870, "tid": 870,
    "ts": 7429770604356.605, "dur": 76.031,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770604357.237, "dur": 1.466,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770604359.043, "dur": 12.150,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770604371.593, "dur": 14.688,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770604398.459, "dur": 10.238,
    "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": 870, "tid": 870,
    "ts": 7429770604419.716, "dur": 0.441,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770604420.558, "dur": 1.322,
    "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": 870, "tid": 870,
    "ts": 7429770604434.491, "dur": 0.366,
    "args": {
      "Python parent id": 838, "Python id": 945, "finished": true, "Ev Idx": 7775
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_3", "pid": 870, "tid": 870,
    "ts": 7429770604437.347, "dur": 163.869,
    "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": 870, "tid": 870,
    "ts": 7429770604437.870, "dur": 163.148,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770604438.160, "dur": 0.281,
    "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": 870, "tid": 870,
    "ts": 7429770604439.138, "dur": 161.784,
    "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": 870, "tid": 870,
    "ts": 7429770604439.559, "dur": 0.328,
    "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": 870, "tid": 870,
    "ts": 7429770604440.283, "dur": 0.357,
    "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": 870, "tid": 870,
    "ts": 7429770604440.892, "dur": 0.240,
    "args": {
      "Python parent id": 949, "Python id": 952, "finished": true, "Ev Idx": 7782
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_25", "pid": 870, "tid": 870,
    "ts": 7429770604443.223, "dur": 42.925,
    "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": 870, "tid": 870,
    "ts": 7429770604443.682, "dur": 42.166,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770604443.809, "dur": 0.339,
    "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": 870, "tid": 870,
    "ts": 7429770604444.748, "dur": 40.993,
    "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": 870, "tid": 870,
    "ts": 7429770604444.943, "dur": 0.115,
    "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": 870, "tid": 870,
    "ts": 7429770604445.302, "dur": 0.164,
    "args": {
      "Python parent id": 956, "Python id": 958, "finished": true, "Ev Idx": 7788
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770604445.579, "dur": 40.113,
    "args": {
      "Python parent id": 956, "Python id": 959, "finished": true, "Ev Idx": 7789
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_3", "pid": 870, "tid": 870,
    "ts": 7429770604488.597, "dur": 14.866,
    "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": 870, "tid": 870,
    "ts": 7429770604489.269, "dur": 13.905,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770604489.472, "dur": 0.241,
    "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": 870, "tid": 870,
    "ts": 7429770604490.785, "dur": 12.280,
    "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": 870, "tid": 870,
    "ts": 7429770604491.525, "dur": 11.414,
    "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": 870, "tid": 870,
    "ts": 7429770604491.643, "dur": 0.172,
    "args": {
      "Python parent id": 964, "Python id": 965, "finished": true, "Ev Idx": 7795
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770604492.426, "dur": 10.418,
    "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": 870, "tid": 870,
    "ts": 7429770604504.175, "dur": 0.337,
    "args": {
      "Python parent id": 949, "Python id": 967, "finished": true, "Ev Idx": 7797
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_26", "pid": 870, "tid": 870,
    "ts": 7429770604507.407, "dur": 37.128,
    "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": 870, "tid": 870,
    "ts": 7429770604507.885, "dur": 36.453,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770604508.053, "dur": 0.186,
    "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": 870, "tid": 870,
    "ts": 7429770604508.987, "dur": 35.252,
    "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": 870, "tid": 870,
    "ts": 7429770604509.185, "dur": 0.139,
    "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": 870, "tid": 870,
    "ts": 7429770604509.513, "dur": 0.128,
    "args": {
      "Python parent id": 971, "Python id": 973, "finished": true, "Ev Idx": 7803
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770604509.755, "dur": 34.434,
    "args": {
      "Python parent id": 971, "Python id": 974, "finished": true, "Ev Idx": 7804
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_27", "pid": 870, "tid": 870,
    "ts": 7429770604556.412, "dur": 44.197,
    "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": 870, "tid": 870,
    "ts": 7429770604556.960, "dur": 43.459,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770604557.139, "dur": 0.296,
    "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": 870, "tid": 870,
    "ts": 7429770604558.311, "dur": 41.991,
    "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": 870, "tid": 870,
    "ts": 7429770604558.500, "dur": 0.320,
    "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": 870, "tid": 870,
    "ts": 7429770604559.119, "dur": 0.150,
    "args": {
      "Python parent id": 978, "Python id": 980, "finished": true, "Ev Idx": 7810
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770604559.385, "dur": 40.854,
    "args": {
      "Python parent id": 978, "Python id": 981, "finished": true, "Ev Idx": 7811
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_4", "pid": 870, "tid": 870,
    "ts": 7429770604618.082, "dur": 1252.031,
    "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": 870, "tid": 870,
    "ts": 7429770604619.109, "dur": 1249.948,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770604619.356, "dur": 0.279,
    "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": 870, "tid": 870,
    "ts": 7429770604620.884, "dur": 1247.211,
    "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": 870, "tid": 870,
    "ts": 7429770604621.201, "dur": 0.538,
    "args": {
      "Python parent id": 985, "Python id": 986, "finished": true, "Ev Idx": 7816
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_8", "pid": 870, "tid": 870,
    "ts": 7429770604623.631, "dur": 73.858,
    "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": 870, "tid": 870,
    "ts": 7429770604624.104, "dur": 73.131,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770604624.257, "dur": 0.165,
    "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": 870, "tid": 870,
    "ts": 7429770604625.138, "dur": 71.501,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770604625.550, "dur": 1.628,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770604627.442, "dur": 9.898,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770604637.727, "dur": 14.766,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770604663.953, "dur": 10.739,
    "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": 870, "tid": 870,
    "ts": 7429770604685.298, "dur": 0.160,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770604685.982, "dur": 1.024,
    "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": 870, "tid": 870,
    "ts": 7429770604698.041, "dur": 0.378,
    "args": {
      "Python parent id": 985, "Python id": 997, "finished": true, "Ev Idx": 7827
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_4", "pid": 870, "tid": 870,
    "ts": 7429770604701.156, "dur": 860.643,
    "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": 870, "tid": 870,
    "ts": 7429770604702.515, "dur": 858.303,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770604702.769, "dur": 0.260,
    "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": 870, "tid": 870,
    "ts": 7429770604704.530, "dur": 855.028,
    "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": 870, "tid": 870,
    "ts": 7429770604706.890, "dur": 0.377,
    "args": {
      "Python parent id": 1001, "Python id": 1002, "finished": true, "Ev Idx": 7832
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_28", "pid": 870, "tid": 870,
    "ts": 7429770604709.509, "dur": 180.452,
    "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": 870, "tid": 870,
    "ts": 7429770604709.951, "dur": 179.586,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770604710.090, "dur": 0.146,
    "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": 870, "tid": 870,
    "ts": 7429770604711.205, "dur": 177.954,
    "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": 870, "tid": 870,
    "ts": 7429770604711.508, "dur": 0.130,
    "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": 870, "tid": 870,
    "ts": 7429770604711.835, "dur": 0.121,
    "args": {
      "Python parent id": 1006, "Python id": 1008, "finished": true, "Ev Idx": 7838
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770604712.060, "dur": 176.965,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770604891.191, "dur": 4.980,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770604896.691, "dur": 4.832,
    "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": 870, "tid": 870,
    "ts": 7429770604902.928, "dur": 0.680,
    "args": {
      "Python parent id": 1001, "Python id": 1012, "finished": true, "Ev Idx": 7842
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_29", "pid": 870, "tid": 870,
    "ts": 7429770604907.910, "dur": 50.628,
    "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": 870, "tid": 870,
    "ts": 7429770604908.825, "dur": 49.509,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770604909.279, "dur": 1.032,
    "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": 870, "tid": 870,
    "ts": 7429770604911.377, "dur": 46.845,
    "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": 870, "tid": 870,
    "ts": 7429770604911.821, "dur": 0.173,
    "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": 870, "tid": 870,
    "ts": 7429770604912.287, "dur": 0.166,
    "args": {
      "Python parent id": 1016, "Python id": 1018, "finished": true, "Ev Idx": 7848
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770604912.615, "dur": 45.517,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770604959.071, "dur": 2.729,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770604962.349, "dur": 3.279,
    "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": 870, "tid": 870,
    "ts": 7429770604966.462, "dur": 0.330,
    "args": {
      "Python parent id": 1001, "Python id": 1022, "finished": true, "Ev Idx": 7852
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_30", "pid": 870, "tid": 870,
    "ts": 7429770604970.000, "dur": 38.713,
    "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": 870, "tid": 870,
    "ts": 7429770604970.930, "dur": 37.535,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770604971.119, "dur": 0.259,
    "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": 870, "tid": 870,
    "ts": 7429770604972.367, "dur": 35.778,
    "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": 870, "tid": 870,
    "ts": 7429770604972.637, "dur": 0.123,
    "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": 870, "tid": 870,
    "ts": 7429770604973.000, "dur": 0.133,
    "args": {
      "Python parent id": 1026, "Python id": 1028, "finished": true, "Ev Idx": 7858
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770604973.273, "dur": 34.804,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770605009.299, "dur": 2.840,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770605012.638, "dur": 2.617,
    "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": 870, "tid": 870,
    "ts": 7429770605016.581, "dur": 157.958,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770605017.622, "dur": 3.627,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770605021.546, "dur": 1.771,
    "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": 870, "tid": 870,
    "ts": 7429770605038.150, "dur": 56.706,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770605067.766, "dur": 25.842,
    "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": 870, "tid": 870,
    "ts": 7429770605128.296, "dur": 28.939,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770605145.412, "dur": 10.778,
    "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": 870, "tid": 870,
    "ts": 7429770605179.968, "dur": 4.245,
    "args": {
      "Python parent id": 1001, "Python id": 1039, "finished": true, "Ev Idx": 7869
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770605180.679, "dur": 0.285,
    "args": {
      "Python parent id": 1039, "Python id": 1040, "finished": true, "Ev Idx": 7870
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770605181.384, "dur": 0.097,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770605182.723, "dur": 0.532,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770605183.706, "dur": 0.271,
    "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": 870, "tid": 870,
    "ts": 7429770605185.481, "dur": 8.316,
    "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": 870, "tid": 870,
    "ts": 7429770605189.228, "dur": 4.455,
    "args": {
      "Python parent id": 1044, "Python id": 1045, "finished": true, "Ev Idx": 7875
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770605189.700, "dur": 1.904,
    "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": 870, "tid": 870,
    "ts": 7429770605190.264, "dur": 0.780,
    "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": 870, "tid": 870,
    "ts": 7429770605192.209, "dur": 0.539,
    "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": 870, "tid": 870,
    "ts": 7429770605193.163, "dur": 0.447,
    "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": 870, "tid": 870,
    "ts": 7429770605197.388, "dur": 301.673,
    "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": 870, "tid": 870,
    "ts": 7429770605198.114, "dur": 81.037,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770605249.499, "dur": 5.172,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770605255.505, "dur": 23.522,
    "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": 870, "tid": 870,
    "ts": 7429770605280.069, "dur": 30.533,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770605293.274, "dur": 3.652,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770605297.481, "dur": 13.051,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770605311.502, "dur": 2.814,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770605314.588, "dur": 50.662,
    "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": 870, "tid": 870,
    "ts": 7429770605400.432, "dur": 17.508,
    "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": 870, "tid": 870,
    "ts": 7429770605400.682, "dur": 0.198,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770605401.595, "dur": 16.142,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770605418.897, "dur": 2.338,
    "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": 870, "tid": 870,
    "ts": 7429770605423.030, "dur": 11.312,
    "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": 870, "tid": 870,
    "ts": 7429770605423.197, "dur": 0.121,
    "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": 870, "tid": 870,
    "ts": 7429770605429.611, "dur": 1.091,
    "args": {
      "Python parent id": 1063, "Python id": 1065, "finished": true, "Ev Idx": 7895
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770605430.175, "dur": 0.451,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770605431.211, "dur": 3.063,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770605435.045, "dur": 44.217,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770605479.939, "dur": 3.034,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770605483.622, "dur": 14.435,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770605502.149, "dur": 3.212,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770605506.084, "dur": 0.238,
    "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": 870, "tid": 870,
    "ts": 7429770605507.287, "dur": 1.011,
    "args": {
      "Python parent id": 1001, "Python id": 1073, "finished": true, "Ev Idx": 7903
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_31", "pid": 870, "tid": 870,
    "ts": 7429770605513.811, "dur": 45.201,
    "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": 870, "tid": 870,
    "ts": 7429770605515.049, "dur": 43.538,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770605515.488, "dur": 1.129,
    "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": 870, "tid": 870,
    "ts": 7429770605517.956, "dur": 40.349,
    "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": 870, "tid": 870,
    "ts": 7429770605518.292, "dur": 0.231,
    "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": 870, "tid": 870,
    "ts": 7429770605518.774, "dur": 0.130,
    "args": {
      "Python parent id": 1077, "Python id": 1079, "finished": true, "Ev Idx": 7909
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770605519.082, "dur": 39.113,
    "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": 870, "tid": 870,
    "ts": 7429770605577.271, "dur": 0.581,
    "args": {
      "Python parent id": 985, "Python id": 1081, "finished": true, "Ev Idx": 7911
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_9", "pid": 870, "tid": 870,
    "ts": 7429770605581.295, "dur": 94.406,
    "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": 870, "tid": 870,
    "ts": 7429770605582.048, "dur": 93.201,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770605582.369, "dur": 0.417,
    "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": 870, "tid": 870,
    "ts": 7429770605584.008, "dur": 90.349,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770605584.783, "dur": 1.806,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770605587.095, "dur": 16.796,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770605604.261, "dur": 19.199,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770605637.231, "dur": 12.814,
    "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": 870, "tid": 870,
    "ts": 7429770605661.526, "dur": 0.308,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770605662.363, "dur": 1.158,
    "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": 870, "tid": 870,
    "ts": 7429770605676.579, "dur": 0.448,
    "args": {
      "Python parent id": 985, "Python id": 1092, "finished": true, "Ev Idx": 7922
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_4", "pid": 870, "tid": 870,
    "ts": 7429770605684.411, "dur": 173.728,
    "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": 870, "tid": 870,
    "ts": 7429770605685.035, "dur": 172.898,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770605685.277, "dur": 0.295,
    "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": 870, "tid": 870,
    "ts": 7429770605686.509, "dur": 171.316,
    "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": 870, "tid": 870,
    "ts": 7429770605686.974, "dur": 0.427,
    "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": 870, "tid": 870,
    "ts": 7429770605687.794, "dur": 0.313,
    "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": 870, "tid": 870,
    "ts": 7429770605688.440, "dur": 0.231,
    "args": {
      "Python parent id": 1096, "Python id": 1099, "finished": true, "Ev Idx": 7929
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_32", "pid": 870, "tid": 870,
    "ts": 7429770605691.008, "dur": 40.183,
    "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": 870, "tid": 870,
    "ts": 7429770605691.686, "dur": 39.273,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770605691.833, "dur": 0.192,
    "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": 870, "tid": 870,
    "ts": 7429770605692.942, "dur": 37.870,
    "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": 870, "tid": 870,
    "ts": 7429770605693.214, "dur": 0.148,
    "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": 870, "tid": 870,
    "ts": 7429770605693.603, "dur": 0.131,
    "args": {
      "Python parent id": 1103, "Python id": 1105, "finished": true, "Ev Idx": 7935
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770605693.866, "dur": 36.862,
    "args": {
      "Python parent id": 1103, "Python id": 1106, "finished": true, "Ev Idx": 7936
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_4", "pid": 870, "tid": 870,
    "ts": 7429770605733.280, "dur": 17.935,
    "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": 870, "tid": 870,
    "ts": 7429770605733.972, "dur": 17.049,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770605734.208, "dur": 0.285,
    "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": 870, "tid": 870,
    "ts": 7429770605735.886, "dur": 15.027,
    "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": 870, "tid": 870,
    "ts": 7429770605736.736, "dur": 14.015,
    "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": 870, "tid": 870,
    "ts": 7429770605736.888, "dur": 0.137,
    "args": {
      "Python parent id": 1111, "Python id": 1112, "finished": true, "Ev Idx": 7942
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770605737.844, "dur": 12.824,
    "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": 870, "tid": 870,
    "ts": 7429770605751.868, "dur": 0.341,
    "args": {
      "Python parent id": 1096, "Python id": 1114, "finished": true, "Ev Idx": 7944
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_33", "pid": 870, "tid": 870,
    "ts": 7429770605754.602, "dur": 38.766,
    "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": 870, "tid": 870,
    "ts": 7429770605755.517, "dur": 37.659,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770605755.695, "dur": 0.182,
    "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": 870, "tid": 870,
    "ts": 7429770605756.614, "dur": 36.465,
    "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": 870, "tid": 870,
    "ts": 7429770605756.923, "dur": 0.123,
    "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": 870, "tid": 870,
    "ts": 7429770605757.253, "dur": 0.114,
    "args": {
      "Python parent id": 1118, "Python id": 1120, "finished": true, "Ev Idx": 7950
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770605757.470, "dur": 35.562,
    "args": {
      "Python parent id": 1118, "Python id": 1121, "finished": true, "Ev Idx": 7951
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_34", "pid": 870, "tid": 870,
    "ts": 7429770605806.218, "dur": 51.318,
    "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": 870, "tid": 870,
    "ts": 7429770605807.113, "dur": 50.211,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770605807.354, "dur": 0.284,
    "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": 870, "tid": 870,
    "ts": 7429770605808.353, "dur": 48.846,
    "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": 870, "tid": 870,
    "ts": 7429770605808.575, "dur": 0.150,
    "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": 870, "tid": 870,
    "ts": 7429770605808.915, "dur": 0.132,
    "args": {
      "Python parent id": 1125, "Python id": 1127, "finished": true, "Ev Idx": 7957
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770605809.150, "dur": 48.012,
    "args": {
      "Python parent id": 1125, "Python id": 1128, "finished": true, "Ev Idx": 7958
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_5", "pid": 870, "tid": 870,
    "ts": 7429770605876.980, "dur": 1031.464,
    "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": 870, "tid": 870,
    "ts": 7429770605878.467, "dur": 1029.190,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770605878.663, "dur": 0.339,
    "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": 870, "tid": 870,
    "ts": 7429770605880.448, "dur": 1026.446,
    "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": 870, "tid": 870,
    "ts": 7429770605880.781, "dur": 0.642,
    "args": {
      "Python parent id": 1132, "Python id": 1133, "finished": true, "Ev Idx": 7963
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_10", "pid": 870, "tid": 870,
    "ts": 7429770605883.802, "dur": 71.909,
    "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": 870, "tid": 870,
    "ts": 7429770605884.378, "dur": 71.091,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770605884.543, "dur": 0.157,
    "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": 870, "tid": 870,
    "ts": 7429770605885.616, "dur": 69.033,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770605886.110, "dur": 1.514,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770605887.933, "dur": 10.873,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770605899.078, "dur": 12.612,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770605923.244, "dur": 10.194,
    "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": 870, "tid": 870,
    "ts": 7429770605943.170, "dur": 0.178,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770605943.740, "dur": 1.104,
    "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": 870, "tid": 870,
    "ts": 7429770605956.350, "dur": 0.287,
    "args": {
      "Python parent id": 1132, "Python id": 1144, "finished": true, "Ev Idx": 7974
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_5", "pid": 870, "tid": 870,
    "ts": 7429770605959.578, "dur": 668.506,
    "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": 870, "tid": 870,
    "ts": 7429770605960.819, "dur": 666.457,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770605961.052, "dur": 0.252,
    "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": 870, "tid": 870,
    "ts": 7429770605963.210, "dur": 663.361,
    "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": 870, "tid": 870,
    "ts": 7429770605965.904, "dur": 0.573,
    "args": {
      "Python parent id": 1148, "Python id": 1149, "finished": true, "Ev Idx": 7979
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_35", "pid": 870, "tid": 870,
    "ts": 7429770605969.479, "dur": 41.171,
    "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": 870, "tid": 870,
    "ts": 7429770605969.994, "dur": 40.387,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770605970.144, "dur": 0.165,
    "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": 870, "tid": 870,
    "ts": 7429770605971.113, "dur": 39.158,
    "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": 870, "tid": 870,
    "ts": 7429770605971.438, "dur": 0.121,
    "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": 870, "tid": 870,
    "ts": 7429770605971.783, "dur": 0.126,
    "args": {
      "Python parent id": 1153, "Python id": 1155, "finished": true, "Ev Idx": 7985
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770605972.013, "dur": 38.166,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770606011.114, "dur": 2.989,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770606015.007, "dur": 3.217,
    "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": 870, "tid": 870,
    "ts": 7429770606018.955, "dur": 0.290,
    "args": {
      "Python parent id": 1148, "Python id": 1159, "finished": true, "Ev Idx": 7989
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_36", "pid": 870, "tid": 870,
    "ts": 7429770606021.702, "dur": 35.126,
    "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": 870, "tid": 870,
    "ts": 7429770606022.341, "dur": 34.194,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770606022.554, "dur": 0.252,
    "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": 870, "tid": 870,
    "ts": 7429770606023.773, "dur": 32.670,
    "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": 870, "tid": 870,
    "ts": 7429770606024.011, "dur": 0.167,
    "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": 870, "tid": 870,
    "ts": 7429770606024.381, "dur": 0.135,
    "args": {
      "Python parent id": 1163, "Python id": 1165, "finished": true, "Ev Idx": 7995
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770606024.618, "dur": 31.769,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770606057.284, "dur": 2.396,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770606060.297, "dur": 2.585,
    "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": 870, "tid": 870,
    "ts": 7429770606063.502, "dur": 0.378,
    "args": {
      "Python parent id": 1148, "Python id": 1169, "finished": true, "Ev Idx": 7999
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_37", "pid": 870, "tid": 870,
    "ts": 7429770606066.001, "dur": 64.758,
    "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": 870, "tid": 870,
    "ts": 7429770606066.588, "dur": 63.814,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770606066.765, "dur": 0.242,
    "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": 870, "tid": 870,
    "ts": 7429770606067.952, "dur": 62.211,
    "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": 870, "tid": 870,
    "ts": 7429770606068.224, "dur": 0.173,
    "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": 870, "tid": 870,
    "ts": 7429770606068.587, "dur": 0.119,
    "args": {
      "Python parent id": 1173, "Python id": 1175, "finished": true, "Ev Idx": 8005
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770606068.808, "dur": 61.263,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770606131.561, "dur": 2.991,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770606135.193, "dur": 2.577,
    "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": 870, "tid": 870,
    "ts": 7429770606139.129, "dur": 148.295,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770606139.373, "dur": 3.363,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770606142.999, "dur": 1.783,
    "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": 870, "tid": 870,
    "ts": 7429770606157.273, "dur": 49.093,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770606184.741, "dur": 20.330,
    "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": 870, "tid": 870,
    "ts": 7429770606238.561, "dur": 31.180,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770606255.744, "dur": 13.283,
    "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": 870, "tid": 870,
    "ts": 7429770606291.621, "dur": 3.745,
    "args": {
      "Python parent id": 1148, "Python id": 1186, "finished": true, "Ev Idx": 8016
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770606292.404, "dur": 0.229,
    "args": {
      "Python parent id": 1186, "Python id": 1187, "finished": true, "Ev Idx": 8017
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770606292.981, "dur": 0.089,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770606294.299, "dur": 0.211,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770606295.010, "dur": 0.078,
    "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": 870, "tid": 870,
    "ts": 7429770606296.564, "dur": 6.450,
    "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": 870, "tid": 870,
    "ts": 7429770606299.810, "dur": 3.093,
    "args": {
      "Python parent id": 1191, "Python id": 1192, "finished": true, "Ev Idx": 8022
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770606300.098, "dur": 1.271,
    "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": 870, "tid": 870,
    "ts": 7429770606300.468, "dur": 0.678,
    "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": 870, "tid": 870,
    "ts": 7429770606301.690, "dur": 0.443,
    "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": 870, "tid": 870,
    "ts": 7429770606302.427, "dur": 0.406,
    "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": 870, "tid": 870,
    "ts": 7429770606305.891, "dur": 263.521,
    "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": 870, "tid": 870,
    "ts": 7429770606306.312, "dur": 46.607,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770606324.769, "dur": 5.810,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770606331.408, "dur": 21.394,
    "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": 870, "tid": 870,
    "ts": 7429770606353.755, "dur": 33.810,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770606369.583, "dur": 2.670,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770606372.693, "dur": 14.811,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770606388.519, "dur": 3.092,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770606391.877, "dur": 48.710,
    "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": 870, "tid": 870,
    "ts": 7429770606473.244, "dur": 18.818,
    "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": 870, "tid": 870,
    "ts": 7429770606473.485, "dur": 0.198,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770606474.329, "dur": 17.584,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770606492.981, "dur": 1.678,
    "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": 870, "tid": 870,
    "ts": 7429770606496.506, "dur": 10.224,
    "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": 870, "tid": 870,
    "ts": 7429770606496.603, "dur": 0.142,
    "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": 870, "tid": 870,
    "ts": 7429770606502.493, "dur": 0.983,
    "args": {
      "Python parent id": 1210, "Python id": 1212, "finished": true, "Ev Idx": 8042
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770606502.939, "dur": 0.454,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770606503.963, "dur": 2.701,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770606507.332, "dur": 42.503,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770606550.513, "dur": 3.147,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770606554.263, "dur": 14.249,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770606572.054, "dur": 2.885,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770606575.749, "dur": 0.391,
    "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": 870, "tid": 870,
    "ts": 7429770606577.038, "dur": 0.926,
    "args": {
      "Python parent id": 1148, "Python id": 1220, "finished": true, "Ev Idx": 8050
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_38", "pid": 870, "tid": 870,
    "ts": 7429770606583.446, "dur": 42.553,
    "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": 870, "tid": 870,
    "ts": 7429770606584.715, "dur": 40.914,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770606585.133, "dur": 0.957,
    "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": 870, "tid": 870,
    "ts": 7429770606587.194, "dur": 38.130,
    "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": 870, "tid": 870,
    "ts": 7429770606587.482, "dur": 0.182,
    "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": 870, "tid": 870,
    "ts": 7429770606587.902, "dur": 0.188,
    "args": {
      "Python parent id": 1224, "Python id": 1226, "finished": true, "Ev Idx": 8056
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770606588.253, "dur": 36.987,
    "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": 870, "tid": 870,
    "ts": 7429770606639.334, "dur": 0.543,
    "args": {
      "Python parent id": 1132, "Python id": 1228, "finished": true, "Ev Idx": 8058
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_11", "pid": 870, "tid": 870,
    "ts": 7429770606642.706, "dur": 83.485,
    "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": 870, "tid": 870,
    "ts": 7429770606643.463, "dur": 82.475,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770606643.765, "dur": 0.397,
    "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": 870, "tid": 870,
    "ts": 7429770606645.811, "dur": 79.338,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770606646.427, "dur": 1.643,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770606648.403, "dur": 12.886,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770606661.570, "dur": 16.203,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770606690.179, "dur": 11.930,
    "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": 870, "tid": 870,
    "ts": 7429770606713.069, "dur": 0.208,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770606713.722, "dur": 1.164,
    "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": 870, "tid": 870,
    "ts": 7429770606727.003, "dur": 0.534,
    "args": {
      "Python parent id": 1132, "Python id": 1239, "finished": true, "Ev Idx": 8069
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_5", "pid": 870, "tid": 870,
    "ts": 7429770606729.407, "dur": 166.676,
    "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": 870, "tid": 870,
    "ts": 7429770606729.945, "dur": 165.953,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770606730.228, "dur": 0.317,
    "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": 870, "tid": 870,
    "ts": 7429770606731.371, "dur": 164.428,
    "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": 870, "tid": 870,
    "ts": 7429770606731.858, "dur": 0.315,
    "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": 870, "tid": 870,
    "ts": 7429770606732.593, "dur": 0.296,
    "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": 870, "tid": 870,
    "ts": 7429770606733.124, "dur": 0.226,
    "args": {
      "Python parent id": 1243, "Python id": 1246, "finished": true, "Ev Idx": 8076
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_39", "pid": 870, "tid": 870,
    "ts": 7429770606735.613, "dur": 39.321,
    "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": 870, "tid": 870,
    "ts": 7429770606736.236, "dur": 38.464,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770606736.390, "dur": 0.166,
    "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": 870, "tid": 870,
    "ts": 7429770606737.415, "dur": 37.192,
    "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": 870, "tid": 870,
    "ts": 7429770606737.767, "dur": 0.149,
    "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": 870, "tid": 870,
    "ts": 7429770606738.155, "dur": 0.127,
    "args": {
      "Python parent id": 1250, "Python id": 1252, "finished": true, "Ev Idx": 8082
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770606738.507, "dur": 36.026,
    "args": {
      "Python parent id": 1250, "Python id": 1253, "finished": true, "Ev Idx": 8083
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_5", "pid": 870, "tid": 870,
    "ts": 7429770606777.552, "dur": 15.995,
    "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": 870, "tid": 870,
    "ts": 7429770606778.206, "dur": 15.139,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770606778.528, "dur": 0.250,
    "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": 870, "tid": 870,
    "ts": 7429770606779.881, "dur": 13.369,
    "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": 870, "tid": 870,
    "ts": 7429770606780.687, "dur": 12.367,
    "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": 870, "tid": 870,
    "ts": 7429770606780.831, "dur": 0.196,
    "args": {
      "Python parent id": 1258, "Python id": 1259, "finished": true, "Ev Idx": 8089
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770606781.765, "dur": 11.209,
    "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": 870, "tid": 870,
    "ts": 7429770606794.213, "dur": 0.376,
    "args": {
      "Python parent id": 1243, "Python id": 1261, "finished": true, "Ev Idx": 8091
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_40", "pid": 870, "tid": 870,
    "ts": 7429770606797.012, "dur": 38.459,
    "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": 870, "tid": 870,
    "ts": 7429770606797.508, "dur": 37.757,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770606797.774, "dur": 0.224,
    "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": 870, "tid": 870,
    "ts": 7429770606798.853, "dur": 36.317,
    "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": 870, "tid": 870,
    "ts": 7429770606799.157, "dur": 0.187,
    "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": 870, "tid": 870,
    "ts": 7429770606799.520, "dur": 0.150,
    "args": {
      "Python parent id": 1265, "Python id": 1267, "finished": true, "Ev Idx": 8097
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770606799.775, "dur": 35.341,
    "args": {
      "Python parent id": 1265, "Python id": 1268, "finished": true, "Ev Idx": 8098
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_41", "pid": 870, "tid": 870,
    "ts": 7429770606848.964, "dur": 46.540,
    "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": 870, "tid": 870,
    "ts": 7429770606849.674, "dur": 45.657,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770606849.899, "dur": 0.275,
    "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": 870, "tid": 870,
    "ts": 7429770606850.929, "dur": 44.203,
    "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": 870, "tid": 870,
    "ts": 7429770606851.129, "dur": 0.137,
    "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": 870, "tid": 870,
    "ts": 7429770606851.477, "dur": 0.116,
    "args": {
      "Python parent id": 1272, "Python id": 1274, "finished": true, "Ev Idx": 8104
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770606851.698, "dur": 43.363,
    "args": {
      "Python parent id": 1272, "Python id": 1275, "finished": true, "Ev Idx": 8105
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_6", "pid": 870, "tid": 870,
    "ts": 7429770606914.739, "dur": 1129.680,
    "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": 870, "tid": 870,
    "ts": 7429770606916.060, "dur": 1127.613,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770606916.270, "dur": 0.275,
    "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": 870, "tid": 870,
    "ts": 7429770606917.808, "dur": 1124.657,
    "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": 870, "tid": 870,
    "ts": 7429770606918.246, "dur": 0.692,
    "args": {
      "Python parent id": 1279, "Python id": 1280, "finished": true, "Ev Idx": 8110
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_12", "pid": 870, "tid": 870,
    "ts": 7429770606921.419, "dur": 75.987,
    "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": 870, "tid": 870,
    "ts": 7429770606921.921, "dur": 75.164,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770606922.055, "dur": 0.170,
    "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": 870, "tid": 870,
    "ts": 7429770606923.009, "dur": 73.263,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770606923.481, "dur": 1.540,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770606925.271, "dur": 13.392,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770606938.936, "dur": 12.946,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770606964.167, "dur": 9.178,
    "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": 870, "tid": 870,
    "ts": 7429770606983.631, "dur": 0.276,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770606984.774, "dur": 1.210,
    "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": 870, "tid": 870,
    "ts": 7429770606998.034, "dur": 0.499,
    "args": {
      "Python parent id": 1279, "Python id": 1291, "finished": true, "Ev Idx": 8121
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_6", "pid": 870, "tid": 870,
    "ts": 7429770607001.095, "dur": 630.053,
    "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": 870, "tid": 870,
    "ts": 7429770607002.272, "dur": 628.228,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770607002.484, "dur": 0.208,
    "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": 870, "tid": 870,
    "ts": 7429770607004.885, "dur": 624.768,
    "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": 870, "tid": 870,
    "ts": 7429770607007.928, "dur": 0.703,
    "args": {
      "Python parent id": 1295, "Python id": 1296, "finished": true, "Ev Idx": 8126
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_42", "pid": 870, "tid": 870,
    "ts": 7429770607011.288, "dur": 45.470,
    "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": 870, "tid": 870,
    "ts": 7429770607011.892, "dur": 44.609,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770607012.028, "dur": 0.160,
    "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": 870, "tid": 870,
    "ts": 7429770607012.858, "dur": 43.511,
    "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": 870, "tid": 870,
    "ts": 7429770607013.081, "dur": 0.143,
    "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": 870, "tid": 870,
    "ts": 7429770607013.517, "dur": 0.167,
    "args": {
      "Python parent id": 1300, "Python id": 1302, "finished": true, "Ev Idx": 8132
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770607013.853, "dur": 42.416,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770607057.439, "dur": 3.294,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770607061.336, "dur": 3.604,
    "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": 870, "tid": 870,
    "ts": 7429770607066.012, "dur": 0.374,
    "args": {
      "Python parent id": 1295, "Python id": 1306, "finished": true, "Ev Idx": 8136
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_43", "pid": 870, "tid": 870,
    "ts": 7429770607068.996, "dur": 35.038,
    "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": 870, "tid": 870,
    "ts": 7429770607069.491, "dur": 34.312,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770607069.665, "dur": 0.271,
    "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": 870, "tid": 870,
    "ts": 7429770607070.546, "dur": 33.152,
    "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": 870, "tid": 870,
    "ts": 7429770607070.792, "dur": 0.145,
    "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": 870, "tid": 870,
    "ts": 7429770607071.159, "dur": 0.134,
    "args": {
      "Python parent id": 1310, "Python id": 1312, "finished": true, "Ev Idx": 8142
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770607071.414, "dur": 32.244,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770607104.490, "dur": 2.452,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770607107.443, "dur": 2.405,
    "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": 870, "tid": 870,
    "ts": 7429770607110.506, "dur": 0.336,
    "args": {
      "Python parent id": 1295, "Python id": 1316, "finished": true, "Ev Idx": 8146
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_44", "pid": 870, "tid": 870,
    "ts": 7429770607113.284, "dur": 35.122,
    "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": 870, "tid": 870,
    "ts": 7429770607113.700, "dur": 34.504,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770607113.878, "dur": 0.221,
    "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": 870, "tid": 870,
    "ts": 7429770607114.810, "dur": 33.159,
    "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": 870, "tid": 870,
    "ts": 7429770607115.027, "dur": 0.169,
    "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": 870, "tid": 870,
    "ts": 7429770607115.377, "dur": 0.141,
    "args": {
      "Python parent id": 1320, "Python id": 1322, "finished": true, "Ev Idx": 8152
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770607115.629, "dur": 32.291,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770607148.956, "dur": 2.249,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770607152.110, "dur": 2.727,
    "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": 870, "tid": 870,
    "ts": 7429770607156.049, "dur": 140.790,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770607156.271, "dur": 2.847,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770607159.413, "dur": 1.641,
    "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": 870, "tid": 870,
    "ts": 7429770607172.082, "dur": 42.206,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770607195.490, "dur": 17.721,
    "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": 870, "tid": 870,
    "ts": 7429770607247.302, "dur": 29.478,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770607263.738, "dur": 12.337,
    "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": 870, "tid": 870,
    "ts": 7429770607300.798, "dur": 3.560,
    "args": {
      "Python parent id": 1295, "Python id": 1333, "finished": true, "Ev Idx": 8163
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770607301.409, "dur": 0.180,
    "args": {
      "Python parent id": 1333, "Python id": 1334, "finished": true, "Ev Idx": 8164
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770607301.941, "dur": 0.065,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770607303.367, "dur": 0.270,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770607304.049, "dur": 0.070,
    "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": 870, "tid": 870,
    "ts": 7429770607305.420, "dur": 6.027,
    "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": 870, "tid": 870,
    "ts": 7429770607308.177, "dur": 3.072,
    "args": {
      "Python parent id": 1338, "Python id": 1339, "finished": true, "Ev Idx": 8169
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770607308.575, "dur": 1.124,
    "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": 870, "tid": 870,
    "ts": 7429770607308.949, "dur": 0.589,
    "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": 870, "tid": 870,
    "ts": 7429770607310.025, "dur": 0.437,
    "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": 870, "tid": 870,
    "ts": 7429770607310.783, "dur": 0.394,
    "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": 870, "tid": 870,
    "ts": 7429770607313.929, "dur": 255.473,
    "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": 870, "tid": 870,
    "ts": 7429770607314.480, "dur": 44.847,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770607333.934, "dur": 4.912,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770607339.676, "dur": 19.504,
    "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": 870, "tid": 870,
    "ts": 7429770607360.340, "dur": 34.872,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770607376.206, "dur": 2.829,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770607379.592, "dur": 15.560,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770607396.118, "dur": 2.771,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770607399.141, "dur": 47.597,
    "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": 870, "tid": 870,
    "ts": 7429770607480.502, "dur": 15.539,
    "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": 870, "tid": 870,
    "ts": 7429770607480.676, "dur": 0.189,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770607481.474, "dur": 14.418,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770607496.964, "dur": 1.488,
    "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": 870, "tid": 870,
    "ts": 7429770607500.005, "dur": 9.090,
    "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": 870, "tid": 870,
    "ts": 7429770607500.130, "dur": 0.073,
    "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": 870, "tid": 870,
    "ts": 7429770607505.507, "dur": 0.945,
    "args": {
      "Python parent id": 1357, "Python id": 1359, "finished": true, "Ev Idx": 8189
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770607506.087, "dur": 0.282,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770607506.878, "dur": 2.131,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770607509.781, "dur": 40.734,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770607551.151, "dur": 3.106,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770607554.750, "dur": 13.616,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770607572.264, "dur": 3.073,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770607576.035, "dur": 0.225,
    "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": 870, "tid": 870,
    "ts": 7429770607577.229, "dur": 0.938,
    "args": {
      "Python parent id": 1295, "Python id": 1367, "finished": true, "Ev Idx": 8197
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_45", "pid": 870, "tid": 870,
    "ts": 7429770607582.108, "dur": 46.933,
    "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": 870, "tid": 870,
    "ts": 7429770607582.982, "dur": 45.642,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770607583.347, "dur": 0.738,
    "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": 870, "tid": 870,
    "ts": 7429770607585.344, "dur": 43.084,
    "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": 870, "tid": 870,
    "ts": 7429770607585.760, "dur": 0.217,
    "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": 870, "tid": 870,
    "ts": 7429770607586.239, "dur": 0.154,
    "args": {
      "Python parent id": 1371, "Python id": 1373, "finished": true, "Ev Idx": 8203
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770607586.547, "dur": 41.778,
    "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": 870, "tid": 870,
    "ts": 7429770607674.676, "dur": 0.435,
    "args": {
      "Python parent id": 1279, "Python id": 1375, "finished": true, "Ev Idx": 8205
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_13", "pid": 870, "tid": 870,
    "ts": 7429770607678.071, "dur": 76.405,
    "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": 870, "tid": 870,
    "ts": 7429770607678.717, "dur": 75.506,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770607679.000, "dur": 0.362,
    "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": 870, "tid": 870,
    "ts": 7429770607680.171, "dur": 73.226,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770607680.619, "dur": 1.416,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770607682.352, "dur": 12.969,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770607695.852, "dur": 13.161,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770607721.006, "dur": 10.565,
    "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": 870, "tid": 870,
    "ts": 7429770607742.792, "dur": 0.291,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770607743.571, "dur": 1.105,
    "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": 870, "tid": 870,
    "ts": 7429770607755.236, "dur": 0.400,
    "args": {
      "Python parent id": 1279, "Python id": 1386, "finished": true, "Ev Idx": 8216
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_6", "pid": 870, "tid": 870,
    "ts": 7429770607758.129, "dur": 273.362,
    "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": 870, "tid": 870,
    "ts": 7429770607758.830, "dur": 272.389,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770607759.051, "dur": 0.297,
    "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": 870, "tid": 870,
    "ts": 7429770607760.376, "dur": 270.745,
    "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": 870, "tid": 870,
    "ts": 7429770607760.849, "dur": 0.366,
    "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": 870, "tid": 870,
    "ts": 7429770607761.615, "dur": 0.321,
    "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": 870, "tid": 870,
    "ts": 7429770607762.240, "dur": 0.256,
    "args": {
      "Python parent id": 1390, "Python id": 1393, "finished": true, "Ev Idx": 8223
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_46", "pid": 870, "tid": 870,
    "ts": 7429770607766.204, "dur": 132.008,
    "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": 870, "tid": 870,
    "ts": 7429770607767.488, "dur": 130.475,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770607767.749, "dur": 0.162,
    "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": 870, "tid": 870,
    "ts": 7429770607768.701, "dur": 129.049,
    "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": 870, "tid": 870,
    "ts": 7429770607769.022, "dur": 0.132,
    "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": 870, "tid": 870,
    "ts": 7429770607769.398, "dur": 0.146,
    "args": {
      "Python parent id": 1397, "Python id": 1399, "finished": true, "Ev Idx": 8229
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770607769.646, "dur": 127.963,
    "args": {
      "Python parent id": 1397, "Python id": 1400, "finished": true, "Ev Idx": 8230
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_6", "pid": 870, "tid": 870,
    "ts": 7429770607900.871, "dur": 18.995,
    "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": 870, "tid": 870,
    "ts": 7429770607901.636, "dur": 18.071,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770607901.952, "dur": 0.505,
    "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": 870, "tid": 870,
    "ts": 7429770607903.585, "dur": 15.949,
    "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": 870, "tid": 870,
    "ts": 7429770607904.586, "dur": 14.753,
    "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": 870, "tid": 870,
    "ts": 7429770607904.753, "dur": 0.205,
    "args": {
      "Python parent id": 1405, "Python id": 1406, "finished": true, "Ev Idx": 8236
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770607905.747, "dur": 13.514,
    "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": 870, "tid": 870,
    "ts": 7429770607920.758, "dur": 0.473,
    "args": {
      "Python parent id": 1390, "Python id": 1408, "finished": true, "Ev Idx": 8238
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_47", "pid": 870, "tid": 870,
    "ts": 7429770607924.243, "dur": 42.259,
    "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": 870, "tid": 870,
    "ts": 7429770607924.920, "dur": 41.356,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770607925.078, "dur": 0.275,
    "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": 870, "tid": 870,
    "ts": 7429770607926.102, "dur": 40.081,
    "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": 870, "tid": 870,
    "ts": 7429770607926.386, "dur": 0.145,
    "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": 870, "tid": 870,
    "ts": 7429770607926.791, "dur": 0.145,
    "args": {
      "Python parent id": 1412, "Python id": 1414, "finished": true, "Ev Idx": 8244
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770607927.076, "dur": 39.056,
    "args": {
      "Python parent id": 1412, "Python id": 1415, "finished": true, "Ev Idx": 8245
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_48", "pid": 870, "tid": 870,
    "ts": 7429770607980.040, "dur": 50.727,
    "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": 870, "tid": 870,
    "ts": 7429770607980.754, "dur": 49.817,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770607980.975, "dur": 0.262,
    "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": 870, "tid": 870,
    "ts": 7429770607982.003, "dur": 48.448,
    "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": 870, "tid": 870,
    "ts": 7429770607982.271, "dur": 0.164,
    "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": 870, "tid": 870,
    "ts": 7429770607982.658, "dur": 0.114,
    "args": {
      "Python parent id": 1419, "Python id": 1421, "finished": true, "Ev Idx": 8251
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770607982.874, "dur": 47.528,
    "args": {
      "Python parent id": 1419, "Python id": 1422, "finished": true, "Ev Idx": 8252
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_7", "pid": 870, "tid": 870,
    "ts": 7429770608051.411, "dur": 1026.111,
    "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": 870, "tid": 870,
    "ts": 7429770608052.800, "dur": 1023.860,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770608053.065, "dur": 0.289,
    "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": 870, "tid": 870,
    "ts": 7429770608054.538, "dur": 1021.303,
    "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": 870, "tid": 870,
    "ts": 7429770608054.886, "dur": 0.593,
    "args": {
      "Python parent id": 1426, "Python id": 1427, "finished": true, "Ev Idx": 8257
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_14", "pid": 870, "tid": 870,
    "ts": 7429770608057.821, "dur": 79.465,
    "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": 870, "tid": 870,
    "ts": 7429770608058.412, "dur": 78.564,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770608058.605, "dur": 0.154,
    "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": 870, "tid": 870,
    "ts": 7429770608059.592, "dur": 76.500,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770608060.117, "dur": 1.773,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770608062.176, "dur": 11.848,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770608074.568, "dur": 15.260,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770608105.833, "dur": 9.776,
    "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": 870, "tid": 870,
    "ts": 7429770608125.639, "dur": 0.237,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770608126.438, "dur": 1.189,
    "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": 870, "tid": 870,
    "ts": 7429770608137.881, "dur": 0.354,
    "args": {
      "Python parent id": 1426, "Python id": 1438, "finished": true, "Ev Idx": 8268
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_7", "pid": 870, "tid": 870,
    "ts": 7429770608141.477, "dur": 637.593,
    "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": 870, "tid": 870,
    "ts": 7429770608142.976, "dur": 635.628,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770608143.178, "dur": 0.259,
    "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": 870, "tid": 870,
    "ts": 7429770608145.597, "dur": 632.167,
    "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": 870, "tid": 870,
    "ts": 7429770608149.164, "dur": 0.528,
    "args": {
      "Python parent id": 1442, "Python id": 1443, "finished": true, "Ev Idx": 8273
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_49", "pid": 870, "tid": 870,
    "ts": 7429770608152.330, "dur": 38.522,
    "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": 870, "tid": 870,
    "ts": 7429770608152.866, "dur": 37.749,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770608153.004, "dur": 0.183,
    "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": 870, "tid": 870,
    "ts": 7429770608153.892, "dur": 36.573,
    "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": 870, "tid": 870,
    "ts": 7429770608154.186, "dur": 0.161,
    "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": 870, "tid": 870,
    "ts": 7429770608154.549, "dur": 0.126,
    "args": {
      "Python parent id": 1447, "Python id": 1449, "finished": true, "Ev Idx": 8279
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770608154.778, "dur": 35.640,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770608191.437, "dur": 3.281,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770608195.304, "dur": 4.012,
    "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": 870, "tid": 870,
    "ts": 7429770608200.412, "dur": 0.357,
    "args": {
      "Python parent id": 1442, "Python id": 1453, "finished": true, "Ev Idx": 8283
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_50", "pid": 870, "tid": 870,
    "ts": 7429770608203.779, "dur": 36.464,
    "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": 870, "tid": 870,
    "ts": 7429770608204.233, "dur": 35.784,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770608204.401, "dur": 0.269,
    "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": 870, "tid": 870,
    "ts": 7429770608205.584, "dur": 34.339,
    "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": 870, "tid": 870,
    "ts": 7429770608205.838, "dur": 0.160,
    "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": 870, "tid": 870,
    "ts": 7429770608206.188, "dur": 0.146,
    "args": {
      "Python parent id": 1457, "Python id": 1459, "finished": true, "Ev Idx": 8289
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770608206.436, "dur": 33.440,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770608240.674, "dur": 2.683,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770608243.970, "dur": 2.710,
    "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": 870, "tid": 870,
    "ts": 7429770608247.669, "dur": 0.318,
    "args": {
      "Python parent id": 1442, "Python id": 1463, "finished": true, "Ev Idx": 8293
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_51", "pid": 870, "tid": 870,
    "ts": 7429770608251.186, "dur": 34.428,
    "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": 870, "tid": 870,
    "ts": 7429770608251.586, "dur": 33.800,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770608251.753, "dur": 0.216,
    "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": 870, "tid": 870,
    "ts": 7429770608252.838, "dur": 32.401,
    "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": 870, "tid": 870,
    "ts": 7429770608253.084, "dur": 0.150,
    "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": 870, "tid": 870,
    "ts": 7429770608253.474, "dur": 0.132,
    "args": {
      "Python parent id": 1467, "Python id": 1469, "finished": true, "Ev Idx": 8299
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770608253.715, "dur": 31.475,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770608286.219, "dur": 2.359,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770608289.410, "dur": 2.418,
    "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": 870, "tid": 870,
    "ts": 7429770608293.198, "dur": 146.991,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770608293.485, "dur": 3.478,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770608297.257, "dur": 1.692,
    "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": 870, "tid": 870,
    "ts": 7429770608309.897, "dur": 51.400,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770608335.860, "dur": 24.421,
    "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": 870, "tid": 870,
    "ts": 7429770608392.021, "dur": 28.847,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770608409.058, "dur": 11.149,
    "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": 870, "tid": 870,
    "ts": 7429770608444.206, "dur": 3.791,
    "args": {
      "Python parent id": 1442, "Python id": 1480, "finished": true, "Ev Idx": 8310
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770608445.025, "dur": 0.233,
    "args": {
      "Python parent id": 1480, "Python id": 1481, "finished": true, "Ev Idx": 8311
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770608445.625, "dur": 0.065,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770608446.923, "dur": 0.237,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770608447.589, "dur": 0.121,
    "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": 870, "tid": 870,
    "ts": 7429770608449.306, "dur": 7.681,
    "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": 870, "tid": 870,
    "ts": 7429770608453.421, "dur": 3.431,
    "args": {
      "Python parent id": 1485, "Python id": 1486, "finished": true, "Ev Idx": 8316
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770608453.849, "dur": 1.369,
    "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": 870, "tid": 870,
    "ts": 7429770608454.222, "dur": 0.752,
    "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": 870, "tid": 870,
    "ts": 7429770608455.567, "dur": 0.433,
    "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": 870, "tid": 870,
    "ts": 7429770608456.358, "dur": 0.417,
    "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": 870, "tid": 870,
    "ts": 7429770608459.937, "dur": 255.769,
    "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": 870, "tid": 870,
    "ts": 7429770608460.383, "dur": 46.200,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770608479.005, "dur": 4.985,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770608484.861, "dur": 21.577,
    "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": 870, "tid": 870,
    "ts": 7429770608507.570, "dur": 30.969,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770608521.650, "dur": 3.630,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770608525.791, "dur": 12.684,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770608539.350, "dur": 2.957,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770608542.555, "dur": 47.722,
    "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": 870, "tid": 870,
    "ts": 7429770608622.384, "dur": 16.428,
    "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": 870, "tid": 870,
    "ts": 7429770608622.632, "dur": 0.197,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770608623.473, "dur": 15.187,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770608639.559, "dur": 1.434,
    "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": 870, "tid": 870,
    "ts": 7429770608642.547, "dur": 10.077,
    "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": 870, "tid": 870,
    "ts": 7429770608642.699, "dur": 0.131,
    "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": 870, "tid": 870,
    "ts": 7429770608648.250, "dur": 1.057,
    "args": {
      "Python parent id": 1504, "Python id": 1506, "finished": true, "Ev Idx": 8336
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770608648.825, "dur": 0.397,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770608649.783, "dur": 2.766,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770608653.258, "dur": 42.552,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770608696.451, "dur": 3.216,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770608700.186, "dur": 14.456,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770608718.485, "dur": 3.298,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770608722.563, "dur": 0.269,
    "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": 870, "tid": 870,
    "ts": 7429770608723.808, "dur": 0.896,
    "args": {
      "Python parent id": 1442, "Python id": 1514, "finished": true, "Ev Idx": 8344
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_52", "pid": 870, "tid": 870,
    "ts": 7429770608729.794, "dur": 47.251,
    "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": 870, "tid": 870,
    "ts": 7429770608730.698, "dur": 45.887,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770608731.146, "dur": 0.957,
    "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": 870, "tid": 870,
    "ts": 7429770608733.476, "dur": 42.755,
    "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": 870, "tid": 870,
    "ts": 7429770608733.857, "dur": 0.251,
    "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": 870, "tid": 870,
    "ts": 7429770608734.412, "dur": 0.125,
    "args": {
      "Python parent id": 1518, "Python id": 1520, "finished": true, "Ev Idx": 8350
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770608734.742, "dur": 41.382,
    "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": 870, "tid": 870,
    "ts": 7429770608791.473, "dur": 0.687,
    "args": {
      "Python parent id": 1426, "Python id": 1522, "finished": true, "Ev Idx": 8352
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_15", "pid": 870, "tid": 870,
    "ts": 7429770608795.311, "dur": 80.766,
    "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": 870, "tid": 870,
    "ts": 7429770608796.173, "dur": 79.614,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770608796.439, "dur": 0.413,
    "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": 870, "tid": 870,
    "ts": 7429770608798.050, "dur": 76.811,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770608798.627, "dur": 1.845,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770608800.791, "dur": 12.211,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770608813.381, "dur": 14.991,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770608840.490, "dur": 12.425,
    "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": 870, "tid": 870,
    "ts": 7429770608863.749, "dur": 0.257,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770608864.305, "dur": 1.127,
    "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": 870, "tid": 870,
    "ts": 7429770608876.943, "dur": 0.410,
    "args": {
      "Python parent id": 1426, "Python id": 1533, "finished": true, "Ev Idx": 8363
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_7", "pid": 870, "tid": 870,
    "ts": 7429770608879.536, "dur": 184.071,
    "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": 870, "tid": 870,
    "ts": 7429770608880.081, "dur": 183.295,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770608880.323, "dur": 0.319,
    "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": 870, "tid": 870,
    "ts": 7429770608881.658, "dur": 181.575,
    "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": 870, "tid": 870,
    "ts": 7429770608882.156, "dur": 0.513,
    "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": 870, "tid": 870,
    "ts": 7429770608883.095, "dur": 0.315,
    "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": 870, "tid": 870,
    "ts": 7429770608883.615, "dur": 0.228,
    "args": {
      "Python parent id": 1537, "Python id": 1540, "finished": true, "Ev Idx": 8370
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_53", "pid": 870, "tid": 870,
    "ts": 7429770608886.400, "dur": 52.507,
    "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": 870, "tid": 870,
    "ts": 7429770608887.024, "dur": 51.566,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770608887.171, "dur": 0.179,
    "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": 870, "tid": 870,
    "ts": 7429770608887.990, "dur": 50.479,
    "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": 870, "tid": 870,
    "ts": 7429770608888.329, "dur": 0.156,
    "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": 870, "tid": 870,
    "ts": 7429770608888.747, "dur": 0.128,
    "args": {
      "Python parent id": 1544, "Python id": 1546, "finished": true, "Ev Idx": 8376
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770608888.994, "dur": 49.413,
    "args": {
      "Python parent id": 1544, "Python id": 1547, "finished": true, "Ev Idx": 8377
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_7", "pid": 870, "tid": 870,
    "ts": 7429770608940.816, "dur": 15.621,
    "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": 870, "tid": 870,
    "ts": 7429770608941.896, "dur": 14.348,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770608942.053, "dur": 0.248,
    "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": 870, "tid": 870,
    "ts": 7429770608943.188, "dur": 12.935,
    "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": 870, "tid": 870,
    "ts": 7429770608944.163, "dur": 11.836,
    "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": 870, "tid": 870,
    "ts": 7429770608944.247, "dur": 0.201,
    "args": {
      "Python parent id": 1552, "Python id": 1553, "finished": true, "Ev Idx": 8383
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770608945.153, "dur": 10.768,
    "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": 870, "tid": 870,
    "ts": 7429770608957.141, "dur": 0.364,
    "args": {
      "Python parent id": 1537, "Python id": 1555, "finished": true, "Ev Idx": 8385
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_54", "pid": 870, "tid": 870,
    "ts": 7429770608960.179, "dur": 37.393,
    "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": 870, "tid": 870,
    "ts": 7429770608960.715, "dur": 36.653,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770608960.871, "dur": 0.200,
    "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": 870, "tid": 870,
    "ts": 7429770608961.936, "dur": 35.345,
    "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": 870, "tid": 870,
    "ts": 7429770608962.161, "dur": 0.131,
    "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": 870, "tid": 870,
    "ts": 7429770608962.485, "dur": 0.131,
    "args": {
      "Python parent id": 1559, "Python id": 1561, "finished": true, "Ev Idx": 8391
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770608962.910, "dur": 34.301,
    "args": {
      "Python parent id": 1559, "Python id": 1562, "finished": true, "Ev Idx": 8392
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_55", "pid": 870, "tid": 870,
    "ts": 7429770609017.860, "dur": 45.022,
    "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": 870, "tid": 870,
    "ts": 7429770609018.525, "dur": 44.160,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770609018.681, "dur": 0.314,
    "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": 870, "tid": 870,
    "ts": 7429770609019.849, "dur": 42.713,
    "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": 870, "tid": 870,
    "ts": 7429770609020.080, "dur": 0.121,
    "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": 870, "tid": 870,
    "ts": 7429770609020.414, "dur": 0.122,
    "args": {
      "Python parent id": 1566, "Python id": 1568, "finished": true, "Ev Idx": 8398
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770609020.670, "dur": 41.845,
    "args": {
      "Python parent id": 1566, "Python id": 1569, "finished": true, "Ev Idx": 8399
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_8", "pid": 870, "tid": 870,
    "ts": 7429770609083.422, "dur": 1092.091,
    "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": 870, "tid": 870,
    "ts": 7429770609084.649, "dur": 1089.945,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770609084.829, "dur": 0.361,
    "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": 870, "tid": 870,
    "ts": 7429770609086.768, "dur": 1086.656,
    "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": 870, "tid": 870,
    "ts": 7429770609087.144, "dur": 0.565,
    "args": {
      "Python parent id": 1573, "Python id": 1574, "finished": true, "Ev Idx": 8404
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_16", "pid": 870, "tid": 870,
    "ts": 7429770609090.363, "dur": 71.999,
    "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": 870, "tid": 870,
    "ts": 7429770609090.924, "dur": 71.012,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770609091.060, "dur": 0.159,
    "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": 870, "tid": 870,
    "ts": 7429770609091.892, "dur": 69.354,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770609092.323, "dur": 1.730,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770609094.321, "dur": 10.726,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770609105.412, "dur": 11.797,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770609128.159, "dur": 11.992,
    "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": 870, "tid": 870,
    "ts": 7429770609149.534, "dur": 0.300,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770609150.170, "dur": 1.061,
    "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": 870, "tid": 870,
    "ts": 7429770609162.908, "dur": 0.352,
    "args": {
      "Python parent id": 1573, "Python id": 1585, "finished": true, "Ev Idx": 8415
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_8", "pid": 870, "tid": 870,
    "ts": 7429770609166.116, "dur": 614.608,
    "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": 870, "tid": 870,
    "ts": 7429770609167.190, "dur": 613.068,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770609167.395, "dur": 0.272,
    "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": 870, "tid": 870,
    "ts": 7429770609169.581, "dur": 609.952,
    "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": 870, "tid": 870,
    "ts": 7429770609172.068, "dur": 0.468,
    "args": {
      "Python parent id": 1589, "Python id": 1590, "finished": true, "Ev Idx": 8420
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_56", "pid": 870, "tid": 870,
    "ts": 7429770609174.663, "dur": 42.306,
    "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": 870, "tid": 870,
    "ts": 7429770609175.326, "dur": 41.408,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770609175.467, "dur": 0.153,
    "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": 870, "tid": 870,
    "ts": 7429770609176.577, "dur": 40.005,
    "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": 870, "tid": 870,
    "ts": 7429770609176.876, "dur": 0.146,
    "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": 870, "tid": 870,
    "ts": 7429770609177.234, "dur": 0.128,
    "args": {
      "Python parent id": 1594, "Python id": 1596, "finished": true, "Ev Idx": 8426
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770609177.505, "dur": 39.030,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770609217.629, "dur": 3.341,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770609221.353, "dur": 3.186,
    "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": 870, "tid": 870,
    "ts": 7429770609225.555, "dur": 0.321,
    "args": {
      "Python parent id": 1589, "Python id": 1600, "finished": true, "Ev Idx": 8430
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_57", "pid": 870, "tid": 870,
    "ts": 7429770609228.489, "dur": 35.469,
    "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": 870, "tid": 870,
    "ts": 7429770609229.003, "dur": 34.723,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770609229.177, "dur": 0.237,
    "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": 870, "tid": 870,
    "ts": 7429770609230.104, "dur": 33.533,
    "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": 870, "tid": 870,
    "ts": 7429770609230.278, "dur": 0.132,
    "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": 870, "tid": 870,
    "ts": 7429770609230.632, "dur": 0.122,
    "args": {
      "Python parent id": 1604, "Python id": 1606, "finished": true, "Ev Idx": 8436
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770609230.884, "dur": 32.690,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770609264.412, "dur": 2.245,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770609267.036, "dur": 2.506,
    "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": 870, "tid": 870,
    "ts": 7429770609270.237, "dur": 0.356,
    "args": {
      "Python parent id": 1589, "Python id": 1610, "finished": true, "Ev Idx": 8440
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_58", "pid": 870, "tid": 870,
    "ts": 7429770609273.731, "dur": 34.428,
    "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": 870, "tid": 870,
    "ts": 7429770609274.461, "dur": 33.471,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770609274.659, "dur": 0.258,
    "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": 870, "tid": 870,
    "ts": 7429770609275.660, "dur": 32.107,
    "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": 870, "tid": 870,
    "ts": 7429770609275.878, "dur": 0.154,
    "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": 870, "tid": 870,
    "ts": 7429770609276.222, "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": 870, "tid": 870,
    "ts": 7429770609276.456, "dur": 31.264,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770609308.726, "dur": 2.448,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770609311.578, "dur": 2.676,
    "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": 870, "tid": 870,
    "ts": 7429770609315.477, "dur": 143.225,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770609315.685, "dur": 3.060,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770609319.082, "dur": 1.634,
    "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": 870, "tid": 870,
    "ts": 7429770609332.138, "dur": 45.490,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770609356.563, "dur": 19.981,
    "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": 870, "tid": 870,
    "ts": 7429770609408.394, "dur": 31.324,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770609426.096, "dur": 12.708,
    "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": 870, "tid": 870,
    "ts": 7429770609462.596, "dur": 3.830,
    "args": {
      "Python parent id": 1589, "Python id": 1627, "finished": true, "Ev Idx": 8457
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770609463.383, "dur": 0.184,
    "args": {
      "Python parent id": 1627, "Python id": 1628, "finished": true, "Ev Idx": 8458
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770609463.937, "dur": 0.055,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770609465.224, "dur": 0.468,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770609466.077, "dur": 0.121,
    "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": 870, "tid": 870,
    "ts": 7429770609467.430, "dur": 6.058,
    "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": 870, "tid": 870,
    "ts": 7429770609470.293, "dur": 3.081,
    "args": {
      "Python parent id": 1632, "Python id": 1633, "finished": true, "Ev Idx": 8463
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770609470.598, "dur": 1.221,
    "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": 870, "tid": 870,
    "ts": 7429770609470.956, "dur": 0.683,
    "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": 870, "tid": 870,
    "ts": 7429770609472.124, "dur": 0.459,
    "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": 870, "tid": 870,
    "ts": 7429770609472.887, "dur": 0.417,
    "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": 870, "tid": 870,
    "ts": 7429770609476.059, "dur": 247.874,
    "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": 870, "tid": 870,
    "ts": 7429770609476.579, "dur": 45.494,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770609494.852, "dur": 4.469,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770609500.150, "dur": 21.784,
    "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": 870, "tid": 870,
    "ts": 7429770609522.925, "dur": 31.136,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770609537.389, "dur": 3.443,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770609541.330, "dur": 12.662,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770609554.868, "dur": 2.787,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770609557.930, "dur": 46.431,
    "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": 870, "tid": 870,
    "ts": 7429770609636.099, "dur": 14.963,
    "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": 870, "tid": 870,
    "ts": 7429770609636.319, "dur": 0.130,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770609636.951, "dur": 13.941,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770609651.729, "dur": 1.575,
    "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": 870, "tid": 870,
    "ts": 7429770609654.799, "dur": 8.862,
    "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": 870, "tid": 870,
    "ts": 7429770609654.949, "dur": 0.136,
    "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": 870, "tid": 870,
    "ts": 7429770609659.988, "dur": 0.946,
    "args": {
      "Python parent id": 1651, "Python id": 1653, "finished": true, "Ev Idx": 8483
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770609660.504, "dur": 0.333,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770609661.357, "dur": 2.238,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770609664.284, "dur": 40.223,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770609705.067, "dur": 3.078,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770609708.624, "dur": 14.307,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770609726.238, "dur": 3.017,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770609729.967, "dur": 0.239,
    "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": 870, "tid": 870,
    "ts": 7429770609731.094, "dur": 0.855,
    "args": {
      "Python parent id": 1589, "Python id": 1661, "finished": true, "Ev Idx": 8491
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_59", "pid": 870, "tid": 870,
    "ts": 7429770609735.881, "dur": 43.085,
    "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": 870, "tid": 870,
    "ts": 7429770609736.692, "dur": 41.818,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770609737.095, "dur": 0.788,
    "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": 870, "tid": 870,
    "ts": 7429770609738.813, "dur": 39.523,
    "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": 870, "tid": 870,
    "ts": 7429770609739.150, "dur": 0.233,
    "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": 870, "tid": 870,
    "ts": 7429770609739.567, "dur": 0.127,
    "args": {
      "Python parent id": 1665, "Python id": 1667, "finished": true, "Ev Idx": 8497
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770609739.942, "dur": 38.287,
    "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": 870, "tid": 870,
    "ts": 7429770609791.449, "dur": 0.461,
    "args": {
      "Python parent id": 1573, "Python id": 1669, "finished": true, "Ev Idx": 8499
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_17", "pid": 870, "tid": 870,
    "ts": 7429770609794.961, "dur": 77.500,
    "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": 870, "tid": 870,
    "ts": 7429770609795.683, "dur": 76.417,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770609795.911, "dur": 0.302,
    "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": 870, "tid": 870,
    "ts": 7429770609797.189, "dur": 74.069,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770609797.807, "dur": 1.454,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770609799.547, "dur": 11.552,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770609811.528, "dur": 13.872,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770609837.214, "dur": 10.855,
    "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": 870, "tid": 870,
    "ts": 7429770609859.359, "dur": 0.223,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770609859.973, "dur": 1.426,
    "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": 870, "tid": 870,
    "ts": 7429770609873.203, "dur": 0.369,
    "args": {
      "Python parent id": 1573, "Python id": 1680, "finished": true, "Ev Idx": 8510
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_8", "pid": 870, "tid": 870,
    "ts": 7429770609875.634, "dur": 282.845,
    "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": 870, "tid": 870,
    "ts": 7429770609876.206, "dur": 281.910,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770609876.465, "dur": 0.248,
    "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": 870, "tid": 870,
    "ts": 7429770609877.713, "dur": 280.143,
    "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": 870, "tid": 870,
    "ts": 7429770609878.111, "dur": 0.349,
    "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": 870, "tid": 870,
    "ts": 7429770609878.901, "dur": 0.304,
    "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": 870, "tid": 870,
    "ts": 7429770609879.446, "dur": 0.239,
    "args": {
      "Python parent id": 1684, "Python id": 1687, "finished": true, "Ev Idx": 8517
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_60", "pid": 870, "tid": 870,
    "ts": 7429770609881.942, "dur": 39.897,
    "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": 870, "tid": 870,
    "ts": 7429770609882.479, "dur": 39.125,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770609882.620, "dur": 0.197,
    "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": 870, "tid": 870,
    "ts": 7429770609883.373, "dur": 38.130,
    "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": 870, "tid": 870,
    "ts": 7429770609883.621, "dur": 0.131,
    "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": 870, "tid": 870,
    "ts": 7429770609884.063, "dur": 0.132,
    "args": {
      "Python parent id": 1691, "Python id": 1693, "finished": true, "Ev Idx": 8523
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770609884.310, "dur": 37.147,
    "args": {
      "Python parent id": 1691, "Python id": 1694, "finished": true, "Ev Idx": 8524
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_8", "pid": 870, "tid": 870,
    "ts": 7429770609923.726, "dur": 14.886,
    "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": 870, "tid": 870,
    "ts": 7429770609924.171, "dur": 14.251,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770609924.387, "dur": 0.227,
    "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": 870, "tid": 870,
    "ts": 7429770609925.416, "dur": 12.947,
    "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": 870, "tid": 870,
    "ts": 7429770609926.183, "dur": 12.022,
    "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": 870, "tid": 870,
    "ts": 7429770609926.289, "dur": 0.162,
    "args": {
      "Python parent id": 1699, "Python id": 1700, "finished": true, "Ev Idx": 8530
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770609927.133, "dur": 10.981,
    "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": 870, "tid": 870,
    "ts": 7429770609939.255, "dur": 0.350,
    "args": {
      "Python parent id": 1684, "Python id": 1702, "finished": true, "Ev Idx": 8532
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_61", "pid": 870, "tid": 870,
    "ts": 7429770609941.606, "dur": 37.785,
    "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": 870, "tid": 870,
    "ts": 7429770609941.983, "dur": 37.176,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770609942.135, "dur": 0.192,
    "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": 870, "tid": 870,
    "ts": 7429770609943.204, "dur": 35.838,
    "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": 870, "tid": 870,
    "ts": 7429770609943.408, "dur": 0.146,
    "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": 870, "tid": 870,
    "ts": 7429770609943.761, "dur": 0.121,
    "args": {
      "Python parent id": 1706, "Python id": 1708, "finished": true, "Ev Idx": 8538
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770609943.984, "dur": 34.999,
    "args": {
      "Python parent id": 1706, "Python id": 1709, "finished": true, "Ev Idx": 8539
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_62", "pid": 870, "tid": 870,
    "ts": 7429770609990.524, "dur": 166.411,
    "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": 870, "tid": 870,
    "ts": 7429770609991.000, "dur": 165.313,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770609991.198, "dur": 0.255,
    "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": 870, "tid": 870,
    "ts": 7429770609992.344, "dur": 163.256,
    "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": 870, "tid": 870,
    "ts": 7429770609992.523, "dur": 0.148,
    "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": 870, "tid": 870,
    "ts": 7429770609992.857, "dur": 0.117,
    "args": {
      "Python parent id": 1713, "Python id": 1715, "finished": true, "Ev Idx": 8545
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770609993.193, "dur": 162.230,
    "args": {
      "Python parent id": 1713, "Python id": 1716, "finished": true, "Ev Idx": 8546
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_9", "pid": 870, "tid": 870,
    "ts": 7429770610183.894, "dur": 1345.613,
    "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": 870, "tid": 870,
    "ts": 7429770610185.532, "dur": 1342.476,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770610185.916, "dur": 0.730,
    "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": 870, "tid": 870,
    "ts": 7429770610188.175, "dur": 1338.982,
    "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": 870, "tid": 870,
    "ts": 7429770610188.623, "dur": 0.831,
    "args": {
      "Python parent id": 1720, "Python id": 1721, "finished": true, "Ev Idx": 8551
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_18", "pid": 870, "tid": 870,
    "ts": 7429770610192.227, "dur": 87.382,
    "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": 870, "tid": 870,
    "ts": 7429770610192.851, "dur": 86.513,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770610193.052, "dur": 0.207,
    "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": 870, "tid": 870,
    "ts": 7429770610194.204, "dur": 84.437,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770610194.826, "dur": 2.344,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770610197.434, "dur": 15.413,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770610213.250, "dur": 17.640,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770610244.820, "dur": 10.746,
    "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": 870, "tid": 870,
    "ts": 7429770610267.145, "dur": 0.307,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770610267.945, "dur": 1.136,
    "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": 870, "tid": 870,
    "ts": 7429770610280.205, "dur": 0.494,
    "args": {
      "Python parent id": 1720, "Python id": 1732, "finished": true, "Ev Idx": 8562
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_9", "pid": 870, "tid": 870,
    "ts": 7429770610284.053, "dur": 936.122,
    "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": 870, "tid": 870,
    "ts": 7429770610285.850, "dur": 933.374,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770610286.121, "dur": 0.233,
    "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": 870, "tid": 870,
    "ts": 7429770610288.216, "dur": 930.007,
    "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": 870, "tid": 870,
    "ts": 7429770610291.773, "dur": 0.609,
    "args": {
      "Python parent id": 1736, "Python id": 1737, "finished": true, "Ev Idx": 8567
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_63", "pid": 870, "tid": 870,
    "ts": 7429770610294.629, "dur": 48.665,
    "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": 870, "tid": 870,
    "ts": 7429770610295.277, "dur": 47.717,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770610295.429, "dur": 0.155,
    "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": 870, "tid": 870,
    "ts": 7429770610296.570, "dur": 46.300,
    "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": 870, "tid": 870,
    "ts": 7429770610296.918, "dur": 0.137,
    "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": 870, "tid": 870,
    "ts": 7429770610297.270, "dur": 0.137,
    "args": {
      "Python parent id": 1741, "Python id": 1743, "finished": true, "Ev Idx": 8573
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770610297.533, "dur": 45.262,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770610343.853, "dur": 3.548,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770610347.775, "dur": 4.519,
    "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": 870, "tid": 870,
    "ts": 7429770610353.106, "dur": 0.332,
    "args": {
      "Python parent id": 1736, "Python id": 1747, "finished": true, "Ev Idx": 8577
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_64", "pid": 870, "tid": 870,
    "ts": 7429770610356.139, "dur": 245.571,
    "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": 870, "tid": 870,
    "ts": 7429770610356.592, "dur": 244.630,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770610356.841, "dur": 0.258,
    "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": 870, "tid": 870,
    "ts": 7429770610357.804, "dur": 243.142,
    "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": 870, "tid": 870,
    "ts": 7429770610550.903, "dur": 0.321,
    "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": 870, "tid": 870,
    "ts": 7429770610551.634, "dur": 0.169,
    "args": {
      "Python parent id": 1751, "Python id": 1753, "finished": true, "Ev Idx": 8583
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770610552.000, "dur": 48.880,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770610602.529, "dur": 3.352,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770610606.417, "dur": 4.096,
    "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": 870, "tid": 870,
    "ts": 7429770610611.606, "dur": 0.531,
    "args": {
      "Python parent id": 1736, "Python id": 1757, "finished": true, "Ev Idx": 8587
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_65", "pid": 870, "tid": 870,
    "ts": 7429770610616.032, "dur": 42.132,
    "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": 870, "tid": 870,
    "ts": 7429770610616.924, "dur": 40.774,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770610617.304, "dur": 0.740,
    "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": 870, "tid": 870,
    "ts": 7429770610619.104, "dur": 38.421,
    "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": 870, "tid": 870,
    "ts": 7429770610619.424, "dur": 0.238,
    "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": 870, "tid": 870,
    "ts": 7429770610619.898, "dur": 0.128,
    "args": {
      "Python parent id": 1761, "Python id": 1763, "finished": true, "Ev Idx": 8593
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770610620.223, "dur": 37.207,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770610658.826, "dur": 2.582,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770610661.834, "dur": 2.753,
    "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": 870, "tid": 870,
    "ts": 7429770610666.013, "dur": 194.975,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770610666.836, "dur": 37.255,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770610704.478, "dur": 1.949,
    "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": 870, "tid": 870,
    "ts": 7429770610721.986, "dur": 55.879,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770610753.069, "dur": 23.105,
    "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": 870, "tid": 870,
    "ts": 7429770610809.221, "dur": 33.140,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770610827.770, "dur": 13.630,
    "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": 870, "tid": 870,
    "ts": 7429770610865.886, "dur": 4.115,
    "args": {
      "Python parent id": 1736, "Python id": 1774, "finished": true, "Ev Idx": 8604
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770610866.615, "dur": 0.366,
    "args": {
      "Python parent id": 1774, "Python id": 1775, "finished": true, "Ev Idx": 8605
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770610867.363, "dur": 0.083,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770610868.930, "dur": 0.281,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770610869.632, "dur": 0.126,
    "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": 870, "tid": 870,
    "ts": 7429770610871.588, "dur": 8.026,
    "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": 870, "tid": 870,
    "ts": 7429770610875.690, "dur": 3.734,
    "args": {
      "Python parent id": 1779, "Python id": 1780, "finished": true, "Ev Idx": 8610
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770610875.973, "dur": 1.712,
    "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": 870, "tid": 870,
    "ts": 7429770610876.473, "dur": 0.991,
    "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": 870, "tid": 870,
    "ts": 7429770610878.017, "dur": 0.517,
    "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": 870, "tid": 870,
    "ts": 7429770610878.894, "dur": 0.451,
    "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": 870, "tid": 870,
    "ts": 7429770610883.318, "dur": 274.213,
    "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": 870, "tid": 870,
    "ts": 7429770610883.947, "dur": 50.691,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770610901.732, "dur": 6.471,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770610909.117, "dur": 25.372,
    "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": 870, "tid": 870,
    "ts": 7429770610935.520, "dur": 31.788,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770610950.382, "dur": 2.735,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770610953.615, "dur": 13.626,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770610968.435, "dur": 3.038,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770610971.723, "dur": 50.681,
    "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": 870, "tid": 870,
    "ts": 7429770611058.243, "dur": 19.406,
    "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": 870, "tid": 870,
    "ts": 7429770611058.581, "dur": 0.225,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770611059.610, "dur": 17.887,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770611078.553, "dur": 2.337,
    "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": 870, "tid": 870,
    "ts": 7429770611082.823, "dur": 11.064,
    "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": 870, "tid": 870,
    "ts": 7429770611082.964, "dur": 0.155,
    "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": 870, "tid": 870,
    "ts": 7429770611089.279, "dur": 1.122,
    "args": {
      "Python parent id": 1798, "Python id": 1800, "finished": true, "Ev Idx": 8630
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770611089.863, "dur": 0.458,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770611090.908, "dur": 2.910,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770611094.574, "dur": 43.221,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770611138.487, "dur": 3.172,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770611142.248, "dur": 14.313,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770611160.637, "dur": 3.028,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770611164.507, "dur": 0.233,
    "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": 870, "tid": 870,
    "ts": 7429770611165.844, "dur": 1.246,
    "args": {
      "Python parent id": 1736, "Python id": 1808, "finished": true, "Ev Idx": 8638
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_66", "pid": 870, "tid": 870,
    "ts": 7429770611171.577, "dur": 45.879,
    "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": 870, "tid": 870,
    "ts": 7429770611172.863, "dur": 44.106,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770611173.293, "dur": 1.044,
    "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": 870, "tid": 870,
    "ts": 7429770611175.907, "dur": 40.798,
    "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": 870, "tid": 870,
    "ts": 7429770611176.308, "dur": 0.214,
    "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": 870, "tid": 870,
    "ts": 7429770611176.776, "dur": 0.149,
    "args": {
      "Python parent id": 1812, "Python id": 1814, "finished": true, "Ev Idx": 8644
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770611177.133, "dur": 39.470,
    "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": 870, "tid": 870,
    "ts": 7429770611233.359, "dur": 0.671,
    "args": {
      "Python parent id": 1720, "Python id": 1816, "finished": true, "Ev Idx": 8646
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_19", "pid": 870, "tid": 870,
    "ts": 7429770611244.766, "dur": 88.598,
    "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": 870, "tid": 870,
    "ts": 7429770611245.613, "dur": 87.303,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770611245.880, "dur": 0.432,
    "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": 870, "tid": 870,
    "ts": 7429770611247.558, "dur": 84.513,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770611248.293, "dur": 1.635,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770611250.251, "dur": 16.692,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770611267.316, "dur": 16.959,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770611298.227, "dur": 11.422,
    "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": 870, "tid": 870,
    "ts": 7429770611321.025, "dur": 0.261,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770611321.764, "dur": 1.072,
    "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": 870, "tid": 870,
    "ts": 7429770611334.214, "dur": 0.435,
    "args": {
      "Python parent id": 1720, "Python id": 1827, "finished": true, "Ev Idx": 8657
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_9", "pid": 870, "tid": 870,
    "ts": 7429770611337.015, "dur": 179.747,
    "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": 870, "tid": 870,
    "ts": 7429770611337.577, "dur": 178.946,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770611337.871, "dur": 0.311,
    "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": 870, "tid": 870,
    "ts": 7429770611339.357, "dur": 177.060,
    "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": 870, "tid": 870,
    "ts": 7429770611339.922, "dur": 0.474,
    "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": 870, "tid": 870,
    "ts": 7429770611340.835, "dur": 0.327,
    "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": 870, "tid": 870,
    "ts": 7429770611341.434, "dur": 0.284,
    "args": {
      "Python parent id": 1831, "Python id": 1834, "finished": true, "Ev Idx": 8664
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_67", "pid": 870, "tid": 870,
    "ts": 7429770611344.592, "dur": 44.211,
    "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": 870, "tid": 870,
    "ts": 7429770611345.925, "dur": 42.527,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770611346.061, "dur": 0.178,
    "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": 870, "tid": 870,
    "ts": 7429770611347.074, "dur": 41.240,
    "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": 870, "tid": 870,
    "ts": 7429770611347.408, "dur": 0.139,
    "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": 870, "tid": 870,
    "ts": 7429770611347.801, "dur": 0.144,
    "args": {
      "Python parent id": 1838, "Python id": 1840, "finished": true, "Ev Idx": 8670
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770611348.111, "dur": 40.141,
    "args": {
      "Python parent id": 1838, "Python id": 1841, "finished": true, "Ev Idx": 8671
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_9", "pid": 870, "tid": 870,
    "ts": 7429770611390.876, "dur": 17.916,
    "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": 870, "tid": 870,
    "ts": 7429770611391.584, "dur": 16.902,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770611391.786, "dur": 0.255,
    "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": 870, "tid": 870,
    "ts": 7429770611393.574, "dur": 14.812,
    "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": 870, "tid": 870,
    "ts": 7429770611394.412, "dur": 13.809,
    "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": 870, "tid": 870,
    "ts": 7429770611394.539, "dur": 0.170,
    "args": {
      "Python parent id": 1846, "Python id": 1847, "finished": true, "Ev Idx": 8677
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770611395.549, "dur": 12.575,
    "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": 870, "tid": 870,
    "ts": 7429770611409.428, "dur": 0.331,
    "args": {
      "Python parent id": 1831, "Python id": 1849, "finished": true, "Ev Idx": 8679
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_68", "pid": 870, "tid": 870,
    "ts": 7429770611412.273, "dur": 38.807,
    "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": 870, "tid": 870,
    "ts": 7429770611412.959, "dur": 37.784,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770611413.153, "dur": 0.214,
    "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": 870, "tid": 870,
    "ts": 7429770611414.126, "dur": 36.524,
    "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": 870, "tid": 870,
    "ts": 7429770611414.378, "dur": 0.144,
    "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": 870, "tid": 870,
    "ts": 7429770611414.713, "dur": 0.117,
    "args": {
      "Python parent id": 1853, "Python id": 1855, "finished": true, "Ev Idx": 8685
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770611414.978, "dur": 35.601,
    "args": {
      "Python parent id": 1853, "Python id": 1856, "finished": true, "Ev Idx": 8686
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_69", "pid": 870, "tid": 870,
    "ts": 7429770611464.042, "dur": 52.039,
    "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": 870, "tid": 870,
    "ts": 7429770611464.624, "dur": 51.140,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770611464.822, "dur": 0.269,
    "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": 870, "tid": 870,
    "ts": 7429770611465.841, "dur": 49.800,
    "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": 870, "tid": 870,
    "ts": 7429770611466.058, "dur": 0.160,
    "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": 870, "tid": 870,
    "ts": 7429770611466.428, "dur": 0.141,
    "args": {
      "Python parent id": 1860, "Python id": 1862, "finished": true, "Ev Idx": 8692
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770611466.688, "dur": 48.907,
    "args": {
      "Python parent id": 1860, "Python id": 1863, "finished": true, "Ev Idx": 8693
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_10", "pid": 870, "tid": 870,
    "ts": 7429770611536.213, "dur": 1010.651,
    "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": 870, "tid": 870,
    "ts": 7429770611537.543, "dur": 1008.485,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770611537.712, "dur": 0.301,
    "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": 870, "tid": 870,
    "ts": 7429770611539.247, "dur": 1005.894,
    "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": 870, "tid": 870,
    "ts": 7429770611539.734, "dur": 0.654,
    "args": {
      "Python parent id": 1867, "Python id": 1868, "finished": true, "Ev Idx": 8698
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_20", "pid": 870, "tid": 870,
    "ts": 7429770611542.610, "dur": 74.553,
    "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": 870, "tid": 870,
    "ts": 7429770611543.109, "dur": 73.654,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770611543.250, "dur": 0.149,
    "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": 870, "tid": 870,
    "ts": 7429770611544.310, "dur": 71.771,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770611544.795, "dur": 1.629,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770611546.752, "dur": 11.159,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770611558.231, "dur": 13.884,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770611583.990, "dur": 11.701,
    "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": 870, "tid": 870,
    "ts": 7429770611605.178, "dur": 0.184,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770611605.860, "dur": 0.986,
    "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": 870, "tid": 870,
    "ts": 7429770611617.846, "dur": 0.352,
    "args": {
      "Python parent id": 1867, "Python id": 1879, "finished": true, "Ev Idx": 8709
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_10", "pid": 870, "tid": 870,
    "ts": 7429770611621.274, "dur": 625.115,
    "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": 870, "tid": 870,
    "ts": 7429770611622.905, "dur": 622.917,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770611623.095, "dur": 0.311,
    "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": 870, "tid": 870,
    "ts": 7429770611625.137, "dur": 619.959,
    "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": 870, "tid": 870,
    "ts": 7429770611627.921, "dur": 0.495,
    "args": {
      "Python parent id": 1883, "Python id": 1884, "finished": true, "Ev Idx": 8714
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_70", "pid": 870, "tid": 870,
    "ts": 7429770611630.424, "dur": 42.063,
    "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": 870, "tid": 870,
    "ts": 7429770611631.107, "dur": 41.111,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770611631.244, "dur": 0.163,
    "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": 870, "tid": 870,
    "ts": 7429770611631.999, "dur": 40.106,
    "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": 870, "tid": 870,
    "ts": 7429770611632.293, "dur": 0.173,
    "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": 870, "tid": 870,
    "ts": 7429770611632.695, "dur": 0.112,
    "args": {
      "Python parent id": 1888, "Python id": 1890, "finished": true, "Ev Idx": 8720
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770611632.912, "dur": 39.139,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770611673.147, "dur": 3.262,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770611676.842, "dur": 3.454,
    "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": 870, "tid": 870,
    "ts": 7429770611681.281, "dur": 0.323,
    "args": {
      "Python parent id": 1883, "Python id": 1894, "finished": true, "Ev Idx": 8724
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_71", "pid": 870, "tid": 870,
    "ts": 7429770611683.698, "dur": 36.952,
    "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": 870, "tid": 870,
    "ts": 7429770611684.457, "dur": 35.913,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770611684.633, "dur": 0.269,
    "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": 870, "tid": 870,
    "ts": 7429770611685.593, "dur": 34.679,
    "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": 870, "tid": 870,
    "ts": 7429770611685.806, "dur": 0.145,
    "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": 870, "tid": 870,
    "ts": 7429770611686.149, "dur": 0.124,
    "args": {
      "Python parent id": 1898, "Python id": 1900, "finished": true, "Ev Idx": 8730
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770611686.377, "dur": 33.841,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770611721.108, "dur": 2.395,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770611723.991, "dur": 2.749,
    "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": 870, "tid": 870,
    "ts": 7429770611727.477, "dur": 0.332,
    "args": {
      "Python parent id": 1883, "Python id": 1904, "finished": true, "Ev Idx": 8734
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_72", "pid": 870, "tid": 870,
    "ts": 7429770611729.798, "dur": 36.753,
    "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": 870, "tid": 870,
    "ts": 7429770611730.384, "dur": 35.939,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770611730.551, "dur": 0.231,
    "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": 870, "tid": 870,
    "ts": 7429770611731.475, "dur": 34.684,
    "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": 870, "tid": 870,
    "ts": 7429770611731.774, "dur": 0.121,
    "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": 870, "tid": 870,
    "ts": 7429770611732.120, "dur": 0.156,
    "args": {
      "Python parent id": 1908, "Python id": 1910, "finished": true, "Ev Idx": 8740
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770611732.389, "dur": 33.728,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770611767.108, "dur": 2.358,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770611770.099, "dur": 2.445,
    "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": 870, "tid": 870,
    "ts": 7429770611773.903, "dur": 148.095,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770611774.461, "dur": 2.995,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770611777.702, "dur": 1.739,
    "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": 870, "tid": 870,
    "ts": 7429770611791.408, "dur": 47.512,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770611817.585, "dur": 20.189,
    "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": 870, "tid": 870,
    "ts": 7429770611870.499, "dur": 30.956,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770611889.683, "dur": 10.946,
    "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": 870, "tid": 870,
    "ts": 7429770611925.918, "dur": 3.577,
    "args": {
      "Python parent id": 1883, "Python id": 1921, "finished": true, "Ev Idx": 8751
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770611926.571, "dur": 0.197,
    "args": {
      "Python parent id": 1921, "Python id": 1922, "finished": true, "Ev Idx": 8752
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770611927.107, "dur": 0.093,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770611928.465, "dur": 0.263,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770611929.111, "dur": 0.104,
    "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": 870, "tid": 870,
    "ts": 7429770611930.743, "dur": 6.074,
    "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": 870, "tid": 870,
    "ts": 7429770611933.558, "dur": 3.126,
    "args": {
      "Python parent id": 1926, "Python id": 1927, "finished": true, "Ev Idx": 8757
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770611933.920, "dur": 1.260,
    "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": 870, "tid": 870,
    "ts": 7429770611934.305, "dur": 0.649,
    "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": 870, "tid": 870,
    "ts": 7429770611935.463, "dur": 0.433,
    "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": 870, "tid": 870,
    "ts": 7429770611936.190, "dur": 0.401,
    "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": 870, "tid": 870,
    "ts": 7429770611939.713, "dur": 249.026,
    "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": 870, "tid": 870,
    "ts": 7429770611940.199, "dur": 44.107,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770611959.077, "dur": 4.518,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770611964.347, "dur": 19.844,
    "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": 870, "tid": 870,
    "ts": 7429770611985.131, "dur": 30.737,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770611999.478, "dur": 3.430,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770612003.378, "dur": 12.425,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770612018.118, "dur": 2.892,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770612021.243, "dur": 45.659,
    "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": 870, "tid": 870,
    "ts": 7429770612098.349, "dur": 16.580,
    "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": 870, "tid": 870,
    "ts": 7429770612098.517, "dur": 0.163,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770612099.315, "dur": 15.463,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770612115.783, "dur": 1.873,
    "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": 870, "tid": 870,
    "ts": 7429770612119.191, "dur": 9.258,
    "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": 870, "tid": 870,
    "ts": 7429770612119.346, "dur": 0.137,
    "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": 870, "tid": 870,
    "ts": 7429770612124.501, "dur": 1.011,
    "args": {
      "Python parent id": 1945, "Python id": 1947, "finished": true, "Ev Idx": 8777
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770612125.101, "dur": 0.314,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770612126.051, "dur": 2.325,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770612129.033, "dur": 40.529,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770612170.251, "dur": 3.072,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770612173.936, "dur": 13.815,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770612191.321, "dur": 3.325,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770612195.268, "dur": 0.272,
    "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": 870, "tid": 870,
    "ts": 7429770612196.392, "dur": 0.787,
    "args": {
      "Python parent id": 1883, "Python id": 1955, "finished": true, "Ev Idx": 8785
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_73", "pid": 870, "tid": 870,
    "ts": 7429770612200.874, "dur": 43.630,
    "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": 870, "tid": 870,
    "ts": 7429770612201.841, "dur": 42.282,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770612202.325, "dur": 0.733,
    "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": 870, "tid": 870,
    "ts": 7429770612204.135, "dur": 39.742,
    "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": 870, "tid": 870,
    "ts": 7429770612204.473, "dur": 0.208,
    "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": 870, "tid": 870,
    "ts": 7429770612204.902, "dur": 0.131,
    "args": {
      "Python parent id": 1959, "Python id": 1961, "finished": true, "Ev Idx": 8791
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770612205.183, "dur": 38.592,
    "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": 870, "tid": 870,
    "ts": 7429770612257.724, "dur": 0.452,
    "args": {
      "Python parent id": 1867, "Python id": 1963, "finished": true, "Ev Idx": 8793
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_21", "pid": 870, "tid": 870,
    "ts": 7429770612260.313, "dur": 94.484,
    "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": 870, "tid": 870,
    "ts": 7429770612260.943, "dur": 93.434,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770612261.181, "dur": 0.310,
    "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": 870, "tid": 870,
    "ts": 7429770612262.639, "dur": 90.754,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770612263.143, "dur": 1.501,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770612264.930, "dur": 12.149,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770612277.435, "dur": 14.326,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770612313.936, "dur": 13.122,
    "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": 870, "tid": 870,
    "ts": 7429770612340.981, "dur": 0.381,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770612341.799, "dur": 1.422,
    "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": 870, "tid": 870,
    "ts": 7429770612355.608, "dur": 0.428,
    "args": {
      "Python parent id": 1867, "Python id": 1974, "finished": true, "Ev Idx": 8804
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_10", "pid": 870, "tid": 870,
    "ts": 7429770612358.218, "dur": 176.940,
    "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": 870, "tid": 870,
    "ts": 7429770612358.849, "dur": 176.097,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770612359.104, "dur": 0.472,
    "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": 870, "tid": 870,
    "ts": 7429770612360.542, "dur": 174.288,
    "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": 870, "tid": 870,
    "ts": 7429770612361.002, "dur": 0.333,
    "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": 870, "tid": 870,
    "ts": 7429770612361.679, "dur": 0.322,
    "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": 870, "tid": 870,
    "ts": 7429770612362.239, "dur": 0.236,
    "args": {
      "Python parent id": 1978, "Python id": 1981, "finished": true, "Ev Idx": 8811
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_74", "pid": 870, "tid": 870,
    "ts": 7429770612365.287, "dur": 48.196,
    "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": 870, "tid": 870,
    "ts": 7429770612366.036, "dur": 47.161,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770612366.172, "dur": 0.150,
    "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": 870, "tid": 870,
    "ts": 7429770612367.089, "dur": 45.953,
    "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": 870, "tid": 870,
    "ts": 7429770612367.345, "dur": 0.140,
    "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": 870, "tid": 870,
    "ts": 7429770612367.698, "dur": 0.168,
    "args": {
      "Python parent id": 1985, "Python id": 1987, "finished": true, "Ev Idx": 8817
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770612368.006, "dur": 44.971,
    "args": {
      "Python parent id": 1985, "Python id": 1988, "finished": true, "Ev Idx": 8818
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_10", "pid": 870, "tid": 870,
    "ts": 7429770612415.292, "dur": 15.657,
    "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": 870, "tid": 870,
    "ts": 7429770612416.100, "dur": 14.605,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770612416.331, "dur": 0.240,
    "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": 870, "tid": 870,
    "ts": 7429770612417.704, "dur": 12.918,
    "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": 870, "tid": 870,
    "ts": 7429770612418.523, "dur": 11.934,
    "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": 870, "tid": 870,
    "ts": 7429770612418.640, "dur": 0.192,
    "args": {
      "Python parent id": 1993, "Python id": 1994, "finished": true, "Ev Idx": 8824
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770612419.450, "dur": 10.908,
    "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": 870, "tid": 870,
    "ts": 7429770612431.560, "dur": 0.345,
    "args": {
      "Python parent id": 1978, "Python id": 1996, "finished": true, "Ev Idx": 8826
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_75", "pid": 870, "tid": 870,
    "ts": 7429770612434.024, "dur": 37.545,
    "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": 870, "tid": 870,
    "ts": 7429770612434.448, "dur": 36.935,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770612434.623, "dur": 0.182,
    "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": 870, "tid": 870,
    "ts": 7429770612435.672, "dur": 35.617,
    "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": 870, "tid": 870,
    "ts": 7429770612435.933, "dur": 0.166,
    "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": 870, "tid": 870,
    "ts": 7429770612436.363, "dur": 0.123,
    "args": {
      "Python parent id": 2000, "Python id": 2002, "finished": true, "Ev Idx": 8832
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770612436.605, "dur": 34.609,
    "args": {
      "Python parent id": 2000, "Python id": 2003, "finished": true, "Ev Idx": 8833
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_76", "pid": 870, "tid": 870,
    "ts": 7429770612482.982, "dur": 51.533,
    "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": 870, "tid": 870,
    "ts": 7429770612483.439, "dur": 50.869,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770612483.618, "dur": 0.300,
    "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": 870, "tid": 870,
    "ts": 7429770612484.769, "dur": 49.415,
    "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": 870, "tid": 870,
    "ts": 7429770612484.989, "dur": 0.150,
    "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": 870, "tid": 870,
    "ts": 7429770612485.356, "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": 870, "tid": 870,
    "ts": 7429770612485.609, "dur": 48.535,
    "args": {
      "Python parent id": 2007, "Python id": 2010, "finished": true, "Ev Idx": 8840
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_11", "pid": 870, "tid": 870,
    "ts": 7429770612552.803, "dur": 1158.008,
    "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": 870, "tid": 870,
    "ts": 7429770612554.071, "dur": 1156.001,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770612554.311, "dur": 0.298,
    "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": 870, "tid": 870,
    "ts": 7429770612556.071, "dur": 1153.152,
    "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": 870, "tid": 870,
    "ts": 7429770612556.415, "dur": 0.529,
    "args": {
      "Python parent id": 2014, "Python id": 2015, "finished": true, "Ev Idx": 8845
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_22", "pid": 870, "tid": 870,
    "ts": 7429770612559.296, "dur": 79.071,
    "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": 870, "tid": 870,
    "ts": 7429770612559.892, "dur": 78.118,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770612560.051, "dur": 0.254,
    "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": 870, "tid": 870,
    "ts": 7429770612560.987, "dur": 76.243,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770612561.541, "dur": 1.573,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770612563.437, "dur": 11.698,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770612575.428, "dur": 14.182,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770612603.172, "dur": 11.479,
    "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": 870, "tid": 870,
    "ts": 7429770612625.500, "dur": 0.218,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770612626.176, "dur": 1.194,
    "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": 870, "tid": 870,
    "ts": 7429770612639.031, "dur": 0.419,
    "args": {
      "Python parent id": 2014, "Python id": 2026, "finished": true, "Ev Idx": 8856
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_11", "pid": 870, "tid": 870,
    "ts": 7429770612642.127, "dur": 762.400,
    "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": 870, "tid": 870,
    "ts": 7429770612643.459, "dur": 760.082,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770612643.658, "dur": 0.244,
    "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": 870, "tid": 870,
    "ts": 7429770612645.775, "dur": 756.594,
    "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": 870, "tid": 870,
    "ts": 7429770612648.734, "dur": 0.352,
    "args": {
      "Python parent id": 2030, "Python id": 2031, "finished": true, "Ev Idx": 8861
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_77", "pid": 870, "tid": 870,
    "ts": 7429770612651.385, "dur": 39.818,
    "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": 870, "tid": 870,
    "ts": 7429770612651.871, "dur": 39.145,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770612652.011, "dur": 0.161,
    "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": 870, "tid": 870,
    "ts": 7429770612652.905, "dur": 37.907,
    "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": 870, "tid": 870,
    "ts": 7429770612653.265, "dur": 0.118,
    "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": 870, "tid": 870,
    "ts": 7429770612653.553, "dur": 0.119,
    "args": {
      "Python parent id": 2035, "Python id": 2037, "finished": true, "Ev Idx": 8867
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770612653.774, "dur": 36.952,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770612691.825, "dur": 3.351,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770612695.554, "dur": 3.929,
    "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": 870, "tid": 870,
    "ts": 7429770612700.253, "dur": 0.344,
    "args": {
      "Python parent id": 2030, "Python id": 2041, "finished": true, "Ev Idx": 8871
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_78", "pid": 870, "tid": 870,
    "ts": 7429770612703.266, "dur": 33.408,
    "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": 870, "tid": 870,
    "ts": 7429770612703.804, "dur": 32.635,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770612703.978, "dur": 0.234,
    "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": 870, "tid": 870,
    "ts": 7429770612704.775, "dur": 31.540,
    "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": 870, "tid": 870,
    "ts": 7429770612704.999, "dur": 0.163,
    "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": 870, "tid": 870,
    "ts": 7429770612705.347, "dur": 0.134,
    "args": {
      "Python parent id": 2045, "Python id": 2047, "finished": true, "Ev Idx": 8877
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770612705.585, "dur": 30.679,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770612737.122, "dur": 2.337,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770612739.840, "dur": 2.747,
    "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": 870, "tid": 870,
    "ts": 7429770612743.270, "dur": 0.336,
    "args": {
      "Python parent id": 2030, "Python id": 2051, "finished": true, "Ev Idx": 8881
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_79", "pid": 870, "tid": 870,
    "ts": 7429770612746.297, "dur": 34.376,
    "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": 870, "tid": 870,
    "ts": 7429770612746.901, "dur": 33.564,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770612747.090, "dur": 0.224,
    "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": 870, "tid": 870,
    "ts": 7429770612748.017, "dur": 32.289,
    "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": 870, "tid": 870,
    "ts": 7429770612748.236, "dur": 0.121,
    "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": 870, "tid": 870,
    "ts": 7429770612748.522, "dur": 0.152,
    "args": {
      "Python parent id": 2055, "Python id": 2057, "finished": true, "Ev Idx": 8887
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770612748.781, "dur": 31.487,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770612781.272, "dur": 2.572,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770612784.255, "dur": 4.449,
    "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": 870, "tid": 870,
    "ts": 7429770612789.711, "dur": 142.716,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770612789.948, "dur": 3.255,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770612793.475, "dur": 1.722,
    "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": 870, "tid": 870,
    "ts": 7429770612808.200, "dur": 45.034,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770612832.609, "dur": 19.278,
    "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": 870, "tid": 870,
    "ts": 7429770612881.195, "dur": 32.684,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770612899.260, "dur": 13.647,
    "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": 870, "tid": 870,
    "ts": 7429770612936.716, "dur": 3.414,
    "args": {
      "Python parent id": 2030, "Python id": 2068, "finished": true, "Ev Idx": 8898
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770612937.403, "dur": 0.183,
    "args": {
      "Python parent id": 2068, "Python id": 2069, "finished": true, "Ev Idx": 8899
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770612937.917, "dur": 0.069,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770612939.177, "dur": 0.179,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770612939.783, "dur": 0.084,
    "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": 870, "tid": 870,
    "ts": 7429770612941.319, "dur": 5.955,
    "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": 870, "tid": 870,
    "ts": 7429770612944.052, "dur": 3.116,
    "args": {
      "Python parent id": 2073, "Python id": 2074, "finished": true, "Ev Idx": 8904
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770612944.368, "dur": 1.192,
    "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": 870, "tid": 870,
    "ts": 7429770612944.728, "dur": 0.702,
    "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": 870, "tid": 870,
    "ts": 7429770612945.868, "dur": 0.486,
    "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": 870, "tid": 870,
    "ts": 7429770612946.649, "dur": 0.444,
    "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": 870, "tid": 870,
    "ts": 7429770612950.031, "dur": 382.166,
    "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": 870, "tid": 870,
    "ts": 7429770612950.484, "dur": 45.928,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770612969.043, "dur": 5.590,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770612975.513, "dur": 20.784,
    "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": 870, "tid": 870,
    "ts": 7429770612997.219, "dur": 33.613,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770613013.199, "dur": 3.141,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770613016.818, "dur": 13.954,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770613031.711, "dur": 3.017,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770613035.041, "dur": 48.477,
    "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": 870, "tid": 870,
    "ts": 7429770613117.302, "dur": 16.960,
    "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": 870, "tid": 870,
    "ts": 7429770613117.555, "dur": 0.209,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770613118.353, "dur": 15.745,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770613135.041, "dur": 1.522,
    "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": 870, "tid": 870,
    "ts": 7429770613138.086, "dur": 9.725,
    "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": 870, "tid": 870,
    "ts": 7429770613138.228, "dur": 0.089,
    "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": 870, "tid": 870,
    "ts": 7429770613143.895, "dur": 0.817,
    "args": {
      "Python parent id": 2092, "Python id": 2094, "finished": true, "Ev Idx": 8924
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770613144.356, "dur": 0.288,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770613145.229, "dur": 2.480,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770613148.525, "dur": 160.495,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770613310.000, "dur": 3.968,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770613314.656, "dur": 16.334,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770613336.057, "dur": 3.606,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770613340.438, "dur": 0.285,
    "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": 870, "tid": 870,
    "ts": 7429770613342.123, "dur": 1.264,
    "args": {
      "Python parent id": 2030, "Python id": 2102, "finished": true, "Ev Idx": 8932
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_80", "pid": 870, "tid": 870,
    "ts": 7429770613348.699, "dur": 53.096,
    "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": 870, "tid": 870,
    "ts": 7429770613350.068, "dur": 51.280,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770613350.626, "dur": 1.363,
    "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": 870, "tid": 870,
    "ts": 7429770613353.390, "dur": 47.614,
    "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": 870, "tid": 870,
    "ts": 7429770613353.746, "dur": 0.235,
    "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": 870, "tid": 870,
    "ts": 7429770613354.407, "dur": 0.308,
    "args": {
      "Python parent id": 2106, "Python id": 2108, "finished": true, "Ev Idx": 8938
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770613354.847, "dur": 46.041,
    "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": 870, "tid": 870,
    "ts": 7429770613418.570, "dur": 0.568,
    "args": {
      "Python parent id": 2014, "Python id": 2110, "finished": true, "Ev Idx": 8940
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_23", "pid": 870, "tid": 870,
    "ts": 7429770613422.206, "dur": 93.141,
    "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": 870, "tid": 870,
    "ts": 7429770613422.863, "dur": 92.170,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770613423.141, "dur": 0.416,
    "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": 870, "tid": 870,
    "ts": 7429770613424.719, "dur": 89.516,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770613425.383, "dur": 1.904,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770613427.710, "dur": 16.167,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770613444.287, "dur": 17.037,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770613474.157, "dur": 14.286,
    "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": 870, "tid": 870,
    "ts": 7429770613501.917, "dur": 0.337,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770613502.752, "dur": 1.272,
    "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": 870, "tid": 870,
    "ts": 7429770613516.094, "dur": 0.422,
    "args": {
      "Python parent id": 2014, "Python id": 2121, "finished": true, "Ev Idx": 8951
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_11", "pid": 870, "tid": 870,
    "ts": 7429770613518.651, "dur": 179.430,
    "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": 870, "tid": 870,
    "ts": 7429770613519.230, "dur": 178.581,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770613519.537, "dur": 0.384,
    "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": 870, "tid": 870,
    "ts": 7429770613521.064, "dur": 176.647,
    "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": 870, "tid": 870,
    "ts": 7429770613521.500, "dur": 0.444,
    "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": 870, "tid": 870,
    "ts": 7429770613522.496, "dur": 0.319,
    "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": 870, "tid": 870,
    "ts": 7429770613523.063, "dur": 0.370,
    "args": {
      "Python parent id": 2125, "Python id": 2128, "finished": true, "Ev Idx": 8958
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_81", "pid": 870, "tid": 870,
    "ts": 7429770613525.923, "dur": 45.121,
    "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": 870, "tid": 870,
    "ts": 7429770613526.480, "dur": 44.286,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770613526.621, "dur": 0.191,
    "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": 870, "tid": 870,
    "ts": 7429770613527.994, "dur": 42.594,
    "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": 870, "tid": 870,
    "ts": 7429770613528.404, "dur": 0.156,
    "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": 870, "tid": 870,
    "ts": 7429770613528.763, "dur": 0.142,
    "args": {
      "Python parent id": 2132, "Python id": 2134, "finished": true, "Ev Idx": 8964
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770613529.025, "dur": 41.500,
    "args": {
      "Python parent id": 2132, "Python id": 2135, "finished": true, "Ev Idx": 8965
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_11", "pid": 870, "tid": 870,
    "ts": 7429770613573.506, "dur": 17.644,
    "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": 870, "tid": 870,
    "ts": 7429770613574.302, "dur": 16.625,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770613574.561, "dur": 0.252,
    "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": 870, "tid": 870,
    "ts": 7429770613575.848, "dur": 14.994,
    "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": 870, "tid": 870,
    "ts": 7429770613576.777, "dur": 13.838,
    "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": 870, "tid": 870,
    "ts": 7429770613577.014, "dur": 0.127,
    "args": {
      "Python parent id": 2140, "Python id": 2141, "finished": true, "Ev Idx": 8971
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770613577.768, "dur": 12.747,
    "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": 870, "tid": 870,
    "ts": 7429770613591.821, "dur": 0.340,
    "args": {
      "Python parent id": 2125, "Python id": 2143, "finished": true, "Ev Idx": 8973
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_82", "pid": 870, "tid": 870,
    "ts": 7429770613593.931, "dur": 39.128,
    "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": 870, "tid": 870,
    "ts": 7429770613594.698, "dur": 38.036,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770613594.900, "dur": 0.208,
    "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": 870, "tid": 870,
    "ts": 7429770613595.978, "dur": 36.652,
    "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": 870, "tid": 870,
    "ts": 7429770613596.222, "dur": 0.144,
    "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": 870, "tid": 870,
    "ts": 7429770613596.574, "dur": 0.134,
    "args": {
      "Python parent id": 2147, "Python id": 2149, "finished": true, "Ev Idx": 8979
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770613596.822, "dur": 35.711,
    "args": {
      "Python parent id": 2147, "Python id": 2150, "finished": true, "Ev Idx": 8980
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_83", "pid": 870, "tid": 870,
    "ts": 7429770613644.992, "dur": 52.384,
    "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": 870, "tid": 870,
    "ts": 7429770613645.506, "dur": 51.628,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770613645.765, "dur": 0.299,
    "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": 870, "tid": 870,
    "ts": 7429770613646.782, "dur": 50.229,
    "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": 870, "tid": 870,
    "ts": 7429770613646.985, "dur": 0.164,
    "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": 870, "tid": 870,
    "ts": 7429770613647.332, "dur": 0.300,
    "args": {
      "Python parent id": 2154, "Python id": 2156, "finished": true, "Ev Idx": 8986
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770613647.779, "dur": 49.177,
    "args": {
      "Python parent id": 2154, "Python id": 2157, "finished": true, "Ev Idx": 8987
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_12", "pid": 870, "tid": 870,
    "ts": 7429770613716.828, "dur": 1042.208,
    "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": 870, "tid": 870,
    "ts": 7429770613718.110, "dur": 1040.121,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770613718.361, "dur": 0.370,
    "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": 870, "tid": 870,
    "ts": 7429770613719.995, "dur": 1037.449,
    "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": 870, "tid": 870,
    "ts": 7429770613720.317, "dur": 0.661,
    "args": {
      "Python parent id": 2161, "Python id": 2162, "finished": true, "Ev Idx": 8992
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_24", "pid": 870, "tid": 870,
    "ts": 7429770613723.126, "dur": 72.180,
    "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": 870, "tid": 870,
    "ts": 7429770613723.658, "dur": 71.328,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770613723.888, "dur": 0.142,
    "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": 870, "tid": 870,
    "ts": 7429770613724.786, "dur": 69.521,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770613725.347, "dur": 1.607,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770613727.245, "dur": 10.328,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770613737.909, "dur": 12.999,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770613761.960, "dur": 11.848,
    "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": 870, "tid": 870,
    "ts": 7429770613782.865, "dur": 0.246,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770613783.425, "dur": 1.092,
    "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": 870, "tid": 870,
    "ts": 7429770613795.862, "dur": 0.304,
    "args": {
      "Python parent id": 2161, "Python id": 2173, "finished": true, "Ev Idx": 9003
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_12", "pid": 870, "tid": 870,
    "ts": 7429770613799.013, "dur": 673.430,
    "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": 870, "tid": 870,
    "ts": 7429770613800.314, "dur": 671.641,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770613800.482, "dur": 0.417,
    "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": 870, "tid": 870,
    "ts": 7429770613802.634, "dur": 668.609,
    "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": 870, "tid": 870,
    "ts": 7429770613805.840, "dur": 0.371,
    "args": {
      "Python parent id": 2177, "Python id": 2178, "finished": true, "Ev Idx": 9008
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_84", "pid": 870, "tid": 870,
    "ts": 7429770613808.440, "dur": 42.271,
    "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": 870, "tid": 870,
    "ts": 7429770613809.105, "dur": 41.311,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770613809.277, "dur": 0.157,
    "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": 870, "tid": 870,
    "ts": 7429770613810.477, "dur": 39.772,
    "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": 870, "tid": 870,
    "ts": 7429770613810.756, "dur": 0.116,
    "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": 870, "tid": 870,
    "ts": 7429770613811.047, "dur": 0.133,
    "args": {
      "Python parent id": 2182, "Python id": 2184, "finished": true, "Ev Idx": 9014
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770613811.287, "dur": 38.904,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770613851.325, "dur": 3.738,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770613855.525, "dur": 3.684,
    "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": 870, "tid": 870,
    "ts": 7429770613860.043, "dur": 0.327,
    "args": {
      "Python parent id": 2177, "Python id": 2188, "finished": true, "Ev Idx": 9018
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_85", "pid": 870, "tid": 870,
    "ts": 7429770613862.831, "dur": 36.277,
    "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": 870, "tid": 870,
    "ts": 7429770613863.611, "dur": 35.278,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770613863.781, "dur": 0.223,
    "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": 870, "tid": 870,
    "ts": 7429770613864.849, "dur": 33.943,
    "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": 870, "tid": 870,
    "ts": 7429770613865.091, "dur": 0.153,
    "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": 870, "tid": 870,
    "ts": 7429770613865.429, "dur": 0.127,
    "args": {
      "Python parent id": 2192, "Python id": 2194, "finished": true, "Ev Idx": 9024
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770613865.663, "dur": 33.082,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770613899.527, "dur": 2.803,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770613902.777, "dur": 2.710,
    "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": 870, "tid": 870,
    "ts": 7429770613906.221, "dur": 0.299,
    "args": {
      "Python parent id": 2177, "Python id": 2198, "finished": true, "Ev Idx": 9028
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_86", "pid": 870, "tid": 870,
    "ts": 7429770613908.678, "dur": 35.287,
    "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": 870, "tid": 870,
    "ts": 7429770613909.295, "dur": 34.392,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770613909.490, "dur": 0.231,
    "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": 870, "tid": 870,
    "ts": 7429770613910.786, "dur": 32.743,
    "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": 870, "tid": 870,
    "ts": 7429770613910.999, "dur": 0.121,
    "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": 870, "tid": 870,
    "ts": 7429770613911.297, "dur": 0.127,
    "args": {
      "Python parent id": 2202, "Python id": 2204, "finished": true, "Ev Idx": 9034
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770613911.527, "dur": 31.954,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770613944.487, "dur": 2.349,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770613947.275, "dur": 2.890,
    "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": 870, "tid": 870,
    "ts": 7429770613951.252, "dur": 174.837,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770613951.532, "dur": 3.811,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770613955.568, "dur": 1.804,
    "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": 870, "tid": 870,
    "ts": 7429770613969.427, "dur": 49.669,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770613996.789, "dur": 21.229,
    "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": 870, "tid": 870,
    "ts": 7429770614048.689, "dur": 53.376,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770614066.727, "dur": 34.011,
    "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": 870, "tid": 870,
    "ts": 7429770614130.810, "dur": 3.893,
    "args": {
      "Python parent id": 2177, "Python id": 2215, "finished": true, "Ev Idx": 9045
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770614131.585, "dur": 0.302,
    "args": {
      "Python parent id": 2215, "Python id": 2216, "finished": true, "Ev Idx": 9046
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770614132.170, "dur": 0.077,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770614133.619, "dur": 0.233,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770614134.303, "dur": 0.084,
    "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": 870, "tid": 870,
    "ts": 7429770614135.979, "dur": 6.988,
    "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": 870, "tid": 870,
    "ts": 7429770614139.534, "dur": 3.287,
    "args": {
      "Python parent id": 2220, "Python id": 2221, "finished": true, "Ev Idx": 9051
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770614139.892, "dur": 1.360,
    "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": 870, "tid": 870,
    "ts": 7429770614140.330, "dur": 0.795,
    "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": 870, "tid": 870,
    "ts": 7429770614141.511, "dur": 0.487,
    "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": 870, "tid": 870,
    "ts": 7429770614142.317, "dur": 0.427,
    "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": 870, "tid": 870,
    "ts": 7429770614145.843, "dur": 264.481,
    "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": 870, "tid": 870,
    "ts": 7429770614146.300, "dur": 47.342,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770614166.103, "dur": 5.554,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770614172.583, "dur": 20.962,
    "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": 870, "tid": 870,
    "ts": 7429770614194.610, "dur": 32.342,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770614208.914, "dur": 3.554,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770614212.977, "dur": 13.911,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770614227.848, "dur": 3.143,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770614231.240, "dur": 50.974,
    "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": 870, "tid": 870,
    "ts": 7429770614316.117, "dur": 17.444,
    "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": 870, "tid": 870,
    "ts": 7429770614316.339, "dur": 0.192,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770614317.200, "dur": 16.200,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770614334.448, "dur": 1.740,
    "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": 870, "tid": 870,
    "ts": 7429770614337.820, "dur": 11.949,
    "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": 870, "tid": 870,
    "ts": 7429770614337.925, "dur": 0.122,
    "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": 870, "tid": 870,
    "ts": 7429770614345.333, "dur": 0.937,
    "args": {
      "Python parent id": 2239, "Python id": 2241, "finished": true, "Ev Idx": 9071
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770614345.798, "dur": 0.401,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770614346.743, "dur": 2.955,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770614350.399, "dur": 40.802,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770614391.874, "dur": 2.964,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770614395.405, "dur": 13.922,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770614412.969, "dur": 3.107,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770614416.834, "dur": 0.277,
    "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": 870, "tid": 870,
    "ts": 7429770614418.146, "dur": 1.032,
    "args": {
      "Python parent id": 2177, "Python id": 2249, "finished": true, "Ev Idx": 9079
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_87", "pid": 870, "tid": 870,
    "ts": 7429770614423.785, "dur": 46.748,
    "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": 870, "tid": 870,
    "ts": 7429770614424.854, "dur": 45.291,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770614425.249, "dur": 1.111,
    "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": 870, "tid": 870,
    "ts": 7429770614427.891, "dur": 42.012,
    "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": 870, "tid": 870,
    "ts": 7429770614428.195, "dur": 0.189,
    "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": 870, "tid": 870,
    "ts": 7429770614428.613, "dur": 0.183,
    "args": {
      "Python parent id": 2253, "Python id": 2255, "finished": true, "Ev Idx": 9085
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770614428.963, "dur": 40.857,
    "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": 870, "tid": 870,
    "ts": 7429770614485.276, "dur": 0.537,
    "args": {
      "Python parent id": 2161, "Python id": 2257, "finished": true, "Ev Idx": 9087
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_25", "pid": 870, "tid": 870,
    "ts": 7429770614488.630, "dur": 82.531,
    "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": 870, "tid": 870,
    "ts": 7429770614489.210, "dur": 81.620,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770614489.434, "dur": 0.445,
    "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": 870, "tid": 870,
    "ts": 7429770614490.888, "dur": 79.190,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770614491.448, "dur": 1.646,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770614493.398, "dur": 12.540,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770614506.262, "dur": 15.439,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770614534.420, "dur": 11.664,
    "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": 870, "tid": 870,
    "ts": 7429770614558.189, "dur": 0.213,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770614558.863, "dur": 1.236,
    "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": 870, "tid": 870,
    "ts": 7429770614571.888, "dur": 0.400,
    "args": {
      "Python parent id": 2161, "Python id": 2268, "finished": true, "Ev Idx": 9098
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_12", "pid": 870, "tid": 870,
    "ts": 7429770614573.969, "dur": 173.034,
    "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": 870, "tid": 870,
    "ts": 7429770614574.446, "dur": 172.341,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770614574.664, "dur": 0.275,
    "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": 870, "tid": 870,
    "ts": 7429770614576.170, "dur": 170.512,
    "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": 870, "tid": 870,
    "ts": 7429770614576.587, "dur": 0.440,
    "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": 870, "tid": 870,
    "ts": 7429770614577.473, "dur": 0.317,
    "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": 870, "tid": 870,
    "ts": 7429770614578.023, "dur": 0.213,
    "args": {
      "Python parent id": 2272, "Python id": 2275, "finished": true, "Ev Idx": 9105
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_88", "pid": 870, "tid": 870,
    "ts": 7429770614580.608, "dur": 40.906,
    "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": 870, "tid": 870,
    "ts": 7429770614581.310, "dur": 39.971,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770614581.451, "dur": 0.148,
    "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": 870, "tid": 870,
    "ts": 7429770614582.477, "dur": 38.716,
    "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": 870, "tid": 870,
    "ts": 7429770614582.736, "dur": 0.344,
    "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": 870, "tid": 870,
    "ts": 7429770614583.267, "dur": 0.125,
    "args": {
      "Python parent id": 2279, "Python id": 2281, "finished": true, "Ev Idx": 9111
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770614583.508, "dur": 37.589,
    "args": {
      "Python parent id": 2279, "Python id": 2282, "finished": true, "Ev Idx": 9112
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_12", "pid": 870, "tid": 870,
    "ts": 7429770614623.766, "dur": 16.406,
    "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": 870, "tid": 870,
    "ts": 7429770614624.220, "dur": 15.712,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770614624.497, "dur": 0.268,
    "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": 870, "tid": 870,
    "ts": 7429770614626.003, "dur": 13.851,
    "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": 870, "tid": 870,
    "ts": 7429770614626.776, "dur": 12.825,
    "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": 870, "tid": 870,
    "ts": 7429770614626.924, "dur": 0.163,
    "args": {
      "Python parent id": 2287, "Python id": 2288, "finished": true, "Ev Idx": 9118
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770614627.814, "dur": 11.691,
    "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": 870, "tid": 870,
    "ts": 7429770614640.788, "dur": 0.334,
    "args": {
      "Python parent id": 2272, "Python id": 2290, "finished": true, "Ev Idx": 9120
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_89", "pid": 870, "tid": 870,
    "ts": 7429770614643.594, "dur": 38.594,
    "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": 870, "tid": 870,
    "ts": 7429770614644.208, "dur": 37.776,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770614644.366, "dur": 0.214,
    "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": 870, "tid": 870,
    "ts": 7429770614645.496, "dur": 36.392,
    "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": 870, "tid": 870,
    "ts": 7429770614645.727, "dur": 0.175,
    "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": 870, "tid": 870,
    "ts": 7429770614646.144, "dur": 0.131,
    "args": {
      "Python parent id": 2294, "Python id": 2296, "finished": true, "Ev Idx": 9126
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770614646.377, "dur": 35.474,
    "args": {
      "Python parent id": 2294, "Python id": 2297, "finished": true, "Ev Idx": 9127
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_90", "pid": 870, "tid": 870,
    "ts": 7429770614694.740, "dur": 51.611,
    "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": 870, "tid": 870,
    "ts": 7429770614695.210, "dur": 50.930,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770614695.385, "dur": 0.445,
    "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": 870, "tid": 870,
    "ts": 7429770614696.409, "dur": 49.595,
    "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": 870, "tid": 870,
    "ts": 7429770614696.621, "dur": 0.193,
    "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": 870, "tid": 870,
    "ts": 7429770614697.011, "dur": 0.176,
    "args": {
      "Python parent id": 2301, "Python id": 2303, "finished": true, "Ev Idx": 9133
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770614697.294, "dur": 48.653,
    "args": {
      "Python parent id": 2301, "Python id": 2304, "finished": true, "Ev Idx": 9134
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_13", "pid": 870, "tid": 870,
    "ts": 7429770614765.004, "dur": 1070.157,
    "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": 870, "tid": 870,
    "ts": 7429770614766.318, "dur": 1067.850,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770614766.508, "dur": 0.426,
    "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": 870, "tid": 870,
    "ts": 7429770614768.417, "dur": 1064.409,
    "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": 870, "tid": 870,
    "ts": 7429770614768.752, "dur": 0.575,
    "args": {
      "Python parent id": 2308, "Python id": 2309, "finished": true, "Ev Idx": 9139
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_26", "pid": 870, "tid": 870,
    "ts": 7429770614771.679, "dur": 73.612,
    "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": 870, "tid": 870,
    "ts": 7429770614772.119, "dur": 72.992,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770614772.251, "dur": 0.165,
    "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": 870, "tid": 870,
    "ts": 7429770614773.123, "dur": 71.263,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770614773.656, "dur": 1.519,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770614775.476, "dur": 10.743,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770614786.554, "dur": 12.707,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770614812.253, "dur": 11.468,
    "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": 870, "tid": 870,
    "ts": 7429770614833.229, "dur": 0.189,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770614834.036, "dur": 1.037,
    "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": 870, "tid": 870,
    "ts": 7429770614845.754, "dur": 0.398,
    "args": {
      "Python parent id": 2308, "Python id": 2320, "finished": true, "Ev Idx": 9150
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_13", "pid": 870, "tid": 870,
    "ts": 7429770614849.046, "dur": 611.245,
    "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": 870, "tid": 870,
    "ts": 7429770614850.468, "dur": 609.351,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770614850.644, "dur": 0.224,
    "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": 870, "tid": 870,
    "ts": 7429770614852.836, "dur": 606.208,
    "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": 870, "tid": 870,
    "ts": 7429770614855.459, "dur": 0.444,
    "args": {
      "Python parent id": 2324, "Python id": 2325, "finished": true, "Ev Idx": 9155
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_91", "pid": 870, "tid": 870,
    "ts": 7429770614858.191, "dur": 42.546,
    "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": 870, "tid": 870,
    "ts": 7429770614858.693, "dur": 41.795,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770614858.837, "dur": 0.165,
    "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": 870, "tid": 870,
    "ts": 7429770614859.736, "dur": 40.586,
    "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": 870, "tid": 870,
    "ts": 7429770614860.055, "dur": 0.129,
    "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": 870, "tid": 870,
    "ts": 7429770614860.456, "dur": 0.122,
    "args": {
      "Python parent id": 2329, "Python id": 2331, "finished": true, "Ev Idx": 9161
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770614860.761, "dur": 39.498,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770614901.411, "dur": 3.231,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770614905.036, "dur": 3.478,
    "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": 870, "tid": 870,
    "ts": 7429770614909.245, "dur": 0.401,
    "args": {
      "Python parent id": 2324, "Python id": 2335, "finished": true, "Ev Idx": 9165
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_92", "pid": 870, "tid": 870,
    "ts": 7429770614912.063, "dur": 35.776,
    "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": 870, "tid": 870,
    "ts": 7429770614912.663, "dur": 34.989,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770614912.845, "dur": 0.275,
    "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": 870, "tid": 870,
    "ts": 7429770614913.792, "dur": 33.769,
    "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": 870, "tid": 870,
    "ts": 7429770614914.047, "dur": 0.178,
    "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": 870, "tid": 870,
    "ts": 7429770614914.438, "dur": 0.118,
    "args": {
      "Python parent id": 2339, "Python id": 2341, "finished": true, "Ev Idx": 9171
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770614914.661, "dur": 32.857,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770614948.256, "dur": 2.291,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770614951.004, "dur": 2.996,
    "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": 870, "tid": 870,
    "ts": 7429770614954.780, "dur": 0.351,
    "args": {
      "Python parent id": 2324, "Python id": 2345, "finished": true, "Ev Idx": 9175
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_93", "pid": 870, "tid": 870,
    "ts": 7429770614957.534, "dur": 33.216,
    "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": 870, "tid": 870,
    "ts": 7429770614958.124, "dur": 32.414,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770614958.330, "dur": 0.227,
    "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": 870, "tid": 870,
    "ts": 7429770614959.389, "dur": 31.052,
    "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": 870, "tid": 870,
    "ts": 7429770614959.620, "dur": 0.132,
    "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": 870, "tid": 870,
    "ts": 7429770614960.000, "dur": 0.111,
    "args": {
      "Python parent id": 2349, "Python id": 2351, "finished": true, "Ev Idx": 9181
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770614960.217, "dur": 30.182,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770614991.297, "dur": 2.610,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770614994.425, "dur": 2.905,
    "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": 870, "tid": 870,
    "ts": 7429770614998.333, "dur": 142.868,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770614998.548, "dur": 2.776,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770615001.572, "dur": 1.804,
    "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": 870, "tid": 870,
    "ts": 7429770615014.245, "dur": 45.340,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770615040.086, "dur": 18.371,
    "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": 870, "tid": 870,
    "ts": 7429770615092.483, "dur": 31.420,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770615110.073, "dur": 12.978,
    "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": 870, "tid": 870,
    "ts": 7429770615145.003, "dur": 3.297,
    "args": {
      "Python parent id": 2324, "Python id": 2362, "finished": true, "Ev Idx": 9192
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770615145.709, "dur": 0.215,
    "args": {
      "Python parent id": 2362, "Python id": 2363, "finished": true, "Ev Idx": 9193
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770615146.221, "dur": 0.067,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770615147.464, "dur": 0.198,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770615148.022, "dur": 0.062,
    "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": 870, "tid": 870,
    "ts": 7429770615149.478, "dur": 5.913,
    "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": 870, "tid": 870,
    "ts": 7429770615152.132, "dur": 3.135,
    "args": {
      "Python parent id": 2367, "Python id": 2368, "finished": true, "Ev Idx": 9198
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770615152.480, "dur": 1.236,
    "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": 870, "tid": 870,
    "ts": 7429770615152.818, "dur": 0.743,
    "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": 870, "tid": 870,
    "ts": 7429770615154.035, "dur": 0.479,
    "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": 870, "tid": 870,
    "ts": 7429770615154.781, "dur": 0.418,
    "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": 870, "tid": 870,
    "ts": 7429770615158.008, "dur": 244.596,
    "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": 870, "tid": 870,
    "ts": 7429770615158.504, "dur": 43.838,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770615177.374, "dur": 5.059,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770615183.264, "dur": 18.953,
    "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": 870, "tid": 870,
    "ts": 7429770615203.273, "dur": 29.484,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770615216.549, "dur": 3.146,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770615220.172, "dur": 12.496,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770615233.713, "dur": 2.835,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770615236.826, "dur": 45.831,
    "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": 870, "tid": 870,
    "ts": 7429770615313.935, "dur": 15.727,
    "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": 870, "tid": 870,
    "ts": 7429770615314.135, "dur": 0.209,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770615315.017, "dur": 14.501,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770615330.462, "dur": 1.557,
    "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": 870, "tid": 870,
    "ts": 7429770615333.552, "dur": 8.682,
    "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": 870, "tid": 870,
    "ts": 7429770615333.668, "dur": 0.117,
    "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": 870, "tid": 870,
    "ts": 7429770615338.480, "dur": 0.912,
    "args": {
      "Python parent id": 2386, "Python id": 2388, "finished": true, "Ev Idx": 9218
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770615339.021, "dur": 0.293,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770615339.839, "dur": 2.326,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770615342.920, "dur": 41.096,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770615384.733, "dur": 3.087,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770615388.266, "dur": 13.433,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770615405.424, "dur": 3.078,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770615409.212, "dur": 0.250,
    "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": 870, "tid": 870,
    "ts": 7429770615410.446, "dur": 0.889,
    "args": {
      "Python parent id": 2324, "Python id": 2396, "finished": true, "Ev Idx": 9226
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_94", "pid": 870, "tid": 870,
    "ts": 7429770615416.308, "dur": 42.125,
    "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": 870, "tid": 870,
    "ts": 7429770615417.192, "dur": 40.917,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770615417.644, "dur": 0.809,
    "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": 870, "tid": 870,
    "ts": 7429770615419.588, "dur": 38.275,
    "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": 870, "tid": 870,
    "ts": 7429770615419.999, "dur": 0.280,
    "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": 870, "tid": 870,
    "ts": 7429770615420.530, "dur": 0.182,
    "args": {
      "Python parent id": 2400, "Python id": 2402, "finished": true, "Ev Idx": 9232
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770615420.893, "dur": 36.899,
    "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": 870, "tid": 870,
    "ts": 7429770615472.053, "dur": 0.565,
    "args": {
      "Python parent id": 2308, "Python id": 2404, "finished": true, "Ev Idx": 9234
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_27", "pid": 870, "tid": 870,
    "ts": 7429770615475.745, "dur": 80.597,
    "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": 870, "tid": 870,
    "ts": 7429770615476.420, "dur": 79.676,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770615476.635, "dur": 0.334,
    "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": 870, "tid": 870,
    "ts": 7429770615478.644, "dur": 76.574,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770615479.160, "dur": 1.589,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770615481.055, "dur": 13.193,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770615494.587, "dur": 13.896,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770615520.016, "dur": 9.531,
    "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": 870, "tid": 870,
    "ts": 7429770615541.751, "dur": 0.237,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770615542.341, "dur": 1.137,
    "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": 870, "tid": 870,
    "ts": 7429770615557.045, "dur": 0.426,
    "args": {
      "Python parent id": 2308, "Python id": 2415, "finished": true, "Ev Idx": 9245
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_13", "pid": 870, "tid": 870,
    "ts": 7429770615559.322, "dur": 261.089,
    "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": 870, "tid": 870,
    "ts": 7429770615559.780, "dur": 260.293,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770615560.027, "dur": 0.310,
    "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": 870, "tid": 870,
    "ts": 7429770615561.490, "dur": 258.491,
    "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": 870, "tid": 870,
    "ts": 7429770615561.922, "dur": 0.304,
    "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": 870, "tid": 870,
    "ts": 7429770615562.575, "dur": 0.307,
    "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": 870, "tid": 870,
    "ts": 7429770615563.097, "dur": 0.212,
    "args": {
      "Python parent id": 2419, "Python id": 2422, "finished": true, "Ev Idx": 9252
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_95", "pid": 870, "tid": 870,
    "ts": 7429770615565.614, "dur": 43.156,
    "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": 870, "tid": 870,
    "ts": 7429770615566.160, "dur": 42.360,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770615566.307, "dur": 0.162,
    "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": 870, "tid": 870,
    "ts": 7429770615567.260, "dur": 41.117,
    "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": 870, "tid": 870,
    "ts": 7429770615567.538, "dur": 0.163,
    "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": 870, "tid": 870,
    "ts": 7429770615567.899, "dur": 0.335,
    "args": {
      "Python parent id": 2426, "Python id": 2428, "finished": true, "Ev Idx": 9258
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770615568.379, "dur": 39.901,
    "args": {
      "Python parent id": 2426, "Python id": 2429, "finished": true, "Ev Idx": 9259
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_13", "pid": 870, "tid": 870,
    "ts": 7429770615610.982, "dur": 15.637,
    "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": 870, "tid": 870,
    "ts": 7429770615611.382, "dur": 14.863,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770615611.597, "dur": 0.330,
    "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": 870, "tid": 870,
    "ts": 7429770615612.935, "dur": 13.210,
    "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": 870, "tid": 870,
    "ts": 7429770615613.702, "dur": 12.246,
    "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": 870, "tid": 870,
    "ts": 7429770615613.877, "dur": 0.124,
    "args": {
      "Python parent id": 2434, "Python id": 2435, "finished": true, "Ev Idx": 9265
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770615614.683, "dur": 11.172,
    "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": 870, "tid": 870,
    "ts": 7429770615627.267, "dur": 0.337,
    "args": {
      "Python parent id": 2419, "Python id": 2437, "finished": true, "Ev Idx": 9267
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_96", "pid": 870, "tid": 870,
    "ts": 7429770615630.378, "dur": 119.552,
    "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": 870, "tid": 870,
    "ts": 7429770615631.119, "dur": 118.526,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770615631.291, "dur": 0.307,
    "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": 870, "tid": 870,
    "ts": 7429770615632.437, "dur": 117.016,
    "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": 870, "tid": 870,
    "ts": 7429770615632.638, "dur": 0.118,
    "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": 870, "tid": 870,
    "ts": 7429770615632.969, "dur": 0.110,
    "args": {
      "Python parent id": 2441, "Python id": 2443, "finished": true, "Ev Idx": 9273
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770615633.185, "dur": 116.159,
    "args": {
      "Python parent id": 2441, "Python id": 2444, "finished": true, "Ev Idx": 9274
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_97", "pid": 870, "tid": 870,
    "ts": 7429770615764.682, "dur": 54.992,
    "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": 870, "tid": 870,
    "ts": 7429770615765.483, "dur": 53.951,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770615765.905, "dur": 0.449,
    "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": 870, "tid": 870,
    "ts": 7429770615767.438, "dur": 51.908,
    "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": 870, "tid": 870,
    "ts": 7429770615767.849, "dur": 0.154,
    "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": 870, "tid": 870,
    "ts": 7429770615768.366, "dur": 0.182,
    "args": {
      "Python parent id": 2448, "Python id": 2450, "finished": true, "Ev Idx": 9280
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770615768.662, "dur": 50.617,
    "args": {
      "Python parent id": 2448, "Python id": 2451, "finished": true, "Ev Idx": 9281
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_14", "pid": 870, "tid": 870,
    "ts": 7429770615840.900, "dur": 1117.539,
    "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": 870, "tid": 870,
    "ts": 7429770615842.419, "dur": 1115.272,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770615842.702, "dur": 0.298,
    "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": 870, "tid": 870,
    "ts": 7429770615844.338, "dur": 1112.681,
    "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": 870, "tid": 870,
    "ts": 7429770615844.688, "dur": 0.691,
    "args": {
      "Python parent id": 2455, "Python id": 2456, "finished": true, "Ev Idx": 9286
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_28", "pid": 870, "tid": 870,
    "ts": 7429770615847.793, "dur": 78.143,
    "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": 870, "tid": 870,
    "ts": 7429770615848.240, "dur": 77.418,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770615848.373, "dur": 0.157,
    "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": 870, "tid": 870,
    "ts": 7429770615849.800, "dur": 75.134,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770615850.354, "dur": 1.856,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770615852.485, "dur": 11.520,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770615864.374, "dur": 13.475,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770615890.394, "dur": 12.279,
    "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": 870, "tid": 870,
    "ts": 7429770615912.897, "dur": 0.241,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770615913.674, "dur": 1.310,
    "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": 870, "tid": 870,
    "ts": 7429770615926.429, "dur": 0.443,
    "args": {
      "Python parent id": 2455, "Python id": 2467, "finished": true, "Ev Idx": 9297
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_14", "pid": 870, "tid": 870,
    "ts": 7429770615929.949, "dur": 732.109,
    "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": 870, "tid": 870,
    "ts": 7429770615931.309, "dur": 729.611,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770615931.592, "dur": 0.251,
    "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": 870, "tid": 870,
    "ts": 7429770615933.707, "dur": 726.198,
    "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": 870, "tid": 870,
    "ts": 7429770615936.919, "dur": 0.388,
    "args": {
      "Python parent id": 2471, "Python id": 2472, "finished": true, "Ev Idx": 9302
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_98", "pid": 870, "tid": 870,
    "ts": 7429770615939.585, "dur": 42.941,
    "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": 870, "tid": 870,
    "ts": 7429770615940.218, "dur": 42.094,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770615940.362, "dur": 0.161,
    "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": 870, "tid": 870,
    "ts": 7429770615941.453, "dur": 40.739,
    "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": 870, "tid": 870,
    "ts": 7429770615941.807, "dur": 0.118,
    "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": 870, "tid": 870,
    "ts": 7429770615942.122, "dur": 0.163,
    "args": {
      "Python parent id": 2476, "Python id": 2478, "finished": true, "Ev Idx": 9308
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770615942.388, "dur": 39.758,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770615983.192, "dur": 3.210,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770615986.783, "dur": 3.977,
    "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": 870, "tid": 870,
    "ts": 7429770615991.759, "dur": 0.451,
    "args": {
      "Python parent id": 2471, "Python id": 2482, "finished": true, "Ev Idx": 9312
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_99", "pid": 870, "tid": 870,
    "ts": 7429770615994.243, "dur": 65.595,
    "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": 870, "tid": 870,
    "ts": 7429770615994.931, "dur": 64.637,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770615995.185, "dur": 0.228,
    "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": 870, "tid": 870,
    "ts": 7429770615996.147, "dur": 63.324,
    "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": 870, "tid": 870,
    "ts": 7429770615996.387, "dur": 0.121,
    "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": 870, "tid": 870,
    "ts": 7429770615996.715, "dur": 0.126,
    "args": {
      "Python parent id": 2486, "Python id": 2488, "finished": true, "Ev Idx": 9318
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770615996.947, "dur": 62.477,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770616060.330, "dur": 2.367,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770616063.030, "dur": 2.619,
    "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": 870, "tid": 870,
    "ts": 7429770616066.381, "dur": 0.418,
    "args": {
      "Python parent id": 2471, "Python id": 2492, "finished": true, "Ev Idx": 9322
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_100", "pid": 870, "tid": 870,
    "ts": 7429770616069.005, "dur": 34.116,
    "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": 870, "tid": 870,
    "ts": 7429770616069.573, "dur": 33.350,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770616069.777, "dur": 0.183,
    "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": 870, "tid": 870,
    "ts": 7429770616070.926, "dur": 31.898,
    "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": 870, "tid": 870,
    "ts": 7429770616071.165, "dur": 0.125,
    "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": 870, "tid": 870,
    "ts": 7429770616071.486, "dur": 0.135,
    "args": {
      "Python parent id": 2496, "Python id": 2498, "finished": true, "Ev Idx": 9328
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770616071.725, "dur": 31.043,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770616103.659, "dur": 2.405,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770616106.786, "dur": 2.426,
    "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": 870, "tid": 870,
    "ts": 7429770616110.264, "dur": 147.986,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770616110.475, "dur": 3.323,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770616114.059, "dur": 1.620,
    "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": 870, "tid": 870,
    "ts": 7429770616126.511, "dur": 48.287,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770616153.364, "dur": 20.360,
    "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": 870, "tid": 870,
    "ts": 7429770616204.527, "dur": 37.236,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770616222.571, "dur": 18.344,
    "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": 870, "tid": 870,
    "ts": 7429770616262.233, "dur": 3.556,
    "args": {
      "Python parent id": 2471, "Python id": 2509, "finished": true, "Ev Idx": 9339
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770616262.896, "dur": 0.234,
    "args": {
      "Python parent id": 2509, "Python id": 2510, "finished": true, "Ev Idx": 9340
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770616263.448, "dur": 0.085,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770616264.765, "dur": 0.273,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770616265.469, "dur": 0.084,
    "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": 870, "tid": 870,
    "ts": 7429770616267.026, "dur": 6.146,
    "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": 870, "tid": 870,
    "ts": 7429770616269.972, "dur": 3.085,
    "args": {
      "Python parent id": 2514, "Python id": 2515, "finished": true, "Ev Idx": 9345
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770616270.302, "dur": 1.068,
    "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": 870, "tid": 870,
    "ts": 7429770616270.632, "dur": 0.610,
    "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": 870, "tid": 870,
    "ts": 7429770616271.688, "dur": 0.444,
    "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": 870, "tid": 870,
    "ts": 7429770616272.574, "dur": 0.407,
    "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": 870, "tid": 870,
    "ts": 7429770616276.083, "dur": 320.992,
    "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": 870, "tid": 870,
    "ts": 7429770616276.615, "dur": 43.292,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770616292.879, "dur": 5.059,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770616299.007, "dur": 20.735,
    "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": 870, "tid": 870,
    "ts": 7429770616320.791, "dur": 28.570,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770616333.325, "dur": 2.650,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770616336.571, "dur": 12.726,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770616350.283, "dur": 2.742,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770616353.321, "dur": 110.265,
    "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": 870, "tid": 870,
    "ts": 7429770616500.348, "dur": 17.982,
    "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": 870, "tid": 870,
    "ts": 7429770616500.602, "dur": 0.229,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770616501.529, "dur": 16.602,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770616519.121, "dur": 1.863,
    "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": 870, "tid": 870,
    "ts": 7429770616522.648, "dur": 10.225,
    "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": 870, "tid": 870,
    "ts": 7429770616522.803, "dur": 0.119,
    "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": 870, "tid": 870,
    "ts": 7429770616528.415, "dur": 1.119,
    "args": {
      "Python parent id": 2533, "Python id": 2535, "finished": true, "Ev Idx": 9365
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770616528.992, "dur": 0.467,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770616530.032, "dur": 2.773,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770616533.679, "dur": 43.130,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770616577.492, "dur": 3.177,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770616581.256, "dur": 14.912,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770616600.214, "dur": 3.632,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770616604.516, "dur": 0.262,
    "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": 870, "tid": 870,
    "ts": 7429770616605.926, "dur": 1.031,
    "args": {
      "Python parent id": 2471, "Python id": 2543, "finished": true, "Ev Idx": 9373
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_101", "pid": 870, "tid": 870,
    "ts": 7429770616611.884, "dur": 47.384,
    "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": 870, "tid": 870,
    "ts": 7429770616613.097, "dur": 45.803,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770616613.602, "dur": 1.365,
    "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": 870, "tid": 870,
    "ts": 7429770616616.447, "dur": 42.193,
    "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": 870, "tid": 870,
    "ts": 7429770616616.830, "dur": 0.213,
    "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": 870, "tid": 870,
    "ts": 7429770616617.302, "dur": 0.142,
    "args": {
      "Python parent id": 2547, "Python id": 2549, "finished": true, "Ev Idx": 9379
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770616617.581, "dur": 40.935,
    "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": 870, "tid": 870,
    "ts": 7429770616675.008, "dur": 0.739,
    "args": {
      "Python parent id": 2455, "Python id": 2551, "finished": true, "Ev Idx": 9381
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_29", "pid": 870, "tid": 870,
    "ts": 7429770616679.470, "dur": 86.933,
    "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": 870, "tid": 870,
    "ts": 7429770616680.170, "dur": 85.916,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770616680.439, "dur": 0.372,
    "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": 870, "tid": 870,
    "ts": 7429770616681.804, "dur": 83.469,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770616682.437, "dur": 1.787,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770616684.528, "dur": 14.078,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770616698.957, "dur": 16.598,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770616728.020, "dur": 11.676,
    "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": 870, "tid": 870,
    "ts": 7429770616751.872, "dur": 0.234,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770616752.480, "dur": 1.291,
    "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": 870, "tid": 870,
    "ts": 7429770616767.113, "dur": 0.353,
    "args": {
      "Python parent id": 2455, "Python id": 2562, "finished": true, "Ev Idx": 9392
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_14", "pid": 870, "tid": 870,
    "ts": 7429770616769.476, "dur": 177.620,
    "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": 870, "tid": 870,
    "ts": 7429770616770.211, "dur": 176.702,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770616770.491, "dur": 0.270,
    "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": 870, "tid": 870,
    "ts": 7429770616771.863, "dur": 174.926,
    "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": 870, "tid": 870,
    "ts": 7429770616772.303, "dur": 0.622,
    "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": 870, "tid": 870,
    "ts": 7429770616773.336, "dur": 0.327,
    "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": 870, "tid": 870,
    "ts": 7429770616773.928, "dur": 0.213,
    "args": {
      "Python parent id": 2566, "Python id": 2569, "finished": true, "Ev Idx": 9399
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_102", "pid": 870, "tid": 870,
    "ts": 7429770616776.624, "dur": 43.391,
    "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": 870, "tid": 870,
    "ts": 7429770616777.025, "dur": 42.773,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770616777.171, "dur": 1.980,
    "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": 870, "tid": 870,
    "ts": 7429770616779.932, "dur": 39.758,
    "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": 870, "tid": 870,
    "ts": 7429770616780.247, "dur": 0.133,
    "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": 870, "tid": 870,
    "ts": 7429770616780.620, "dur": 0.126,
    "args": {
      "Python parent id": 2573, "Python id": 2575, "finished": true, "Ev Idx": 9405
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770616780.866, "dur": 38.751,
    "args": {
      "Python parent id": 2573, "Python id": 2576, "finished": true, "Ev Idx": 9406
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_14", "pid": 870, "tid": 870,
    "ts": 7429770616821.970, "dur": 19.020,
    "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": 870, "tid": 870,
    "ts": 7429770616822.632, "dur": 18.155,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770616822.836, "dur": 0.273,
    "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": 870, "tid": 870,
    "ts": 7429770616824.278, "dur": 16.413,
    "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": 870, "tid": 870,
    "ts": 7429770616825.314, "dur": 15.143,
    "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": 870, "tid": 870,
    "ts": 7429770616825.445, "dur": 0.197,
    "args": {
      "Python parent id": 2581, "Python id": 2582, "finished": true, "Ev Idx": 9412
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770616826.400, "dur": 13.966,
    "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": 870, "tid": 870,
    "ts": 7429770616841.611, "dur": 0.409,
    "args": {
      "Python parent id": 2566, "Python id": 2584, "finished": true, "Ev Idx": 9414
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_103", "pid": 870, "tid": 870,
    "ts": 7429770616844.979, "dur": 37.294,
    "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": 870, "tid": 870,
    "ts": 7429770616845.461, "dur": 36.582,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770616845.620, "dur": 0.195,
    "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": 870, "tid": 870,
    "ts": 7429770616846.510, "dur": 35.120,
    "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": 870, "tid": 870,
    "ts": 7429770616846.757, "dur": 0.137,
    "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": 870, "tid": 870,
    "ts": 7429770616847.105, "dur": 0.132,
    "args": {
      "Python parent id": 2588, "Python id": 2590, "finished": true, "Ev Idx": 9420
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770616847.354, "dur": 34.229,
    "args": {
      "Python parent id": 2588, "Python id": 2591, "finished": true, "Ev Idx": 9421
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_104", "pid": 870, "tid": 870,
    "ts": 7429770616894.707, "dur": 51.698,
    "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": 870, "tid": 870,
    "ts": 7429770616895.273, "dur": 50.866,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770616895.455, "dur": 0.237,
    "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": 870, "tid": 870,
    "ts": 7429770616896.650, "dur": 49.369,
    "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": 870, "tid": 870,
    "ts": 7429770616896.912, "dur": 0.151,
    "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": 870, "tid": 870,
    "ts": 7429770616897.250, "dur": 0.136,
    "args": {
      "Python parent id": 2595, "Python id": 2597, "finished": true, "Ev Idx": 9427
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770616897.497, "dur": 48.452,
    "args": {
      "Python parent id": 2595, "Python id": 2598, "finished": true, "Ev Idx": 9428
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_15", "pid": 870, "tid": 870,
    "ts": 7429770616964.216, "dur": 1099.603,
    "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": 870, "tid": 870,
    "ts": 7429770616965.419, "dur": 1097.328,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770616965.588, "dur": 0.257,
    "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": 870, "tid": 870,
    "ts": 7429770616967.065, "dur": 1095.024,
    "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": 870, "tid": 870,
    "ts": 7429770616967.429, "dur": 0.664,
    "args": {
      "Python parent id": 2602, "Python id": 2603, "finished": true, "Ev Idx": 9433
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_30", "pid": 870, "tid": 870,
    "ts": 7429770616970.468, "dur": 74.702,
    "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": 870, "tid": 870,
    "ts": 7429770616971.001, "dur": 73.881,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770616971.120, "dur": 0.156,
    "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": 870, "tid": 870,
    "ts": 7429770616971.951, "dur": 72.202,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770616972.388, "dur": 1.583,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770616974.248, "dur": 11.240,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770616985.859, "dur": 14.198,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770617011.851, "dur": 10.759,
    "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": 870, "tid": 870,
    "ts": 7429770617032.551, "dur": 0.212,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770617033.172, "dur": 1.249,
    "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": 870, "tid": 870,
    "ts": 7429770617045.752, "dur": 0.330,
    "args": {
      "Python parent id": 2602, "Python id": 2614, "finished": true, "Ev Idx": 9444
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_15", "pid": 870, "tid": 870,
    "ts": 7429770617049.262, "dur": 720.058,
    "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": 870, "tid": 870,
    "ts": 7429770617050.433, "dur": 718.025,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770617050.621, "dur": 0.262,
    "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": 870, "tid": 870,
    "ts": 7429770617052.610, "dur": 714.942,
    "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": 870, "tid": 870,
    "ts": 7429770617055.684, "dur": 0.382,
    "args": {
      "Python parent id": 2618, "Python id": 2619, "finished": true, "Ev Idx": 9449
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_105", "pid": 870, "tid": 870,
    "ts": 7429770617058.161, "dur": 43.506,
    "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": 870, "tid": 870,
    "ts": 7429770617058.845, "dur": 42.435,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770617059.007, "dur": 0.206,
    "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": 870, "tid": 870,
    "ts": 7429770617059.969, "dur": 41.197,
    "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": 870, "tid": 870,
    "ts": 7429770617060.252, "dur": 0.118,
    "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": 870, "tid": 870,
    "ts": 7429770617060.607, "dur": 0.118,
    "args": {
      "Python parent id": 2623, "Python id": 2625, "finished": true, "Ev Idx": 9455
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770617060.849, "dur": 40.232,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770617102.620, "dur": 3.794,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770617106.866, "dur": 3.358,
    "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": 870, "tid": 870,
    "ts": 7429770617111.163, "dur": 0.364,
    "args": {
      "Python parent id": 2618, "Python id": 2629, "finished": true, "Ev Idx": 9459
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_106", "pid": 870, "tid": 870,
    "ts": 7429770617113.998, "dur": 35.768,
    "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": 870, "tid": 870,
    "ts": 7429770617114.430, "dur": 35.137,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770617114.588, "dur": 0.284,
    "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": 870, "tid": 870,
    "ts": 7429770617115.435, "dur": 34.017,
    "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": 870, "tid": 870,
    "ts": 7429770617115.659, "dur": 0.120,
    "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": 870, "tid": 870,
    "ts": 7429770617115.969, "dur": 0.136,
    "args": {
      "Python parent id": 2633, "Python id": 2635, "finished": true, "Ev Idx": 9465
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770617116.205, "dur": 33.201,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770617150.261, "dur": 2.444,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770617153.462, "dur": 2.877,
    "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": 870, "tid": 870,
    "ts": 7429770617157.196, "dur": 0.409,
    "args": {
      "Python parent id": 2618, "Python id": 2639, "finished": true, "Ev Idx": 9469
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_107", "pid": 870, "tid": 870,
    "ts": 7429770617160.325, "dur": 34.886,
    "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": 870, "tid": 870,
    "ts": 7429770617160.725, "dur": 34.275,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770617160.885, "dur": 0.236,
    "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": 870, "tid": 870,
    "ts": 7429770617161.991, "dur": 32.910,
    "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": 870, "tid": 870,
    "ts": 7429770617162.235, "dur": 0.185,
    "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": 870, "tid": 870,
    "ts": 7429770617162.601, "dur": 0.123,
    "args": {
      "Python parent id": 2643, "Python id": 2645, "finished": true, "Ev Idx": 9475
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770617162.831, "dur": 31.991,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770617195.868, "dur": 2.435,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770617198.854, "dur": 2.435,
    "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": 870, "tid": 870,
    "ts": 7429770617202.281, "dur": 149.319,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770617202.538, "dur": 3.316,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770617206.115, "dur": 1.948,
    "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": 870, "tid": 870,
    "ts": 7429770617220.625, "dur": 49.162,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770617247.619, "dur": 20.875,
    "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": 870, "tid": 870,
    "ts": 7429770617301.620, "dur": 31.082,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770617318.566, "dur": 13.276,
    "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": 870, "tid": 870,
    "ts": 7429770617355.954, "dur": 3.962,
    "args": {
      "Python parent id": 2618, "Python id": 2656, "finished": true, "Ev Idx": 9486
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770617356.869, "dur": 0.218,
    "args": {
      "Python parent id": 2656, "Python id": 2657, "finished": true, "Ev Idx": 9487
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770617357.467, "dur": 0.062,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770617358.916, "dur": 0.246,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770617359.590, "dur": 0.078,
    "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": 870, "tid": 870,
    "ts": 7429770617361.208, "dur": 6.084,
    "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": 870, "tid": 870,
    "ts": 7429770617364.012, "dur": 3.132,
    "args": {
      "Python parent id": 2661, "Python id": 2662, "finished": true, "Ev Idx": 9492
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770617364.352, "dur": 1.163,
    "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": 870, "tid": 870,
    "ts": 7429770617364.719, "dur": 0.668,
    "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": 870, "tid": 870,
    "ts": 7429770617365.889, "dur": 0.436,
    "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": 870, "tid": 870,
    "ts": 7429770617366.668, "dur": 0.394,
    "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": 870, "tid": 870,
    "ts": 7429770617370.086, "dur": 333.686,
    "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": 870, "tid": 870,
    "ts": 7429770617370.667, "dur": 45.956,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770617389.713, "dur": 5.112,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770617395.785, "dur": 20.696,
    "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": 870, "tid": 870,
    "ts": 7429770617417.465, "dur": 30.441,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770617430.872, "dur": 3.374,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770617434.849, "dur": 12.992,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770617448.860, "dur": 2.921,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770617452.110, "dur": 114.441,
    "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": 870, "tid": 870,
    "ts": 7429770617604.842, "dur": 17.785,
    "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": 870, "tid": 870,
    "ts": 7429770617605.109, "dur": 0.250,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770617606.131, "dur": 16.300,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770617623.547, "dur": 1.867,
    "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": 870, "tid": 870,
    "ts": 7429770617626.826, "dur": 10.428,
    "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": 870, "tid": 870,
    "ts": 7429770617627.003, "dur": 0.076,
    "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": 870, "tid": 870,
    "ts": 7429770617632.841, "dur": 1.101,
    "args": {
      "Python parent id": 2680, "Python id": 2682, "finished": true, "Ev Idx": 9512
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770617633.464, "dur": 0.387,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770617634.423, "dur": 2.747,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770617637.959, "dur": 45.437,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770617684.110, "dur": 3.232,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770617687.853, "dur": 14.865,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770617707.148, "dur": 3.526,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770617711.462, "dur": 0.284,
    "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": 870, "tid": 870,
    "ts": 7429770617712.896, "dur": 1.246,
    "args": {
      "Python parent id": 2618, "Python id": 2690, "finished": true, "Ev Idx": 9520
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_108", "pid": 870, "tid": 870,
    "ts": 7429770617720.416, "dur": 46.488,
    "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": 870, "tid": 870,
    "ts": 7429770617721.469, "dur": 44.973,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770617722.030, "dur": 1.066,
    "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": 870, "tid": 870,
    "ts": 7429770617724.304, "dur": 41.836,
    "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": 870, "tid": 870,
    "ts": 7429770617724.675, "dur": 0.192,
    "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": 870, "tid": 870,
    "ts": 7429770617725.132, "dur": 0.157,
    "args": {
      "Python parent id": 2694, "Python id": 2696, "finished": true, "Ev Idx": 9526
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770617725.449, "dur": 40.600,
    "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": 870, "tid": 870,
    "ts": 7429770617782.979, "dur": 0.678,
    "args": {
      "Python parent id": 2602, "Python id": 2698, "finished": true, "Ev Idx": 9528
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_31", "pid": 870, "tid": 870,
    "ts": 7429770617786.216, "dur": 85.962,
    "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": 870, "tid": 870,
    "ts": 7429770617786.839, "dur": 85.017,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770617787.108, "dur": 0.347,
    "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": 870, "tid": 870,
    "ts": 7429770617788.251, "dur": 82.809,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770617788.992, "dur": 1.518,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770617790.835, "dur": 13.392,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770617804.570, "dur": 15.857,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770617834.020, "dur": 12.643,
    "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": 870, "tid": 870,
    "ts": 7429770617858.158, "dur": 0.319,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770617858.891, "dur": 1.369,
    "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": 870, "tid": 870,
    "ts": 7429770617872.912, "dur": 0.545,
    "args": {
      "Python parent id": 2602, "Python id": 2709, "finished": true, "Ev Idx": 9539
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_15", "pid": 870, "tid": 870,
    "ts": 7429770617875.441, "dur": 176.309,
    "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": 870, "tid": 870,
    "ts": 7429770617876.182, "dur": 175.401,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770617876.421, "dur": 0.284,
    "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": 870, "tid": 870,
    "ts": 7429770617877.849, "dur": 173.641,
    "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": 870, "tid": 870,
    "ts": 7429770617878.354, "dur": 0.632,
    "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": 870, "tid": 870,
    "ts": 7429770617879.253, "dur": 0.297,
    "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": 870, "tid": 870,
    "ts": 7429770617879.799, "dur": 0.226,
    "args": {
      "Python parent id": 2713, "Python id": 2716, "finished": true, "Ev Idx": 9546
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_109", "pid": 870, "tid": 870,
    "ts": 7429770617883.667, "dur": 44.952,
    "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": 870, "tid": 870,
    "ts": 7429770617884.361, "dur": 43.968,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770617884.518, "dur": 0.151,
    "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": 870, "tid": 870,
    "ts": 7429770617885.604, "dur": 42.607,
    "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": 870, "tid": 870,
    "ts": 7429770617885.894, "dur": 0.122,
    "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": 870, "tid": 870,
    "ts": 7429770617886.241, "dur": 0.124,
    "args": {
      "Python parent id": 2720, "Python id": 2722, "finished": true, "Ev Idx": 9552
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770617886.831, "dur": 41.309,
    "args": {
      "Python parent id": 2720, "Python id": 2723, "finished": true, "Ev Idx": 9553
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_15", "pid": 870, "tid": 870,
    "ts": 7429770617930.976, "dur": 17.147,
    "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": 870, "tid": 870,
    "ts": 7429770617931.457, "dur": 16.409,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770617931.676, "dur": 0.211,
    "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": 870, "tid": 870,
    "ts": 7429770617933.008, "dur": 14.739,
    "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": 870, "tid": 870,
    "ts": 7429770617934.077, "dur": 13.509,
    "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": 870, "tid": 870,
    "ts": 7429770617934.249, "dur": 0.167,
    "args": {
      "Python parent id": 2728, "Python id": 2729, "finished": true, "Ev Idx": 9559
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770617935.222, "dur": 12.271,
    "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": 870, "tid": 870,
    "ts": 7429770617948.719, "dur": 0.380,
    "args": {
      "Python parent id": 2713, "Python id": 2731, "finished": true, "Ev Idx": 9561
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_110", "pid": 870, "tid": 870,
    "ts": 7429770617951.280, "dur": 39.773,
    "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": 870, "tid": 870,
    "ts": 7429770617951.729, "dur": 39.042,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770617951.910, "dur": 0.207,
    "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": 870, "tid": 870,
    "ts": 7429770617953.164, "dur": 37.478,
    "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": 870, "tid": 870,
    "ts": 7429770617953.388, "dur": 0.142,
    "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": 870, "tid": 870,
    "ts": 7429770617953.704, "dur": 0.150,
    "args": {
      "Python parent id": 2735, "Python id": 2737, "finished": true, "Ev Idx": 9567
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770617953.962, "dur": 36.633,
    "args": {
      "Python parent id": 2735, "Python id": 2738, "finished": true, "Ev Idx": 9568
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_111", "pid": 870, "tid": 870,
    "ts": 7429770618003.352, "dur": 47.836,
    "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": 870, "tid": 870,
    "ts": 7429770618003.879, "dur": 47.086,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770618004.080, "dur": 0.204,
    "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": 870, "tid": 870,
    "ts": 7429770618004.902, "dur": 45.913,
    "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": 870, "tid": 870,
    "ts": 7429770618005.162, "dur": 0.164,
    "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": 870, "tid": 870,
    "ts": 7429770618005.523, "dur": 0.119,
    "args": {
      "Python parent id": 2742, "Python id": 2744, "finished": true, "Ev Idx": 9574
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770618005.746, "dur": 45.013,
    "args": {
      "Python parent id": 2742, "Python id": 2745, "finished": true, "Ev Idx": 9575
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_16", "pid": 870, "tid": 870,
    "ts": 7429770618086.288, "dur": 1222.203,
    "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": 870, "tid": 870,
    "ts": 7429770618088.177, "dur": 1219.480,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770618088.479, "dur": 0.575,
    "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": 870, "tid": 870,
    "ts": 7429770618090.610, "dur": 1216.355,
    "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": 870, "tid": 870,
    "ts": 7429770618091.030, "dur": 0.748,
    "args": {
      "Python parent id": 2749, "Python id": 2750, "finished": true, "Ev Idx": 9580
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_32", "pid": 870, "tid": 870,
    "ts": 7429770618094.926, "dur": 85.125,
    "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": 870, "tid": 870,
    "ts": 7429770618095.672, "dur": 84.117,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770618095.815, "dur": 0.163,
    "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": 870, "tid": 870,
    "ts": 7429770618096.867, "dur": 82.129,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770618097.402, "dur": 2.871,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770618100.638, "dur": 15.080,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770618116.059, "dur": 15.531,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770618145.111, "dur": 9.302,
    "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": 870, "tid": 870,
    "ts": 7429770618166.092, "dur": 0.268,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770618166.709, "dur": 1.253,
    "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": 870, "tid": 870,
    "ts": 7429770618180.625, "dur": 0.508,
    "args": {
      "Python parent id": 2749, "Python id": 2761, "finished": true, "Ev Idx": 9591
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_16", "pid": 870, "tid": 870,
    "ts": 7429770618183.917, "dur": 831.377,
    "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": 870, "tid": 870,
    "ts": 7429770618185.386, "dur": 829.034,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770618185.647, "dur": 0.284,
    "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": 870, "tid": 870,
    "ts": 7429770618187.888, "dur": 825.824,
    "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": 870, "tid": 870,
    "ts": 7429770618191.075, "dur": 0.368,
    "args": {
      "Python parent id": 2765, "Python id": 2766, "finished": true, "Ev Idx": 9596
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_112", "pid": 870, "tid": 870,
    "ts": 7429770618194.147, "dur": 118.060,
    "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": 870, "tid": 870,
    "ts": 7429770618194.788, "dur": 116.944,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770618194.932, "dur": 0.160,
    "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": 870, "tid": 870,
    "ts": 7429770618196.257, "dur": 115.258,
    "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": 870, "tid": 870,
    "ts": 7429770618196.631, "dur": 0.161,
    "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": 870, "tid": 870,
    "ts": 7429770618196.990, "dur": 0.142,
    "args": {
      "Python parent id": 2770, "Python id": 2772, "finished": true, "Ev Idx": 9602
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770618197.250, "dur": 114.177,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770618312.977, "dur": 4.058,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770618317.430, "dur": 3.961,
    "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": 870, "tid": 870,
    "ts": 7429770618322.587, "dur": 0.392,
    "args": {
      "Python parent id": 2765, "Python id": 2776, "finished": true, "Ev Idx": 9606
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_113", "pid": 870, "tid": 870,
    "ts": 7429770618326.817, "dur": 42.195,
    "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": 870, "tid": 870,
    "ts": 7429770618327.695, "dur": 41.062,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770618327.991, "dur": 0.408,
    "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": 870, "tid": 870,
    "ts": 7429770618329.360, "dur": 39.172,
    "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": 870, "tid": 870,
    "ts": 7429770618329.622, "dur": 0.163,
    "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": 870, "tid": 870,
    "ts": 7429770618329.990, "dur": 0.139,
    "args": {
      "Python parent id": 2780, "Python id": 2782, "finished": true, "Ev Idx": 9612
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770618330.259, "dur": 38.208,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770618369.489, "dur": 2.607,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770618372.748, "dur": 3.295,
    "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": 870, "tid": 870,
    "ts": 7429770618376.776, "dur": 0.373,
    "args": {
      "Python parent id": 2765, "Python id": 2786, "finished": true, "Ev Idx": 9616
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_114", "pid": 870, "tid": 870,
    "ts": 7429770618379.969, "dur": 36.675,
    "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": 870, "tid": 870,
    "ts": 7429770618380.359, "dur": 36.004,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770618380.594, "dur": 0.273,
    "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": 870, "tid": 870,
    "ts": 7429770618381.429, "dur": 34.828,
    "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": 870, "tid": 870,
    "ts": 7429770618381.672, "dur": 0.160,
    "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": 870, "tid": 870,
    "ts": 7429770618382.029, "dur": 0.154,
    "args": {
      "Python parent id": 2790, "Python id": 2792, "finished": true, "Ev Idx": 9622
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770618382.287, "dur": 33.921,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770618417.305, "dur": 2.706,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770618420.433, "dur": 2.781,
    "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": 870, "tid": 870,
    "ts": 7429770618424.226, "dur": 155.517,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770618424.566, "dur": 4.125,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770618428.988, "dur": 1.947,
    "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": 870, "tid": 870,
    "ts": 7429770618443.202, "dur": 54.590,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770618473.932, "dur": 22.611,
    "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": 870, "tid": 870,
    "ts": 7429770618528.827, "dur": 33.133,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770618547.692, "dur": 13.293,
    "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": 870, "tid": 870,
    "ts": 7429770618584.055, "dur": 4.067,
    "args": {
      "Python parent id": 2765, "Python id": 2803, "finished": true, "Ev Idx": 9633
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770618584.835, "dur": 0.245,
    "args": {
      "Python parent id": 2803, "Python id": 2804, "finished": true, "Ev Idx": 9634
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770618585.447, "dur": 0.051,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770618586.851, "dur": 0.377,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770618587.682, "dur": 0.138,
    "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": 870, "tid": 870,
    "ts": 7429770618589.706, "dur": 7.154,
    "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": 870, "tid": 870,
    "ts": 7429770618593.122, "dur": 3.647,
    "args": {
      "Python parent id": 2808, "Python id": 2809, "finished": true, "Ev Idx": 9639
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770618593.433, "dur": 1.546,
    "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": 870, "tid": 870,
    "ts": 7429770618594.057, "dur": 0.718,
    "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": 870, "tid": 870,
    "ts": 7429770618595.284, "dur": 0.481,
    "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": 870, "tid": 870,
    "ts": 7429770618596.119, "dur": 0.566,
    "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": 870, "tid": 870,
    "ts": 7429770618599.710, "dur": 353.137,
    "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": 870, "tid": 870,
    "ts": 7429770618600.266, "dur": 47.163,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770618619.751, "dur": 5.020,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770618625.646, "dur": 21.639,
    "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": 870, "tid": 870,
    "ts": 7429770618648.295, "dur": 32.612,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770618663.858, "dur": 2.887,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770618667.275, "dur": 13.573,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770618681.881, "dur": 2.728,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770618684.865, "dur": 100.513,
    "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": 870, "tid": 870,
    "ts": 7429770618855.892, "dur": 17.750,
    "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": 870, "tid": 870,
    "ts": 7429770618856.155, "dur": 0.254,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770618857.076, "dur": 16.402,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770618874.633, "dur": 1.926,
    "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": 870, "tid": 870,
    "ts": 7429770618878.181, "dur": 10.694,
    "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": 870, "tid": 870,
    "ts": 7429770618878.342, "dur": 0.132,
    "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": 870, "tid": 870,
    "ts": 7429770618884.266, "dur": 1.169,
    "args": {
      "Python parent id": 2827, "Python id": 2829, "finished": true, "Ev Idx": 9659
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770618884.906, "dur": 0.441,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770618885.919, "dur": 2.885,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770618889.544, "dur": 43.346,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770618933.522, "dur": 3.237,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770618937.335, "dur": 14.505,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770618956.657, "dur": 3.005,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770618960.413, "dur": 0.272,
    "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": 870, "tid": 870,
    "ts": 7429770618961.638, "dur": 1.059,
    "args": {
      "Python parent id": 2765, "Python id": 2837, "finished": true, "Ev Idx": 9667
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_115", "pid": 870, "tid": 870,
    "ts": 7429770618967.841, "dur": 45.321,
    "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": 870, "tid": 870,
    "ts": 7429770618969.211, "dur": 43.536,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770618969.677, "dur": 1.228,
    "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": 870, "tid": 870,
    "ts": 7429770618972.216, "dur": 40.216,
    "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": 870, "tid": 870,
    "ts": 7429770618972.652, "dur": 0.222,
    "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": 870, "tid": 870,
    "ts": 7429770618973.147, "dur": 0.146,
    "args": {
      "Python parent id": 2841, "Python id": 2843, "finished": true, "Ev Idx": 9673
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770618973.549, "dur": 38.771,
    "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": 870, "tid": 870,
    "ts": 7429770619028.059, "dur": 0.752,
    "args": {
      "Python parent id": 2749, "Python id": 2845, "finished": true, "Ev Idx": 9675
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_33", "pid": 870, "tid": 870,
    "ts": 7429770619031.901, "dur": 85.185,
    "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": 870, "tid": 870,
    "ts": 7429770619032.472, "dur": 84.341,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770619032.696, "dur": 0.333,
    "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": 870, "tid": 870,
    "ts": 7429770619034.448, "dur": 81.538,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770619035.119, "dur": 1.840,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770619037.358, "dur": 14.069,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770619051.775, "dur": 15.504,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770619079.405, "dur": 12.828,
    "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": 870, "tid": 870,
    "ts": 7429770619104.887, "dur": 0.369,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770619105.660, "dur": 1.129,
    "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": 870, "tid": 870,
    "ts": 7429770619117.812, "dur": 0.393,
    "args": {
      "Python parent id": 2749, "Python id": 2856, "finished": true, "Ev Idx": 9686
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_16", "pid": 870, "tid": 870,
    "ts": 7429770619133.892, "dur": 162.188,
    "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": 870, "tid": 870,
    "ts": 7429770619134.583, "dur": 161.298,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770619134.851, "dur": 0.397,
    "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": 870, "tid": 870,
    "ts": 7429770619136.457, "dur": 159.336,
    "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": 870, "tid": 870,
    "ts": 7429770619136.946, "dur": 0.592,
    "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": 870, "tid": 870,
    "ts": 7429770619137.910, "dur": 0.288,
    "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": 870, "tid": 870,
    "ts": 7429770619138.474, "dur": 0.207,
    "args": {
      "Python parent id": 2860, "Python id": 2863, "finished": true, "Ev Idx": 9693
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_116", "pid": 870, "tid": 870,
    "ts": 7429770619141.211, "dur": 38.617,
    "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": 870, "tid": 870,
    "ts": 7429770619141.754, "dur": 37.792,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770619141.979, "dur": 0.155,
    "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": 870, "tid": 870,
    "ts": 7429770619143.184, "dur": 36.192,
    "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": 870, "tid": 870,
    "ts": 7429770619143.407, "dur": 0.180,
    "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": 870, "tid": 870,
    "ts": 7429770619143.806, "dur": 0.134,
    "args": {
      "Python parent id": 2867, "Python id": 2869, "finished": true, "Ev Idx": 9699
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770619144.057, "dur": 35.272,
    "args": {
      "Python parent id": 2867, "Python id": 2870, "finished": true, "Ev Idx": 9700
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_16", "pid": 870, "tid": 870,
    "ts": 7429770619182.076, "dur": 17.692,
    "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": 870, "tid": 870,
    "ts": 7429770619182.535, "dur": 16.929,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770619182.741, "dur": 0.259,
    "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": 870, "tid": 870,
    "ts": 7429770619184.265, "dur": 15.103,
    "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": 870, "tid": 870,
    "ts": 7429770619185.062, "dur": 14.129,
    "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": 870, "tid": 870,
    "ts": 7429770619185.193, "dur": 0.196,
    "args": {
      "Python parent id": 2875, "Python id": 2876, "finished": true, "Ev Idx": 9706
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770619186.306, "dur": 12.789,
    "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": 870, "tid": 870,
    "ts": 7429770619200.425, "dur": 0.351,
    "args": {
      "Python parent id": 2860, "Python id": 2878, "finished": true, "Ev Idx": 9708
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_117", "pid": 870, "tid": 870,
    "ts": 7429770619202.780, "dur": 32.003,
    "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": 870, "tid": 870,
    "ts": 7429770619203.239, "dur": 31.322,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770619203.395, "dur": 0.363,
    "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": 870, "tid": 870,
    "ts": 7429770619204.747, "dur": 29.662,
    "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": 870, "tid": 870,
    "ts": 7429770619204.971, "dur": 0.145,
    "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": 870, "tid": 870,
    "ts": 7429770619205.317, "dur": 0.149,
    "args": {
      "Python parent id": 2882, "Python id": 2884, "finished": true, "Ev Idx": 9714
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770619205.575, "dur": 28.791,
    "args": {
      "Python parent id": 2882, "Python id": 2885, "finished": true, "Ev Idx": 9715
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_118", "pid": 870, "tid": 870,
    "ts": 7429770619245.981, "dur": 49.479,
    "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": 870, "tid": 870,
    "ts": 7429770619246.489, "dur": 48.747,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770619246.643, "dur": 0.245,
    "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": 870, "tid": 870,
    "ts": 7429770619247.673, "dur": 47.454,
    "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": 870, "tid": 870,
    "ts": 7429770619247.920, "dur": 0.150,
    "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": 870, "tid": 870,
    "ts": 7429770619248.273, "dur": 0.128,
    "args": {
      "Python parent id": 2889, "Python id": 2891, "finished": true, "Ev Idx": 9721
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770619248.507, "dur": 46.562,
    "args": {
      "Python parent id": 2889, "Python id": 2892, "finished": true, "Ev Idx": 9722
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_17", "pid": 870, "tid": 870,
    "ts": 7429770619314.457, "dur": 1100.230,
    "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": 870, "tid": 870,
    "ts": 7429770619315.867, "dur": 1097.814,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770619316.076, "dur": 0.254,
    "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": 870, "tid": 870,
    "ts": 7429770619317.778, "dur": 1095.147,
    "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": 870, "tid": 870,
    "ts": 7429770619318.125, "dur": 0.571,
    "args": {
      "Python parent id": 2896, "Python id": 2897, "finished": true, "Ev Idx": 9727
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_34", "pid": 870, "tid": 870,
    "ts": 7429770619320.861, "dur": 73.423,
    "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": 870, "tid": 870,
    "ts": 7429770619321.445, "dur": 72.481,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770619321.623, "dur": 0.143,
    "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": 870, "tid": 870,
    "ts": 7429770619322.835, "dur": 70.348,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770619323.319, "dur": 1.574,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770619325.194, "dur": 10.363,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770619335.918, "dur": 13.381,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770619361.786, "dur": 9.413,
    "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": 870, "tid": 870,
    "ts": 7429770619381.315, "dur": 0.228,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770619382.164, "dur": 1.104,
    "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": 870, "tid": 870,
    "ts": 7429770619394.981, "dur": 0.351,
    "args": {
      "Python parent id": 2896, "Python id": 2908, "finished": true, "Ev Idx": 9738
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_17", "pid": 870, "tid": 870,
    "ts": 7429770619397.919, "dur": 725.615,
    "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": 870, "tid": 870,
    "ts": 7429770619399.510, "dur": 723.218,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770619399.711, "dur": 0.240,
    "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": 870, "tid": 870,
    "ts": 7429770619401.899, "dur": 720.067,
    "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": 870, "tid": 870,
    "ts": 7429770619405.086, "dur": 0.386,
    "args": {
      "Python parent id": 2912, "Python id": 2913, "finished": true, "Ev Idx": 9743
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_119", "pid": 870, "tid": 870,
    "ts": 7429770619407.368, "dur": 42.713,
    "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": 870, "tid": 870,
    "ts": 7429770619407.899, "dur": 41.884,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770619408.064, "dur": 0.165,
    "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": 870, "tid": 870,
    "ts": 7429770619409.174, "dur": 40.458,
    "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": 870, "tid": 870,
    "ts": 7429770619409.657, "dur": 0.143,
    "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": 870, "tid": 870,
    "ts": 7429770619410.018, "dur": 0.148,
    "args": {
      "Python parent id": 2917, "Python id": 2919, "finished": true, "Ev Idx": 9749
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770619410.266, "dur": 39.318,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770619450.784, "dur": 3.638,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770619454.786, "dur": 3.476,
    "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": 870, "tid": 870,
    "ts": 7429770619459.241, "dur": 0.327,
    "args": {
      "Python parent id": 2912, "Python id": 2923, "finished": true, "Ev Idx": 9753
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_120", "pid": 870, "tid": 870,
    "ts": 7429770619461.823, "dur": 35.710,
    "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": 870, "tid": 870,
    "ts": 7429770619462.469, "dur": 34.856,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770619462.726, "dur": 0.238,
    "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": 870, "tid": 870,
    "ts": 7429770619463.847, "dur": 33.385,
    "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": 870, "tid": 870,
    "ts": 7429770619464.035, "dur": 0.169,
    "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": 870, "tid": 870,
    "ts": 7429770619464.390, "dur": 0.303,
    "args": {
      "Python parent id": 2927, "Python id": 2929, "finished": true, "Ev Idx": 9759
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770619464.796, "dur": 32.376,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770619497.971, "dur": 2.469,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770619500.939, "dur": 2.377,
    "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": 870, "tid": 870,
    "ts": 7429770619504.292, "dur": 0.552,
    "args": {
      "Python parent id": 2912, "Python id": 2933, "finished": true, "Ev Idx": 9763
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_121", "pid": 870, "tid": 870,
    "ts": 7429770619506.676, "dur": 36.689,
    "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": 870, "tid": 870,
    "ts": 7429770619507.070, "dur": 36.089,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770619507.244, "dur": 0.219,
    "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": 870, "tid": 870,
    "ts": 7429770619508.510, "dur": 34.480,
    "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": 870, "tid": 870,
    "ts": 7429770619508.763, "dur": 0.124,
    "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": 870, "tid": 870,
    "ts": 7429770619509.073, "dur": 0.162,
    "args": {
      "Python parent id": 2937, "Python id": 2939, "finished": true, "Ev Idx": 9769
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770619509.342, "dur": 33.589,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770619543.933, "dur": 2.274,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770619546.787, "dur": 2.541,
    "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": 870, "tid": 870,
    "ts": 7429770619550.447, "dur": 153.328,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770619550.666, "dur": 3.516,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770619554.465, "dur": 1.688,
    "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": 870, "tid": 870,
    "ts": 7429770619570.202, "dur": 49.379,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770619597.894, "dur": 20.527,
    "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": 870, "tid": 870,
    "ts": 7429770619650.674, "dur": 33.814,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770619669.633, "dur": 13.828,
    "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": 870, "tid": 870,
    "ts": 7429770619708.006, "dur": 4.060,
    "args": {
      "Python parent id": 2912, "Python id": 2950, "finished": true, "Ev Idx": 9780
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770619708.983, "dur": 0.306,
    "args": {
      "Python parent id": 2950, "Python id": 2951, "finished": true, "Ev Idx": 9781
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770619709.657, "dur": 0.059,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770619711.055, "dur": 0.235,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770619711.684, "dur": 0.112,
    "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": 870, "tid": 870,
    "ts": 7429770619713.413, "dur": 6.587,
    "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": 870, "tid": 870,
    "ts": 7429770619716.683, "dur": 3.223,
    "args": {
      "Python parent id": 2955, "Python id": 2956, "finished": true, "Ev Idx": 9786
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770619716.952, "dur": 1.288,
    "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": 870, "tid": 870,
    "ts": 7429770619717.402, "dur": 0.641,
    "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": 870, "tid": 870,
    "ts": 7429770619718.595, "dur": 0.445,
    "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": 870, "tid": 870,
    "ts": 7429770619719.431, "dur": 0.392,
    "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": 870, "tid": 870,
    "ts": 7429770619722.592, "dur": 336.742,
    "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": 870, "tid": 870,
    "ts": 7429770619723.238, "dur": 44.251,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770619740.861, "dur": 4.694,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770619746.394, "dur": 20.965,
    "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": 870, "tid": 870,
    "ts": 7429770619768.448, "dur": 29.129,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770619782.654, "dur": 2.574,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770619785.693, "dur": 11.824,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770619798.455, "dur": 3.817,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770619802.537, "dur": 117.633,
    "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": 870, "tid": 870,
    "ts": 7429770619960.136, "dur": 17.959,
    "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": 870, "tid": 870,
    "ts": 7429770619960.359, "dur": 0.208,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770619961.244, "dur": 16.675,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770619979.089, "dur": 1.807,
    "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": 870, "tid": 870,
    "ts": 7429770619982.317, "dur": 10.494,
    "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": 870, "tid": 870,
    "ts": 7429770619982.475, "dur": 0.131,
    "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": 870, "tid": 870,
    "ts": 7429770619988.489, "dur": 1.087,
    "args": {
      "Python parent id": 2974, "Python id": 2976, "finished": true, "Ev Idx": 9806
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770619989.097, "dur": 0.392,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770619990.043, "dur": 2.698,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770619993.411, "dur": 45.630,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770620039.697, "dur": 3.283,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770620043.465, "dur": 14.838,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770620062.516, "dur": 3.445,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770620066.791, "dur": 0.272,
    "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": 870, "tid": 870,
    "ts": 7429770620068.054, "dur": 1.001,
    "args": {
      "Python parent id": 2912, "Python id": 2984, "finished": true, "Ev Idx": 9814
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_122", "pid": 870, "tid": 870,
    "ts": 7429770620074.139, "dur": 47.063,
    "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": 870, "tid": 870,
    "ts": 7429770620075.510, "dur": 45.310,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770620076.047, "dur": 1.070,
    "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": 870, "tid": 870,
    "ts": 7429770620078.435, "dur": 42.131,
    "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": 870, "tid": 870,
    "ts": 7429770620078.828, "dur": 0.195,
    "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": 870, "tid": 870,
    "ts": 7429770620079.258, "dur": 0.145,
    "args": {
      "Python parent id": 2988, "Python id": 2990, "finished": true, "Ev Idx": 9820
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770620079.593, "dur": 40.811,
    "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": 870, "tid": 870,
    "ts": 7429770620135.923, "dur": 0.617,
    "args": {
      "Python parent id": 2896, "Python id": 2992, "finished": true, "Ev Idx": 9822
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_35", "pid": 870, "tid": 870,
    "ts": 7429770620139.280, "dur": 84.545,
    "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": 870, "tid": 870,
    "ts": 7429770620139.828, "dur": 83.653,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770620140.094, "dur": 0.383,
    "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": 870, "tid": 870,
    "ts": 7429770620142.197, "dur": 80.571,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770620142.892, "dur": 1.918,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770620145.202, "dur": 12.806,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770620158.336, "dur": 17.107,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770620187.608, "dur": 10.744,
    "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": 870, "tid": 870,
    "ts": 7429770620209.408, "dur": 0.280,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770620210.114, "dur": 1.083,
    "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": 870, "tid": 870,
    "ts": 7429770620224.559, "dur": 0.350,
    "args": {
      "Python parent id": 2896, "Python id": 3003, "finished": true, "Ev Idx": 9833
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_17", "pid": 870, "tid": 870,
    "ts": 7429770620227.020, "dur": 175.543,
    "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": 870, "tid": 870,
    "ts": 7429770620227.468, "dur": 174.939,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770620227.668, "dur": 0.326,
    "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": 870, "tid": 870,
    "ts": 7429770620229.150, "dur": 173.138,
    "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": 870, "tid": 870,
    "ts": 7429770620229.544, "dur": 0.367,
    "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": 870, "tid": 870,
    "ts": 7429770620230.236, "dur": 0.321,
    "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": 870, "tid": 870,
    "ts": 7429770620230.828, "dur": 0.250,
    "args": {
      "Python parent id": 3007, "Python id": 3010, "finished": true, "Ev Idx": 9840
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_123", "pid": 870, "tid": 870,
    "ts": 7429770620233.397, "dur": 43.798,
    "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": 870, "tid": 870,
    "ts": 7429770620234.072, "dur": 42.791,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770620234.236, "dur": 0.183,
    "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": 870, "tid": 870,
    "ts": 7429770620235.294, "dur": 41.442,
    "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": 870, "tid": 870,
    "ts": 7429770620235.556, "dur": 0.122,
    "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": 870, "tid": 870,
    "ts": 7429770620235.871, "dur": 0.147,
    "args": {
      "Python parent id": 3014, "Python id": 3016, "finished": true, "Ev Idx": 9846
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770620236.165, "dur": 40.475,
    "args": {
      "Python parent id": 3014, "Python id": 3017, "finished": true, "Ev Idx": 9847
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_17", "pid": 870, "tid": 870,
    "ts": 7429770620278.951, "dur": 16.936,
    "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": 870, "tid": 870,
    "ts": 7429770620279.481, "dur": 16.187,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770620279.661, "dur": 0.243,
    "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": 870, "tid": 870,
    "ts": 7429770620281.201, "dur": 14.357,
    "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": 870, "tid": 870,
    "ts": 7429770620282.036, "dur": 13.378,
    "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": 870, "tid": 870,
    "ts": 7429770620282.138, "dur": 0.207,
    "args": {
      "Python parent id": 3022, "Python id": 3023, "finished": true, "Ev Idx": 9853
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770620283.213, "dur": 12.113,
    "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": 870, "tid": 870,
    "ts": 7429770620296.456, "dur": 0.417,
    "args": {
      "Python parent id": 3007, "Python id": 3025, "finished": true, "Ev Idx": 9855
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_124", "pid": 870, "tid": 870,
    "ts": 7429770620298.773, "dur": 40.225,
    "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": 870, "tid": 870,
    "ts": 7429770620299.144, "dur": 39.586,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770620299.301, "dur": 0.194,
    "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": 870, "tid": 870,
    "ts": 7429770620300.268, "dur": 38.340,
    "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": 870, "tid": 870,
    "ts": 7429770620300.617, "dur": 0.123,
    "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": 870, "tid": 870,
    "ts": 7429770620300.925, "dur": 0.114,
    "args": {
      "Python parent id": 3029, "Python id": 3031, "finished": true, "Ev Idx": 9861
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770620301.310, "dur": 37.255,
    "args": {
      "Python parent id": 3029, "Python id": 3032, "finished": true, "Ev Idx": 9862
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_125", "pid": 870, "tid": 870,
    "ts": 7429770620350.570, "dur": 51.425,
    "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": 870, "tid": 870,
    "ts": 7429770620351.160, "dur": 50.646,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770620351.417, "dur": 0.259,
    "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": 870, "tid": 870,
    "ts": 7429770620352.318, "dur": 49.367,
    "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": 870, "tid": 870,
    "ts": 7429770620352.540, "dur": 0.302,
    "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": 870, "tid": 870,
    "ts": 7429770620353.050, "dur": 0.134,
    "args": {
      "Python parent id": 3036, "Python id": 3038, "finished": true, "Ev Idx": 9868
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770620353.291, "dur": 48.346,
    "args": {
      "Python parent id": 3036, "Python id": 3039, "finished": true, "Ev Idx": 9869
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_18", "pid": 870, "tid": 870,
    "ts": 7429770620419.982, "dur": 1139.270,
    "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": 870, "tid": 870,
    "ts": 7429770620421.123, "dur": 1137.224,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770620421.414, "dur": 0.280,
    "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": 870, "tid": 870,
    "ts": 7429770620422.911, "dur": 1134.568,
    "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": 870, "tid": 870,
    "ts": 7429770620423.206, "dur": 0.525,
    "args": {
      "Python parent id": 3043, "Python id": 3044, "finished": true, "Ev Idx": 9874
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_36", "pid": 870, "tid": 870,
    "ts": 7429770620426.085, "dur": 70.855,
    "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": 870, "tid": 870,
    "ts": 7429770620426.512, "dur": 70.176,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770620426.643, "dur": 0.148,
    "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": 870, "tid": 870,
    "ts": 7429770620427.587, "dur": 68.328,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770620428.043, "dur": 1.571,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770620429.819, "dur": 9.946,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770620440.183, "dur": 13.127,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770620464.674, "dur": 8.427,
    "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": 870, "tid": 870,
    "ts": 7429770620483.927, "dur": 0.182,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770620484.524, "dur": 0.966,
    "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": 870, "tid": 870,
    "ts": 7429770620497.496, "dur": 0.364,
    "args": {
      "Python parent id": 3043, "Python id": 3055, "finished": true, "Ev Idx": 9885
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_18", "pid": 870, "tid": 870,
    "ts": 7429770620500.455, "dur": 695.043,
    "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": 870, "tid": 870,
    "ts": 7429770620501.887, "dur": 692.890,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770620502.107, "dur": 0.262,
    "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": 870, "tid": 870,
    "ts": 7429770620504.053, "dur": 690.066,
    "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": 870, "tid": 870,
    "ts": 7429770620506.784, "dur": 0.469,
    "args": {
      "Python parent id": 3059, "Python id": 3060, "finished": true, "Ev Idx": 9890
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_126", "pid": 870, "tid": 870,
    "ts": 7429770620509.523, "dur": 37.049,
    "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": 870, "tid": 870,
    "ts": 7429770620509.995, "dur": 36.329,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770620510.149, "dur": 0.178,
    "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": 870, "tid": 870,
    "ts": 7429770620511.042, "dur": 35.173,
    "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": 870, "tid": 870,
    "ts": 7429770620511.369, "dur": 0.121,
    "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": 870, "tid": 870,
    "ts": 7429770620511.660, "dur": 0.132,
    "args": {
      "Python parent id": 3064, "Python id": 3066, "finished": true, "Ev Idx": 9896
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770620511.895, "dur": 34.273,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770620547.190, "dur": 3.188,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770620550.800, "dur": 3.394,
    "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": 870, "tid": 870,
    "ts": 7429770620555.162, "dur": 0.337,
    "args": {
      "Python parent id": 3059, "Python id": 3070, "finished": true, "Ev Idx": 9900
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_127", "pid": 870, "tid": 870,
    "ts": 7429770620557.398, "dur": 36.733,
    "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": 870, "tid": 870,
    "ts": 7429770620557.830, "dur": 36.094,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770620558.014, "dur": 0.251,
    "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": 870, "tid": 870,
    "ts": 7429770620559.174, "dur": 34.644,
    "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": 870, "tid": 870,
    "ts": 7429770620559.401, "dur": 0.122,
    "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": 870, "tid": 870,
    "ts": 7429770620559.725, "dur": 0.122,
    "args": {
      "Python parent id": 3074, "Python id": 3076, "finished": true, "Ev Idx": 9906
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770620559.955, "dur": 33.823,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770620594.845, "dur": 2.492,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770620597.761, "dur": 2.608,
    "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": 870, "tid": 870,
    "ts": 7429770620601.033, "dur": 0.382,
    "args": {
      "Python parent id": 3059, "Python id": 3080, "finished": true, "Ev Idx": 9910
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_128", "pid": 870, "tid": 870,
    "ts": 7429770620603.864, "dur": 35.287,
    "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": 870, "tid": 870,
    "ts": 7429770620604.517, "dur": 34.313,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770620604.706, "dur": 0.206,
    "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": 870, "tid": 870,
    "ts": 7429770620605.492, "dur": 33.221,
    "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": 870, "tid": 870,
    "ts": 7429770620606.950, "dur": 0.131,
    "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": 870, "tid": 870,
    "ts": 7429770620607.269, "dur": 0.110,
    "args": {
      "Python parent id": 3084, "Python id": 3086, "finished": true, "Ev Idx": 9916
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770620607.483, "dur": 31.169,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770620639.721, "dur": 2.793,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770620642.939, "dur": 2.560,
    "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": 870, "tid": 870,
    "ts": 7429770620646.858, "dur": 139.572,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770620647.113, "dur": 2.997,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770620650.407, "dur": 1.672,
    "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": 870, "tid": 870,
    "ts": 7429770620663.215, "dur": 45.398,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770620686.631, "dur": 20.995,
    "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": 870, "tid": 870,
    "ts": 7429770620736.700, "dur": 31.200,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770620753.352, "dur": 13.502,
    "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": 870, "tid": 870,
    "ts": 7429770620790.417, "dur": 3.423,
    "args": {
      "Python parent id": 3059, "Python id": 3097, "finished": true, "Ev Idx": 9927
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770620791.127, "dur": 0.231,
    "args": {
      "Python parent id": 3097, "Python id": 3098, "finished": true, "Ev Idx": 9928
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770620791.641, "dur": 0.101,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770620792.926, "dur": 0.209,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770620793.520, "dur": 0.110,
    "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": 870, "tid": 870,
    "ts": 7429770620795.172, "dur": 6.333,
    "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": 870, "tid": 870,
    "ts": 7429770620798.028, "dur": 3.346,
    "args": {
      "Python parent id": 3102, "Python id": 3103, "finished": true, "Ev Idx": 9933
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770620798.292, "dur": 1.401,
    "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": 870, "tid": 870,
    "ts": 7429770620798.857, "dur": 0.603,
    "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": 870, "tid": 870,
    "ts": 7429770620800.020, "dur": 0.505,
    "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": 870, "tid": 870,
    "ts": 7429770620800.852, "dur": 0.437,
    "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": 870, "tid": 870,
    "ts": 7429770620803.987, "dur": 328.185,
    "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": 870, "tid": 870,
    "ts": 7429770620804.567, "dur": 42.884,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770620823.470, "dur": 4.399,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770620828.676, "dur": 18.618,
    "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": 870, "tid": 870,
    "ts": 7429770620848.334, "dur": 32.114,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770620862.576, "dur": 2.930,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770620866.003, "dur": 14.378,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770620881.294, "dur": 2.824,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770620884.374, "dur": 110.037,
    "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": 870, "tid": 870,
    "ts": 7429770621034.447, "dur": 16.496,
    "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": 870, "tid": 870,
    "ts": 7429770621034.621, "dur": 0.203,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770621035.471, "dur": 15.290,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770621051.876, "dur": 1.918,
    "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": 870, "tid": 870,
    "ts": 7429770621055.223, "dur": 10.212,
    "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": 870, "tid": 870,
    "ts": 7429770621055.384, "dur": 0.087,
    "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": 870, "tid": 870,
    "ts": 7429770621061.152, "dur": 1.160,
    "args": {
      "Python parent id": 3121, "Python id": 3123, "finished": true, "Ev Idx": 9953
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770621061.732, "dur": 0.519,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770621062.645, "dur": 2.718,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770621066.041, "dur": 45.870,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770621112.544, "dur": 3.400,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770621116.576, "dur": 14.612,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770621135.450, "dur": 3.190,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770621139.421, "dur": 0.300,
    "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": 870, "tid": 870,
    "ts": 7429770621140.672, "dur": 1.066,
    "args": {
      "Python parent id": 3059, "Python id": 3131, "finished": true, "Ev Idx": 9961
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_129", "pid": 870, "tid": 870,
    "ts": 7429770621146.234, "dur": 47.213,
    "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": 870, "tid": 870,
    "ts": 7429770621147.516, "dur": 45.590,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770621147.994, "dur": 1.199,
    "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": 870, "tid": 870,
    "ts": 7429770621150.594, "dur": 42.205,
    "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": 870, "tid": 870,
    "ts": 7429770621150.942, "dur": 0.208,
    "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": 870, "tid": 870,
    "ts": 7429770621151.419, "dur": 0.164,
    "args": {
      "Python parent id": 3135, "Python id": 3137, "finished": true, "Ev Idx": 9967
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770621151.747, "dur": 40.942,
    "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": 870, "tid": 870,
    "ts": 7429770621207.618, "dur": 0.429,
    "args": {
      "Python parent id": 3043, "Python id": 3139, "finished": true, "Ev Idx": 9969
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_37", "pid": 870, "tid": 870,
    "ts": 7429770621210.757, "dur": 82.047,
    "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": 870, "tid": 870,
    "ts": 7429770621211.379, "dur": 81.027,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770621211.669, "dur": 0.420,
    "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": 870, "tid": 870,
    "ts": 7429770621212.927, "dur": 78.711,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770621213.472, "dur": 1.671,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770621215.425, "dur": 13.956,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770621229.770, "dur": 15.223,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770621256.764, "dur": 12.838,
    "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": 870, "tid": 870,
    "ts": 7429770621280.702, "dur": 0.273,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770621281.452, "dur": 1.221,
    "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": 870, "tid": 870,
    "ts": 7429770621293.572, "dur": 0.346,
    "args": {
      "Python parent id": 3043, "Python id": 3150, "finished": true, "Ev Idx": 9980
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_18", "pid": 870, "tid": 870,
    "ts": 7429770621295.799, "dur": 249.810,
    "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": 870, "tid": 870,
    "ts": 7429770621296.328, "dur": 249.113,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770621296.566, "dur": 0.279,
    "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": 870, "tid": 870,
    "ts": 7429770621297.748, "dur": 247.574,
    "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": 870, "tid": 870,
    "ts": 7429770621298.125, "dur": 0.364,
    "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": 870, "tid": 870,
    "ts": 7429770621298.895, "dur": 0.313,
    "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": 870, "tid": 870,
    "ts": 7429770621299.467, "dur": 0.229,
    "args": {
      "Python parent id": 3154, "Python id": 3157, "finished": true, "Ev Idx": 9987
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_130", "pid": 870, "tid": 870,
    "ts": 7429770621301.993, "dur": 115.993,
    "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": 870, "tid": 870,
    "ts": 7429770621302.686, "dur": 114.777,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770621302.819, "dur": 0.154,
    "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": 870, "tid": 870,
    "ts": 7429770621303.807, "dur": 113.452,
    "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": 870, "tid": 870,
    "ts": 7429770621304.068, "dur": 0.126,
    "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": 870, "tid": 870,
    "ts": 7429770621304.566, "dur": 0.134,
    "args": {
      "Python parent id": 3161, "Python id": 3163, "finished": true, "Ev Idx": 9993
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770621304.858, "dur": 112.298,
    "args": {
      "Python parent id": 3161, "Python id": 3164, "finished": true, "Ev Idx": 9994
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_18", "pid": 870, "tid": 870,
    "ts": 7429770621420.655, "dur": 18.331,
    "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": 870, "tid": 870,
    "ts": 7429770621421.233, "dur": 17.513,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770621421.503, "dur": 0.424,
    "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": 870, "tid": 870,
    "ts": 7429770621422.946, "dur": 15.673,
    "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": 870, "tid": 870,
    "ts": 7429770621424.104, "dur": 14.243,
    "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": 870, "tid": 870,
    "ts": 7429770621424.301, "dur": 0.144,
    "args": {
      "Python parent id": 3169, "Python id": 3170, "finished": true, "Ev Idx": 10000
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770621425.271, "dur": 12.992,
    "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": 870, "tid": 870,
    "ts": 7429770621439.792, "dur": 0.438,
    "args": {
      "Python parent id": 3154, "Python id": 3172, "finished": true, "Ev Idx": 10002
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_131", "pid": 870, "tid": 870,
    "ts": 7429770621442.368, "dur": 41.421,
    "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": 870, "tid": 870,
    "ts": 7429770621442.886, "dur": 40.665,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770621443.058, "dur": 0.346,
    "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": 870, "tid": 870,
    "ts": 7429770621444.165, "dur": 39.285,
    "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": 870, "tid": 870,
    "ts": 7429770621444.405, "dur": 0.174,
    "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": 870, "tid": 870,
    "ts": 7429770621444.830, "dur": 0.140,
    "args": {
      "Python parent id": 3176, "Python id": 3178, "finished": true, "Ev Idx": 10008
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770621445.079, "dur": 38.333,
    "args": {
      "Python parent id": 3176, "Python id": 3179, "finished": true, "Ev Idx": 10009
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_132", "pid": 870, "tid": 870,
    "ts": 7429770621495.753, "dur": 49.258,
    "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": 870, "tid": 870,
    "ts": 7429770621496.309, "dur": 48.467,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770621496.500, "dur": 0.219,
    "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": 870, "tid": 870,
    "ts": 7429770621497.347, "dur": 47.289,
    "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": 870, "tid": 870,
    "ts": 7429770621497.585, "dur": 0.165,
    "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": 870, "tid": 870,
    "ts": 7429770621497.956, "dur": 0.147,
    "args": {
      "Python parent id": 3183, "Python id": 3185, "finished": true, "Ev Idx": 10015
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770621498.403, "dur": 46.186,
    "args": {
      "Python parent id": 3183, "Python id": 3186, "finished": true, "Ev Idx": 10016
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_19", "pid": 870, "tid": 870,
    "ts": 7429770621565.368, "dur": 1232.885,
    "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": 870, "tid": 870,
    "ts": 7429770621566.619, "dur": 1230.880,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770621566.807, "dur": 0.291,
    "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": 870, "tid": 870,
    "ts": 7429770621568.671, "dur": 1228.059,
    "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": 870, "tid": 870,
    "ts": 7429770621569.116, "dur": 0.517,
    "args": {
      "Python parent id": 3190, "Python id": 3191, "finished": true, "Ev Idx": 10021
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_38", "pid": 870, "tid": 870,
    "ts": 7429770621571.703, "dur": 115.852,
    "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": 870, "tid": 870,
    "ts": 7429770621572.280, "dur": 115.003,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770621572.425, "dur": 0.172,
    "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": 870, "tid": 870,
    "ts": 7429770621573.374, "dur": 113.159,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770621573.935, "dur": 1.681,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770621575.889, "dur": 11.842,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770621588.245, "dur": 14.284,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770621615.079, "dur": 9.545,
    "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": 870, "tid": 870,
    "ts": 7429770621635.939, "dur": 0.243,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770621636.681, "dur": 39.073,
    "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": 870, "tid": 870,
    "ts": 7429770621688.282, "dur": 0.334,
    "args": {
      "Python parent id": 3190, "Python id": 3202, "finished": true, "Ev Idx": 10032
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_19", "pid": 870, "tid": 870,
    "ts": 7429770621691.401, "dur": 797.736,
    "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": 870, "tid": 870,
    "ts": 7429770621693.106, "dur": 795.274,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770621693.327, "dur": 0.242,
    "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": 870, "tid": 870,
    "ts": 7429770621695.631, "dur": 791.854,
    "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": 870, "tid": 870,
    "ts": 7429770621698.987, "dur": 0.547,
    "args": {
      "Python parent id": 3206, "Python id": 3207, "finished": true, "Ev Idx": 10037
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_133", "pid": 870, "tid": 870,
    "ts": 7429770621701.918, "dur": 43.415,
    "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": 870, "tid": 870,
    "ts": 7429770621702.386, "dur": 42.742,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770621702.528, "dur": 0.163,
    "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": 870, "tid": 870,
    "ts": 7429770621703.588, "dur": 41.337,
    "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": 870, "tid": 870,
    "ts": 7429770621703.931, "dur": 0.116,
    "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": 870, "tid": 870,
    "ts": 7429770621704.271, "dur": 0.129,
    "args": {
      "Python parent id": 3211, "Python id": 3213, "finished": true, "Ev Idx": 10043
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770621704.502, "dur": 40.358,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770621746.074, "dur": 3.579,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770621750.234, "dur": 3.689,
    "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": 870, "tid": 870,
    "ts": 7429770621755.120, "dur": 0.288,
    "args": {
      "Python parent id": 3206, "Python id": 3217, "finished": true, "Ev Idx": 10047
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_134", "pid": 870, "tid": 870,
    "ts": 7429770621758.020, "dur": 39.408,
    "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": 870, "tid": 870,
    "ts": 7429770621758.822, "dur": 38.379,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770621759.022, "dur": 0.264,
    "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": 870, "tid": 870,
    "ts": 7429770621760.184, "dur": 36.904,
    "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": 870, "tid": 870,
    "ts": 7429770621760.418, "dur": 0.123,
    "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": 870, "tid": 870,
    "ts": 7429770621760.712, "dur": 0.135,
    "args": {
      "Python parent id": 3221, "Python id": 3223, "finished": true, "Ev Idx": 10053
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770621760.949, "dur": 36.078,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770621797.911, "dur": 2.629,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770621801.069, "dur": 2.478,
    "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": 870, "tid": 870,
    "ts": 7429770621804.748, "dur": 0.306,
    "args": {
      "Python parent id": 3206, "Python id": 3227, "finished": true, "Ev Idx": 10057
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_135", "pid": 870, "tid": 870,
    "ts": 7429770621821.157, "dur": 32.192,
    "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": 870, "tid": 870,
    "ts": 7429770621821.759, "dur": 31.364,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770621821.938, "dur": 0.236,
    "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": 870, "tid": 870,
    "ts": 7429770621823.009, "dur": 30.014,
    "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": 870, "tid": 870,
    "ts": 7429770621823.298, "dur": 0.141,
    "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": 870, "tid": 870,
    "ts": 7429770621823.676, "dur": 0.118,
    "args": {
      "Python parent id": 3231, "Python id": 3233, "finished": true, "Ev Idx": 10063
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770621823.900, "dur": 29.082,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770621853.943, "dur": 2.480,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770621857.051, "dur": 2.909,
    "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": 870, "tid": 870,
    "ts": 7429770621861.115, "dur": 151.835,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770621861.373, "dur": 3.698,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770621865.343, "dur": 1.867,
    "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": 870, "tid": 870,
    "ts": 7429770621881.348, "dur": 50.033,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770621907.693, "dur": 22.313,
    "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": 870, "tid": 870,
    "ts": 7429770621960.310, "dur": 31.750,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770621977.633, "dur": 13.626,
    "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": 870, "tid": 870,
    "ts": 7429770622017.544, "dur": 4.033,
    "args": {
      "Python parent id": 3206, "Python id": 3244, "finished": true, "Ev Idx": 10074
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770622018.403, "dur": 0.267,
    "args": {
      "Python parent id": 3244, "Python id": 3245, "finished": true, "Ev Idx": 10075
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770622019.045, "dur": 0.080,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770622020.543, "dur": 0.270,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770622021.220, "dur": 0.114,
    "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": 870, "tid": 870,
    "ts": 7429770622022.873, "dur": 6.993,
    "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": 870, "tid": 870,
    "ts": 7429770622026.365, "dur": 3.407,
    "args": {
      "Python parent id": 3249, "Python id": 3250, "finished": true, "Ev Idx": 10080
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770622026.664, "dur": 1.455,
    "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": 870, "tid": 870,
    "ts": 7429770622027.186, "dur": 0.705,
    "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": 870, "tid": 870,
    "ts": 7429770622028.456, "dur": 0.438,
    "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": 870, "tid": 870,
    "ts": 7429770622029.244, "dur": 0.456,
    "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": 870, "tid": 870,
    "ts": 7429770622032.632, "dur": 388.586,
    "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": 870, "tid": 870,
    "ts": 7429770622033.185, "dur": 69.097,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770622052.035, "dur": 5.855,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770622058.832, "dur": 43.332,
    "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": 870, "tid": 870,
    "ts": 7429770622103.609, "dur": 36.128,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770622119.789, "dur": 3.551,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770622123.939, "dur": 15.700,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770622140.766, "dur": 3.301,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770622144.333, "dur": 117.964,
    "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": 870, "tid": 870,
    "ts": 7429770622303.259, "dur": 27.336,
    "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": 870, "tid": 870,
    "ts": 7429770622303.542, "dur": 0.198,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770622304.505, "dur": 25.897,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770622331.935, "dur": 2.225,
    "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": 870, "tid": 870,
    "ts": 7429770622336.260, "dur": 11.997,
    "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": 870, "tid": 870,
    "ts": 7429770622336.456, "dur": 0.143,
    "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": 870, "tid": 870,
    "ts": 7429770622343.551, "dur": 1.225,
    "args": {
      "Python parent id": 3268, "Python id": 3270, "finished": true, "Ev Idx": 10100
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770622344.161, "dur": 0.538,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770622345.216, "dur": 2.973,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770622349.019, "dur": 48.780,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770622398.489, "dur": 3.688,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770622402.731, "dur": 17.470,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770622425.045, "dur": 3.362,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770622429.191, "dur": 0.247,
    "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": 870, "tid": 870,
    "ts": 7429770622430.640, "dur": 1.185,
    "args": {
      "Python parent id": 3206, "Python id": 3278, "finished": true, "Ev Idx": 10108
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_136", "pid": 870, "tid": 870,
    "ts": 7429770622437.048, "dur": 49.839,
    "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": 870, "tid": 870,
    "ts": 7429770622438.885, "dur": 47.596,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770622439.691, "dur": 1.440,
    "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": 870, "tid": 870,
    "ts": 7429770622442.811, "dur": 43.411,
    "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": 870, "tid": 870,
    "ts": 7429770622443.233, "dur": 0.258,
    "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": 870, "tid": 870,
    "ts": 7429770622443.762, "dur": 0.143,
    "args": {
      "Python parent id": 3282, "Python id": 3284, "finished": true, "Ev Idx": 10114
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770622444.046, "dur": 42.086,
    "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": 870, "tid": 870,
    "ts": 7429770622502.915, "dur": 0.505,
    "args": {
      "Python parent id": 3190, "Python id": 3286, "finished": true, "Ev Idx": 10116
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_39", "pid": 870, "tid": 870,
    "ts": 7429770622506.413, "dur": 92.297,
    "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": 870, "tid": 870,
    "ts": 7429770622507.130, "dur": 91.246,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770622507.403, "dur": 0.381,
    "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": 870, "tid": 870,
    "ts": 7429770622508.577, "dur": 88.994,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770622509.200, "dur": 1.800,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770622511.349, "dur": 16.160,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770622527.921, "dur": 18.357,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770622559.972, "dur": 11.742,
    "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": 870, "tid": 870,
    "ts": 7429770622585.014, "dur": 0.318,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770622585.769, "dur": 1.156,
    "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": 870, "tid": 870,
    "ts": 7429770622599.404, "dur": 0.407,
    "args": {
      "Python parent id": 3190, "Python id": 3297, "finished": true, "Ev Idx": 10127
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_19", "pid": 870, "tid": 870,
    "ts": 7429770622601.705, "dur": 184.401,
    "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": 870, "tid": 870,
    "ts": 7429770622602.251, "dur": 183.638,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770622602.568, "dur": 0.367,
    "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": 870, "tid": 870,
    "ts": 7429770622603.984, "dur": 181.782,
    "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": 870, "tid": 870,
    "ts": 7429770622604.421, "dur": 0.397,
    "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": 870, "tid": 870,
    "ts": 7429770622605.317, "dur": 0.344,
    "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": 870, "tid": 870,
    "ts": 7429770622605.948, "dur": 0.228,
    "args": {
      "Python parent id": 3301, "Python id": 3304, "finished": true, "Ev Idx": 10134
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_137", "pid": 870, "tid": 870,
    "ts": 7429770622608.554, "dur": 45.515,
    "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": 870, "tid": 870,
    "ts": 7429770622609.154, "dur": 44.691,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770622609.285, "dur": 0.185,
    "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": 870, "tid": 870,
    "ts": 7429770622610.204, "dur": 43.529,
    "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": 870, "tid": 870,
    "ts": 7429770622610.510, "dur": 0.163,
    "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": 870, "tid": 870,
    "ts": 7429770622610.905, "dur": 0.145,
    "args": {
      "Python parent id": 3308, "Python id": 3310, "finished": true, "Ev Idx": 10140
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770622611.172, "dur": 42.479,
    "args": {
      "Python parent id": 3308, "Python id": 3311, "finished": true, "Ev Idx": 10141
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_19", "pid": 870, "tid": 870,
    "ts": 7429770622655.826, "dur": 16.323,
    "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": 870, "tid": 870,
    "ts": 7429770622656.269, "dur": 15.721,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770622656.490, "dur": 0.243,
    "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": 870, "tid": 870,
    "ts": 7429770622657.731, "dur": 14.173,
    "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": 870, "tid": 870,
    "ts": 7429770622658.556, "dur": 13.199,
    "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": 870, "tid": 870,
    "ts": 7429770622658.707, "dur": 0.151,
    "args": {
      "Python parent id": 3316, "Python id": 3317, "finished": true, "Ev Idx": 10147
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770622659.771, "dur": 11.894,
    "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": 870, "tid": 870,
    "ts": 7429770622672.664, "dur": 0.503,
    "args": {
      "Python parent id": 3301, "Python id": 3319, "finished": true, "Ev Idx": 10149
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_138", "pid": 870, "tid": 870,
    "ts": 7429770622676.325, "dur": 43.905,
    "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": 870, "tid": 870,
    "ts": 7429770622677.164, "dur": 42.758,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770622677.346, "dur": 0.197,
    "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": 870, "tid": 870,
    "ts": 7429770622678.228, "dur": 41.603,
    "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": 870, "tid": 870,
    "ts": 7429770622678.569, "dur": 0.183,
    "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": 870, "tid": 870,
    "ts": 7429770622678.938, "dur": 0.135,
    "args": {
      "Python parent id": 3323, "Python id": 3325, "finished": true, "Ev Idx": 10155
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770622679.180, "dur": 40.589,
    "args": {
      "Python parent id": 3323, "Python id": 3326, "finished": true, "Ev Idx": 10156
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_139", "pid": 870, "tid": 870,
    "ts": 7429770622732.549, "dur": 52.937,
    "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": 870, "tid": 870,
    "ts": 7429770622733.252, "dur": 51.887,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770622733.546, "dur": 0.247,
    "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": 870, "tid": 870,
    "ts": 7429770622734.450, "dur": 50.573,
    "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": 870, "tid": 870,
    "ts": 7429770622734.691, "dur": 0.309,
    "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": 870, "tid": 870,
    "ts": 7429770622735.249, "dur": 0.148,
    "args": {
      "Python parent id": 3330, "Python id": 3332, "finished": true, "Ev Idx": 10162
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770622735.515, "dur": 49.447,
    "args": {
      "Python parent id": 3330, "Python id": 3333, "finished": true, "Ev Idx": 10163
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_20", "pid": 870, "tid": 870,
    "ts": 7429770622804.461, "dur": 1162.186,
    "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": 870, "tid": 870,
    "ts": 7429770622805.690, "dur": 1160.092,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770622805.881, "dur": 0.240,
    "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": 870, "tid": 870,
    "ts": 7429770622807.378, "dur": 1157.518,
    "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": 870, "tid": 870,
    "ts": 7429770622807.700, "dur": 0.617,
    "args": {
      "Python parent id": 3337, "Python id": 3338, "finished": true, "Ev Idx": 10168
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_40", "pid": 870, "tid": 870,
    "ts": 7429770622810.156, "dur": 75.327,
    "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": 870, "tid": 870,
    "ts": 7429770622810.718, "dur": 74.480,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770622810.866, "dur": 0.143,
    "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": 870, "tid": 870,
    "ts": 7429770622811.905, "dur": 72.562,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770622812.347, "dur": 1.693,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770622814.308, "dur": 10.596,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770622825.169, "dur": 14.895,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770622851.641, "dur": 9.140,
    "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": 870, "tid": 870,
    "ts": 7429770622871.123, "dur": 0.231,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770622871.904, "dur": 1.353,
    "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": 870, "tid": 870,
    "ts": 7429770622886.097, "dur": 0.394,
    "args": {
      "Python parent id": 3337, "Python id": 3349, "finished": true, "Ev Idx": 10179
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_20", "pid": 870, "tid": 870,
    "ts": 7429770622889.210, "dur": 718.396,
    "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": 870, "tid": 870,
    "ts": 7429770622890.687, "dur": 716.020,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770622890.912, "dur": 0.318,
    "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": 870, "tid": 870,
    "ts": 7429770622893.010, "dur": 712.855,
    "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": 870, "tid": 870,
    "ts": 7429770622895.973, "dur": 0.409,
    "args": {
      "Python parent id": 3353, "Python id": 3354, "finished": true, "Ev Idx": 10184
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_140", "pid": 870, "tid": 870,
    "ts": 7429770622898.450, "dur": 42.507,
    "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": 870, "tid": 870,
    "ts": 7429770622898.885, "dur": 41.836,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770622899.030, "dur": 0.146,
    "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": 870, "tid": 870,
    "ts": 7429770622900.180, "dur": 40.426,
    "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": 870, "tid": 870,
    "ts": 7429770622900.453, "dur": 0.140,
    "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": 870, "tid": 870,
    "ts": 7429770622900.833, "dur": 0.124,
    "args": {
      "Python parent id": 3358, "Python id": 3360, "finished": true, "Ev Idx": 10190
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770622901.076, "dur": 39.455,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770622941.560, "dur": 3.241,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770622945.188, "dur": 3.195,
    "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": 870, "tid": 870,
    "ts": 7429770622949.029, "dur": 0.334,
    "args": {
      "Python parent id": 3353, "Python id": 3364, "finished": true, "Ev Idx": 10194
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_141", "pid": 870, "tid": 870,
    "ts": 7429770622951.306, "dur": 37.084,
    "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": 870, "tid": 870,
    "ts": 7429770622951.854, "dur": 36.284,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770622952.048, "dur": 0.266,
    "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": 870, "tid": 870,
    "ts": 7429770622953.413, "dur": 34.630,
    "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": 870, "tid": 870,
    "ts": 7429770622953.652, "dur": 0.122,
    "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": 870, "tid": 870,
    "ts": 7429770622953.969, "dur": 0.142,
    "args": {
      "Python parent id": 3368, "Python id": 3370, "finished": true, "Ev Idx": 10200
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770622954.215, "dur": 33.779,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770622988.752, "dur": 2.435,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770622991.547, "dur": 2.703,
    "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": 870, "tid": 870,
    "ts": 7429770622994.903, "dur": 0.345,
    "args": {
      "Python parent id": 3353, "Python id": 3374, "finished": true, "Ev Idx": 10204
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_142", "pid": 870, "tid": 870,
    "ts": 7429770622997.190, "dur": 37.168,
    "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": 870, "tid": 870,
    "ts": 7429770622997.751, "dur": 36.371,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770622997.943, "dur": 0.235,
    "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": 870, "tid": 870,
    "ts": 7429770622998.799, "dur": 35.234,
    "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": 870, "tid": 870,
    "ts": 7429770622998.990, "dur": 0.151,
    "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": 870, "tid": 870,
    "ts": 7429770622999.469, "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": 870, "tid": 870,
    "ts": 7429770622999.684, "dur": 34.301,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770623034.958, "dur": 2.365,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770623037.945, "dur": 2.800,
    "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": 870, "tid": 870,
    "ts": 7429770623041.828, "dur": 144.242,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770623042.083, "dur": 3.376,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770623045.713, "dur": 1.700,
    "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": 870, "tid": 870,
    "ts": 7429770623059.593, "dur": 49.095,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770623086.537, "dur": 21.117,
    "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": 870, "tid": 870,
    "ts": 7429770623138.788, "dur": 29.907,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770623156.973, "dur": 10.928,
    "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": 870, "tid": 870,
    "ts": 7429770623190.220, "dur": 3.716,
    "args": {
      "Python parent id": 3353, "Python id": 3391, "finished": true, "Ev Idx": 10221
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770623190.918, "dur": 0.244,
    "args": {
      "Python parent id": 3391, "Python id": 3392, "finished": true, "Ev Idx": 10222
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770623191.586, "dur": 0.056,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770623192.926, "dur": 0.233,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770623193.564, "dur": 0.118,
    "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": 870, "tid": 870,
    "ts": 7429770623195.364, "dur": 6.349,
    "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": 870, "tid": 870,
    "ts": 7429770623198.448, "dur": 3.140,
    "args": {
      "Python parent id": 3396, "Python id": 3397, "finished": true, "Ev Idx": 10227
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770623198.735, "dur": 1.295,
    "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": 870, "tid": 870,
    "ts": 7429770623199.231, "dur": 0.672,
    "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": 870, "tid": 870,
    "ts": 7429770623200.315, "dur": 0.431,
    "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": 870, "tid": 870,
    "ts": 7429770623201.099, "dur": 0.413,
    "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": 870, "tid": 870,
    "ts": 7429770623204.593, "dur": 335.201,
    "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": 870, "tid": 870,
    "ts": 7429770623205.096, "dur": 42.014,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770623222.355, "dur": 5.046,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770623228.233, "dur": 18.749,
    "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": 870, "tid": 870,
    "ts": 7429770623248.060, "dur": 31.241,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770623262.980, "dur": 2.946,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770623266.370, "dur": 12.870,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770623280.109, "dur": 2.782,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770623283.174, "dur": 113.355,
    "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": 870, "tid": 870,
    "ts": 7429770623440.097, "dur": 17.640,
    "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": 870, "tid": 870,
    "ts": 7429770623440.339, "dur": 0.216,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770623441.232, "dur": 16.332,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770623458.551, "dur": 2.010,
    "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": 870, "tid": 870,
    "ts": 7429770623461.995, "dur": 9.634,
    "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": 870, "tid": 870,
    "ts": 7429770623462.207, "dur": 0.118,
    "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": 870, "tid": 870,
    "ts": 7429770623467.325, "dur": 0.980,
    "args": {
      "Python parent id": 3415, "Python id": 3417, "finished": true, "Ev Idx": 10247
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770623467.852, "dur": 0.355,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770623468.743, "dur": 2.820,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770623472.275, "dur": 46.728,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770623519.703, "dur": 3.425,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770623523.640, "dur": 15.074,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770623543.165, "dur": 3.452,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770623547.337, "dur": 0.275,
    "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": 870, "tid": 870,
    "ts": 7429770623548.636, "dur": 1.125,
    "args": {
      "Python parent id": 3353, "Python id": 3425, "finished": true, "Ev Idx": 10255
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_143", "pid": 870, "tid": 870,
    "ts": 7429770623555.116, "dur": 50.096,
    "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": 870, "tid": 870,
    "ts": 7429770623556.444, "dur": 48.219,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770623557.013, "dur": 1.310,
    "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": 870, "tid": 870,
    "ts": 7429770623559.883, "dur": 44.560,
    "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": 870, "tid": 870,
    "ts": 7429770623560.283, "dur": 0.300,
    "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": 870, "tid": 870,
    "ts": 7429770623560.897, "dur": 0.176,
    "args": {
      "Python parent id": 3429, "Python id": 3431, "finished": true, "Ev Idx": 10261
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770623561.279, "dur": 43.074,
    "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": 870, "tid": 870,
    "ts": 7429770623620.009, "dur": 0.549,
    "args": {
      "Python parent id": 3337, "Python id": 3433, "finished": true, "Ev Idx": 10263
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_41", "pid": 870, "tid": 870,
    "ts": 7429770623623.308, "dur": 85.712,
    "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": 870, "tid": 870,
    "ts": 7429770623623.974, "dur": 84.684,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770623624.236, "dur": 0.366,
    "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": 870, "tid": 870,
    "ts": 7429770623625.671, "dur": 82.268,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770623626.339, "dur": 1.818,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770623628.461, "dur": 12.997,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770623641.819, "dur": 16.712,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770623671.187, "dur": 13.368,
    "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": 870, "tid": 870,
    "ts": 7429770623695.770, "dur": 0.249,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770623696.447, "dur": 1.163,
    "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": 870, "tid": 870,
    "ts": 7429770623709.825, "dur": 0.317,
    "args": {
      "Python parent id": 3337, "Python id": 3444, "finished": true, "Ev Idx": 10274
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_20", "pid": 870, "tid": 870,
    "ts": 7429770623712.177, "dur": 240.234,
    "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": 870, "tid": 870,
    "ts": 7429770623712.691, "dur": 239.540,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770623712.997, "dur": 0.299,
    "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": 870, "tid": 870,
    "ts": 7429770623714.382, "dur": 237.757,
    "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": 870, "tid": 870,
    "ts": 7429770623714.802, "dur": 0.397,
    "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": 870, "tid": 870,
    "ts": 7429770623715.630, "dur": 0.327,
    "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": 870, "tid": 870,
    "ts": 7429770623716.166, "dur": 0.208,
    "args": {
      "Python parent id": 3448, "Python id": 3451, "finished": true, "Ev Idx": 10281
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_144", "pid": 870, "tid": 870,
    "ts": 7429770623718.673, "dur": 43.354,
    "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": 870, "tid": 870,
    "ts": 7429770623719.363, "dur": 42.429,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770623719.486, "dur": 0.143,
    "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": 870, "tid": 870,
    "ts": 7429770623720.498, "dur": 41.193,
    "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": 870, "tid": 870,
    "ts": 7429770623720.726, "dur": 0.136,
    "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": 870, "tid": 870,
    "ts": 7429770623721.073, "dur": 0.146,
    "args": {
      "Python parent id": 3455, "Python id": 3457, "finished": true, "Ev Idx": 10287
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770623721.372, "dur": 40.217,
    "args": {
      "Python parent id": 3455, "Python id": 3458, "finished": true, "Ev Idx": 10288
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_20", "pid": 870, "tid": 870,
    "ts": 7429770623764.560, "dur": 17.849,
    "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": 870, "tid": 870,
    "ts": 7429770623765.448, "dur": 16.689,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770623765.675, "dur": 0.247,
    "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": 870, "tid": 870,
    "ts": 7429770623766.868, "dur": 15.165,
    "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": 870, "tid": 870,
    "ts": 7429770623767.844, "dur": 14.061,
    "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": 870, "tid": 870,
    "ts": 7429770623768.007, "dur": 0.170,
    "args": {
      "Python parent id": 3463, "Python id": 3464, "finished": true, "Ev Idx": 10294
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770623769.059, "dur": 12.756,
    "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": 870, "tid": 870,
    "ts": 7429770623783.017, "dur": 0.351,
    "args": {
      "Python parent id": 3448, "Python id": 3466, "finished": true, "Ev Idx": 10296
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_145", "pid": 870, "tid": 870,
    "ts": 7429770623785.180, "dur": 96.161,
    "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": 870, "tid": 870,
    "ts": 7429770623786.144, "dur": 94.887,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770623786.299, "dur": 0.196,
    "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": 870, "tid": 870,
    "ts": 7429770623787.320, "dur": 93.547,
    "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": 870, "tid": 870,
    "ts": 7429770623787.525, "dur": 0.446,
    "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": 870, "tid": 870,
    "ts": 7429770623788.403, "dur": 0.152,
    "args": {
      "Python parent id": 3470, "Python id": 3472, "finished": true, "Ev Idx": 10302
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770623788.658, "dur": 92.110,
    "args": {
      "Python parent id": 3470, "Python id": 3473, "finished": true, "Ev Idx": 10303
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_146", "pid": 870, "tid": 870,
    "ts": 7429770623895.207, "dur": 56.562,
    "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": 870, "tid": 870,
    "ts": 7429770623896.035, "dur": 55.434,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770623896.324, "dur": 0.507,
    "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": 870, "tid": 870,
    "ts": 7429770623897.547, "dur": 53.790,
    "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": 870, "tid": 870,
    "ts": 7429770623897.813, "dur": 0.160,
    "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": 870, "tid": 870,
    "ts": 7429770623898.204, "dur": 0.141,
    "args": {
      "Python parent id": 3477, "Python id": 3479, "finished": true, "Ev Idx": 10309
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770623898.470, "dur": 52.797,
    "args": {
      "Python parent id": 3477, "Python id": 3480, "finished": true, "Ev Idx": 10310
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_21", "pid": 870, "tid": 870,
    "ts": 7429770623973.116, "dur": 1150.515,
    "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": 870, "tid": 870,
    "ts": 7429770623974.418, "dur": 1148.302,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770623974.618, "dur": 0.290,
    "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": 870, "tid": 870,
    "ts": 7429770623976.293, "dur": 1145.634,
    "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": 870, "tid": 870,
    "ts": 7429770623976.611, "dur": 0.758,
    "args": {
      "Python parent id": 3484, "Python id": 3485, "finished": true, "Ev Idx": 10315
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_42", "pid": 870, "tid": 870,
    "ts": 7429770623980.245, "dur": 81.085,
    "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": 870, "tid": 870,
    "ts": 7429770623980.721, "dur": 80.350,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770623980.877, "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": 870, "tid": 870,
    "ts": 7429770623981.726, "dur": 78.624,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770623982.313, "dur": 1.766,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770623984.379, "dur": 12.557,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770623997.235, "dur": 14.912,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770624024.011, "dur": 10.898,
    "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": 870, "tid": 870,
    "ts": 7429770624047.553, "dur": 0.159,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770624048.265, "dur": 1.157,
    "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": 870, "tid": 870,
    "ts": 7429770624061.916, "dur": 0.335,
    "args": {
      "Python parent id": 3484, "Python id": 3496, "finished": true, "Ev Idx": 10326
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_21", "pid": 870, "tid": 870,
    "ts": 7429770624065.129, "dur": 764.289,
    "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": 870, "tid": 870,
    "ts": 7429770624066.645, "dur": 761.976,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770624066.872, "dur": 0.334,
    "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": 870, "tid": 870,
    "ts": 7429770624069.102, "dur": 758.687,
    "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": 870, "tid": 870,
    "ts": 7429770624072.257, "dur": 0.465,
    "args": {
      "Python parent id": 3500, "Python id": 3501, "finished": true, "Ev Idx": 10331
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_147", "pid": 870, "tid": 870,
    "ts": 7429770624074.945, "dur": 48.360,
    "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": 870, "tid": 870,
    "ts": 7429770624075.420, "dur": 47.594,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770624075.577, "dur": 0.165,
    "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": 870, "tid": 870,
    "ts": 7429770624076.706, "dur": 46.203,
    "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": 870, "tid": 870,
    "ts": 7429770624076.981, "dur": 0.120,
    "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": 870, "tid": 870,
    "ts": 7429770624077.322, "dur": 0.138,
    "args": {
      "Python parent id": 3505, "Python id": 3507, "finished": true, "Ev Idx": 10337
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770624077.591, "dur": 45.272,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770624124.123, "dur": 3.787,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770624128.299, "dur": 3.708,
    "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": 870, "tid": 870,
    "ts": 7429770624133.034, "dur": 0.332,
    "args": {
      "Python parent id": 3500, "Python id": 3511, "finished": true, "Ev Idx": 10341
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_148", "pid": 870, "tid": 870,
    "ts": 7429770624135.548, "dur": 41.133,
    "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": 870, "tid": 870,
    "ts": 7429770624136.026, "dur": 40.421,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770624136.281, "dur": 0.282,
    "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": 870, "tid": 870,
    "ts": 7429770624137.253, "dur": 39.083,
    "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": 870, "tid": 870,
    "ts": 7429770624137.607, "dur": 0.160,
    "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": 870, "tid": 870,
    "ts": 7429770624138.010, "dur": 0.182,
    "args": {
      "Python parent id": 3515, "Python id": 3517, "finished": true, "Ev Idx": 10347
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770624138.296, "dur": 37.993,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770624177.148, "dur": 2.591,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770624180.261, "dur": 2.585,
    "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": 870, "tid": 870,
    "ts": 7429770624183.514, "dur": 0.409,
    "args": {
      "Python parent id": 3500, "Python id": 3521, "finished": true, "Ev Idx": 10351
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_149", "pid": 870, "tid": 870,
    "ts": 7429770624185.923, "dur": 39.510,
    "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": 870, "tid": 870,
    "ts": 7429770624186.670, "dur": 38.517,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770624186.873, "dur": 0.246,
    "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": 870, "tid": 870,
    "ts": 7429770624187.901, "dur": 37.175,
    "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": 870, "tid": 870,
    "ts": 7429770624188.116, "dur": 0.149,
    "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": 870, "tid": 870,
    "ts": 7429770624188.452, "dur": 0.119,
    "args": {
      "Python parent id": 3525, "Python id": 3527, "finished": true, "Ev Idx": 10357
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770624188.675, "dur": 36.351,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770624226.061, "dur": 2.262,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770624228.940, "dur": 2.552,
    "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": 870, "tid": 870,
    "ts": 7429770624232.869, "dur": 151.486,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770624233.119, "dur": 3.991,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770624237.392, "dur": 1.657,
    "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": 870, "tid": 870,
    "ts": 7429770624254.291, "dur": 51.320,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770624282.199, "dur": 22.218,
    "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": 870, "tid": 870,
    "ts": 7429770624334.600, "dur": 31.693,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770624353.279, "dur": 12.084,
    "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": 870, "tid": 870,
    "ts": 7429770624388.794, "dur": 3.728,
    "args": {
      "Python parent id": 3500, "Python id": 3538, "finished": true, "Ev Idx": 10368
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770624389.528, "dur": 0.245,
    "args": {
      "Python parent id": 3538, "Python id": 3539, "finished": true, "Ev Idx": 10369
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770624390.170, "dur": 0.086,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770624391.592, "dur": 0.224,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770624392.183, "dur": 0.112,
    "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": 870, "tid": 870,
    "ts": 7429770624393.936, "dur": 6.763,
    "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": 870, "tid": 870,
    "ts": 7429770624397.153, "dur": 3.411,
    "args": {
      "Python parent id": 3543, "Python id": 3544, "finished": true, "Ev Idx": 10374
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770624397.451, "dur": 1.416,
    "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": 870, "tid": 870,
    "ts": 7429770624397.915, "dur": 0.784,
    "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": 870, "tid": 870,
    "ts": 7429770624399.226, "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": 870, "tid": 870,
    "ts": 7429770624400.083, "dur": 0.410,
    "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": 870, "tid": 870,
    "ts": 7429770624403.464, "dur": 363.830,
    "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": 870, "tid": 870,
    "ts": 7429770624404.086, "dur": 48.198,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770624423.560, "dur": 6.253,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770624430.705, "dur": 21.458,
    "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": 870, "tid": 870,
    "ts": 7429770624453.198, "dur": 66.423,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770624469.052, "dur": 2.865,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770624472.473, "dur": 47.067,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770624520.452, "dur": 2.949,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770624523.717, "dur": 104.552,
    "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": 870, "tid": 870,
    "ts": 7429770624668.119, "dur": 20.184,
    "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": 870, "tid": 870,
    "ts": 7429770624668.372, "dur": 0.224,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770624669.257, "dur": 18.900,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770624689.150, "dur": 1.827,
    "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": 870, "tid": 870,
    "ts": 7429770624692.468, "dur": 10.600,
    "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": 870, "tid": 870,
    "ts": 7429770624692.605, "dur": 0.105,
    "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": 870, "tid": 870,
    "ts": 7429770624698.577, "dur": 1.069,
    "args": {
      "Python parent id": 3562, "Python id": 3564, "finished": true, "Ev Idx": 10394
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770624699.114, "dur": 0.443,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770624700.079, "dur": 2.918,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770624703.676, "dur": 43.765,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770624748.062, "dur": 3.259,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770624751.832, "dur": 14.524,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770624770.401, "dur": 3.244,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770624774.376, "dur": 0.216,
    "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": 870, "tid": 870,
    "ts": 7429770624775.540, "dur": 1.296,
    "args": {
      "Python parent id": 3500, "Python id": 3572, "finished": true, "Ev Idx": 10402
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_150", "pid": 870, "tid": 870,
    "ts": 7429770624781.793, "dur": 45.414,
    "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": 870, "tid": 870,
    "ts": 7429770624783.076, "dur": 43.710,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770624783.685, "dur": 1.418,
    "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": 870, "tid": 870,
    "ts": 7429770624786.792, "dur": 39.755,
    "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": 870, "tid": 870,
    "ts": 7429770624787.175, "dur": 0.159,
    "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": 870, "tid": 870,
    "ts": 7429770624787.568, "dur": 0.147,
    "args": {
      "Python parent id": 3576, "Python id": 3578, "finished": true, "Ev Idx": 10408
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770624787.954, "dur": 38.525,
    "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": 870, "tid": 870,
    "ts": 7429770624843.228, "dur": 0.483,
    "args": {
      "Python parent id": 3484, "Python id": 3580, "finished": true, "Ev Idx": 10410
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_43", "pid": 870, "tid": 870,
    "ts": 7429770624846.478, "dur": 84.330,
    "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": 870, "tid": 870,
    "ts": 7429770624847.135, "dur": 83.407,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770624847.356, "dur": 0.383,
    "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": 870, "tid": 870,
    "ts": 7429770624849.551, "dur": 80.307,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770624850.395, "dur": 1.659,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770624852.389, "dur": 14.348,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770624867.062, "dur": 15.308,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770624895.481, "dur": 12.137,
    "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": 870, "tid": 870,
    "ts": 7429770624918.689, "dur": 0.179,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770624919.261, "dur": 1.043,
    "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": 870, "tid": 870,
    "ts": 7429770624931.503, "dur": 0.391,
    "args": {
      "Python parent id": 3484, "Python id": 3591, "finished": true, "Ev Idx": 10421
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_21", "pid": 870, "tid": 870,
    "ts": 7429770624933.945, "dur": 177.221,
    "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": 870, "tid": 870,
    "ts": 7429770624934.553, "dur": 176.405,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770624934.825, "dur": 0.301,
    "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": 870, "tid": 870,
    "ts": 7429770624936.046, "dur": 174.802,
    "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": 870, "tid": 870,
    "ts": 7429770624936.505, "dur": 2.193,
    "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": 870, "tid": 870,
    "ts": 7429770624939.152, "dur": 0.317,
    "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": 870, "tid": 870,
    "ts": 7429770624939.695, "dur": 0.274,
    "args": {
      "Python parent id": 3595, "Python id": 3598, "finished": true, "Ev Idx": 10428
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_151", "pid": 870, "tid": 870,
    "ts": 7429770624942.455, "dur": 46.799,
    "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": 870, "tid": 870,
    "ts": 7429770624943.308, "dur": 45.725,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770624943.470, "dur": 0.166,
    "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": 870, "tid": 870,
    "ts": 7429770624944.659, "dur": 44.261,
    "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": 870, "tid": 870,
    "ts": 7429770624944.908, "dur": 0.120,
    "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": 870, "tid": 870,
    "ts": 7429770624945.220, "dur": 0.148,
    "args": {
      "Python parent id": 3602, "Python id": 3604, "finished": true, "Ev Idx": 10434
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770624945.498, "dur": 43.371,
    "args": {
      "Python parent id": 3602, "Python id": 3605, "finished": true, "Ev Idx": 10435
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_21", "pid": 870, "tid": 870,
    "ts": 7429770624991.334, "dur": 17.146,
    "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": 870, "tid": 870,
    "ts": 7429770624991.816, "dur": 16.468,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770624991.981, "dur": 0.258,
    "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": 870, "tid": 870,
    "ts": 7429770624993.360, "dur": 14.808,
    "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": 870, "tid": 870,
    "ts": 7429770624994.354, "dur": 13.702,
    "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": 870, "tid": 870,
    "ts": 7429770624994.525, "dur": 0.133,
    "args": {
      "Python parent id": 3610, "Python id": 3611, "finished": true, "Ev Idx": 10441
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770624995.429, "dur": 12.528,
    "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": 870, "tid": 870,
    "ts": 7429770625009.157, "dur": 0.373,
    "args": {
      "Python parent id": 3595, "Python id": 3613, "finished": true, "Ev Idx": 10443
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_152", "pid": 870, "tid": 870,
    "ts": 7429770625011.612, "dur": 37.905,
    "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": 870, "tid": 870,
    "ts": 7429770625011.986, "dur": 37.301,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770625012.159, "dur": 0.186,
    "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": 870, "tid": 870,
    "ts": 7429770625013.077, "dur": 36.116,
    "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": 870, "tid": 870,
    "ts": 7429770625013.364, "dur": 0.132,
    "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": 870, "tid": 870,
    "ts": 7429770625013.725, "dur": 0.130,
    "args": {
      "Python parent id": 3617, "Python id": 3619, "finished": true, "Ev Idx": 10449
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770625013.960, "dur": 35.165,
    "args": {
      "Python parent id": 3617, "Python id": 3620, "finished": true, "Ev Idx": 10450
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_153", "pid": 870, "tid": 870,
    "ts": 7429770625060.566, "dur": 49.931,
    "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": 870, "tid": 870,
    "ts": 7429770625061.093, "dur": 49.066,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770625061.258, "dur": 0.219,
    "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": 870, "tid": 870,
    "ts": 7429770625062.536, "dur": 47.502,
    "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": 870, "tid": 870,
    "ts": 7429770625062.760, "dur": 0.128,
    "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": 870, "tid": 870,
    "ts": 7429770625063.065, "dur": 0.121,
    "args": {
      "Python parent id": 3624, "Python id": 3626, "finished": true, "Ev Idx": 10456
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770625063.287, "dur": 46.697,
    "args": {
      "Python parent id": 3624, "Python id": 3627, "finished": true, "Ev Idx": 10457
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_22", "pid": 870, "tid": 870,
    "ts": 7429770625129.427, "dur": 1150.020,
    "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": 870, "tid": 870,
    "ts": 7429770625130.825, "dur": 1147.710,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770625131.067, "dur": 0.224,
    "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": 870, "tid": 870,
    "ts": 7429770625132.645, "dur": 1145.050,
    "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": 870, "tid": 870,
    "ts": 7429770625132.968, "dur": 0.596,
    "args": {
      "Python parent id": 3631, "Python id": 3632, "finished": true, "Ev Idx": 10462
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_44", "pid": 870, "tid": 870,
    "ts": 7429770625135.454, "dur": 72.280,
    "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": 870, "tid": 870,
    "ts": 7429770625135.910, "dur": 71.529,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770625136.046, "dur": 0.152,
    "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": 870, "tid": 870,
    "ts": 7429770625136.928, "dur": 69.849,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770625137.376, "dur": 1.572,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770625139.237, "dur": 11.925,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770625151.608, "dur": 12.134,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770625174.642, "dur": 8.597,
    "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": 870, "tid": 870,
    "ts": 7429770625193.995, "dur": 0.149,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770625194.575, "dur": 1.059,
    "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": 870, "tid": 870,
    "ts": 7429770625208.361, "dur": 0.350,
    "args": {
      "Python parent id": 3631, "Python id": 3643, "finished": true, "Ev Idx": 10473
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_22", "pid": 870, "tid": 870,
    "ts": 7429770625211.304, "dur": 724.373,
    "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": 870, "tid": 870,
    "ts": 7429770625212.997, "dur": 721.955,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770625213.178, "dur": 0.273,
    "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": 870, "tid": 870,
    "ts": 7429770625215.293, "dur": 718.666,
    "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": 870, "tid": 870,
    "ts": 7429770625218.215, "dur": 0.340,
    "args": {
      "Python parent id": 3647, "Python id": 3648, "finished": true, "Ev Idx": 10478
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_154", "pid": 870, "tid": 870,
    "ts": 7429770625220.543, "dur": 45.152,
    "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": 870, "tid": 870,
    "ts": 7429770625221.098, "dur": 44.310,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770625221.256, "dur": 0.165,
    "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": 870, "tid": 870,
    "ts": 7429770625222.555, "dur": 42.711,
    "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": 870, "tid": 870,
    "ts": 7429770625222.887, "dur": 0.129,
    "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": 870, "tid": 870,
    "ts": 7429770625223.235, "dur": 0.160,
    "args": {
      "Python parent id": 3652, "Python id": 3654, "finished": true, "Ev Idx": 10484
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770625223.501, "dur": 41.704,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770625266.520, "dur": 3.119,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770625270.239, "dur": 3.039,
    "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": 870, "tid": 870,
    "ts": 7429770625274.276, "dur": 0.333,
    "args": {
      "Python parent id": 3647, "Python id": 3658, "finished": true, "Ev Idx": 10488
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_155", "pid": 870, "tid": 870,
    "ts": 7429770625276.498, "dur": 37.412,
    "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": 870, "tid": 870,
    "ts": 7429770625276.936, "dur": 36.692,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770625277.167, "dur": 0.239,
    "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": 870, "tid": 870,
    "ts": 7429770625278.250, "dur": 35.246,
    "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": 870, "tid": 870,
    "ts": 7429770625278.540, "dur": 0.142,
    "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": 870, "tid": 870,
    "ts": 7429770625279.032, "dur": 0.156,
    "args": {
      "Python parent id": 3662, "Python id": 3664, "finished": true, "Ev Idx": 10494
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770625279.296, "dur": 34.153,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770625314.401, "dur": 2.442,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770625317.360, "dur": 2.319,
    "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": 870, "tid": 870,
    "ts": 7429770625320.598, "dur": 0.311,
    "args": {
      "Python parent id": 3647, "Python id": 3668, "finished": true, "Ev Idx": 10498
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_156", "pid": 870, "tid": 870,
    "ts": 7429770625322.964, "dur": 35.759,
    "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": 870, "tid": 870,
    "ts": 7429770625323.481, "dur": 35.006,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770625323.676, "dur": 0.191,
    "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": 870, "tid": 870,
    "ts": 7429770625324.624, "dur": 33.740,
    "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": 870, "tid": 870,
    "ts": 7429770625324.879, "dur": 0.152,
    "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": 870, "tid": 870,
    "ts": 7429770625325.234, "dur": 0.133,
    "args": {
      "Python parent id": 3672, "Python id": 3674, "finished": true, "Ev Idx": 10504
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770625325.483, "dur": 32.821,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770625359.364, "dur": 2.213,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770625362.102, "dur": 2.709,
    "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": 870, "tid": 870,
    "ts": 7429770625365.843, "dur": 145.256,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770625366.080, "dur": 4.052,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770625370.417, "dur": 1.687,
    "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": 870, "tid": 870,
    "ts": 7429770625383.190, "dur": 46.917,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770625407.734, "dur": 21.304,
    "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": 870, "tid": 870,
    "ts": 7429770625461.158, "dur": 29.951,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770625478.678, "dur": 11.557,
    "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": 870, "tid": 870,
    "ts": 7429770625515.099, "dur": 3.722,
    "args": {
      "Python parent id": 3647, "Python id": 3685, "finished": true, "Ev Idx": 10515
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770625515.804, "dur": 0.255,
    "args": {
      "Python parent id": 3685, "Python id": 3686, "finished": true, "Ev Idx": 10516
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770625516.370, "dur": 0.079,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770625517.662, "dur": 0.230,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770625518.265, "dur": 0.302,
    "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": 870, "tid": 870,
    "ts": 7429770625520.225, "dur": 6.905,
    "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": 870, "tid": 870,
    "ts": 7429770625523.587, "dur": 3.390,
    "args": {
      "Python parent id": 3690, "Python id": 3691, "finished": true, "Ev Idx": 10521
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770625523.886, "dur": 1.268,
    "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": 870, "tid": 870,
    "ts": 7429770625524.305, "dur": 0.699,
    "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": 870, "tid": 870,
    "ts": 7429770625525.523, "dur": 0.595,
    "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": 870, "tid": 870,
    "ts": 7429770625526.454, "dur": 0.445,
    "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": 870, "tid": 870,
    "ts": 7429770625529.775, "dur": 338.706,
    "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": 870, "tid": 870,
    "ts": 7429770625530.260, "dur": 42.731,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770625548.542, "dur": 4.950,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770625554.289, "dur": 18.572,
    "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": 870, "tid": 870,
    "ts": 7429770625573.816, "dur": 31.914,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770625588.523, "dur": 2.831,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770625591.822, "dur": 13.842,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770625606.435, "dur": 2.619,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770625609.396, "dur": 115.368,
    "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": 870, "tid": 870,
    "ts": 7429770625765.786, "dur": 18.019,
    "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": 870, "tid": 870,
    "ts": 7429770625766.104, "dur": 0.228,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770625766.997, "dur": 16.665,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770625784.708, "dur": 2.002,
    "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": 870, "tid": 870,
    "ts": 7429770625788.261, "dur": 9.886,
    "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": 870, "tid": 870,
    "ts": 7429770625788.497, "dur": 0.109,
    "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": 870, "tid": 870,
    "ts": 7429770625794.109, "dur": 0.987,
    "args": {
      "Python parent id": 3709, "Python id": 3711, "finished": true, "Ev Idx": 10541
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770625794.640, "dur": 0.367,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770625795.406, "dur": 2.647,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770625798.804, "dur": 48.406,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770625847.856, "dur": 3.924,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770625852.335, "dur": 15.097,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770625872.092, "dur": 3.542,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770625876.376, "dur": 0.277,
    "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": 870, "tid": 870,
    "ts": 7429770625877.704, "dur": 1.192,
    "args": {
      "Python parent id": 3647, "Python id": 3719, "finished": true, "Ev Idx": 10549
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_157", "pid": 870, "tid": 870,
    "ts": 7429770625885.162, "dur": 48.133,
    "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": 870, "tid": 870,
    "ts": 7429770625886.665, "dur": 46.178,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770625887.189, "dur": 1.371,
    "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": 870, "tid": 870,
    "ts": 7429770625890.030, "dur": 42.505,
    "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": 870, "tid": 870,
    "ts": 7429770625890.364, "dur": 0.291,
    "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": 870, "tid": 870,
    "ts": 7429770625890.922, "dur": 0.162,
    "args": {
      "Python parent id": 3723, "Python id": 3725, "finished": true, "Ev Idx": 10555
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770625891.266, "dur": 41.160,
    "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": 870, "tid": 870,
    "ts": 7429770625948.499, "dur": 0.612,
    "args": {
      "Python parent id": 3631, "Python id": 3727, "finished": true, "Ev Idx": 10557
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_45", "pid": 870, "tid": 870,
    "ts": 7429770625952.292, "dur": 85.469,
    "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": 870, "tid": 870,
    "ts": 7429770625953.162, "dur": 84.220,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770625953.466, "dur": 0.409,
    "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": 870, "tid": 870,
    "ts": 7429770625955.040, "dur": 81.507,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770625955.705, "dur": 1.781,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770625957.873, "dur": 13.656,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770625971.879, "dur": 17.388,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770626002.521, "dur": 12.243,
    "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": 870, "tid": 870,
    "ts": 7429770626025.337, "dur": 0.300,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770626025.994, "dur": 1.125,
    "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": 870, "tid": 870,
    "ts": 7429770626038.682, "dur": 0.391,
    "args": {
      "Python parent id": 3631, "Python id": 3738, "finished": true, "Ev Idx": 10568
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_22", "pid": 870, "tid": 870,
    "ts": 7429770626041.055, "dur": 225.602,
    "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": 870, "tid": 870,
    "ts": 7429770626041.546, "dur": 224.878,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770626041.752, "dur": 0.302,
    "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": 870, "tid": 870,
    "ts": 7429770626042.870, "dur": 223.461,
    "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": 870, "tid": 870,
    "ts": 7429770626043.280, "dur": 0.369,
    "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": 870, "tid": 870,
    "ts": 7429770626044.062, "dur": 0.290,
    "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": 870, "tid": 870,
    "ts": 7429770626044.564, "dur": 0.459,
    "args": {
      "Python parent id": 3742, "Python id": 3745, "finished": true, "Ev Idx": 10575
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_158", "pid": 870, "tid": 870,
    "ts": 7429770626047.296, "dur": 80.001,
    "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": 870, "tid": 870,
    "ts": 7429770626047.896, "dur": 79.016,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770626048.028, "dur": 0.171,
    "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": 870, "tid": 870,
    "ts": 7429770626049.138, "dur": 77.504,
    "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": 870, "tid": 870,
    "ts": 7429770626049.406, "dur": 0.143,
    "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": 870, "tid": 870,
    "ts": 7429770626049.751, "dur": 0.164,
    "args": {
      "Python parent id": 3749, "Python id": 3751, "finished": true, "Ev Idx": 10581
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770626050.079, "dur": 76.456,
    "args": {
      "Python parent id": 3749, "Python id": 3752, "finished": true, "Ev Idx": 10582
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_22", "pid": 870, "tid": 870,
    "ts": 7429770626130.104, "dur": 21.973,
    "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": 870, "tid": 870,
    "ts": 7429770626131.021, "dur": 20.773,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770626131.284, "dur": 0.471,
    "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": 870, "tid": 870,
    "ts": 7429770626132.998, "dur": 18.627,
    "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": 870, "tid": 870,
    "ts": 7429770626134.033, "dur": 16.066,
    "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": 870, "tid": 870,
    "ts": 7429770626134.200, "dur": 0.221,
    "args": {
      "Python parent id": 3757, "Python id": 3758, "finished": true, "Ev Idx": 10588
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770626135.267, "dur": 14.749,
    "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": 870, "tid": 870,
    "ts": 7429770626152.884, "dur": 0.483,
    "args": {
      "Python parent id": 3742, "Python id": 3760, "finished": true, "Ev Idx": 10590
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_159", "pid": 870, "tid": 870,
    "ts": 7429770626155.708, "dur": 44.350,
    "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": 870, "tid": 870,
    "ts": 7429770626156.224, "dur": 43.493,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770626156.419, "dur": 0.186,
    "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": 870, "tid": 870,
    "ts": 7429770626157.494, "dur": 42.108,
    "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": 870, "tid": 870,
    "ts": 7429770626157.738, "dur": 0.366,
    "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": 870, "tid": 870,
    "ts": 7429770626158.290, "dur": 0.158,
    "args": {
      "Python parent id": 3764, "Python id": 3766, "finished": true, "Ev Idx": 10596
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770626158.577, "dur": 40.965,
    "args": {
      "Python parent id": 3764, "Python id": 3767, "finished": true, "Ev Idx": 10597
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_160", "pid": 870, "tid": 870,
    "ts": 7429770626212.819, "dur": 53.183,
    "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": 870, "tid": 870,
    "ts": 7429770626213.360, "dur": 52.379,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770626213.604, "dur": 0.254,
    "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": 870, "tid": 870,
    "ts": 7429770626214.742, "dur": 50.872,
    "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": 870, "tid": 870,
    "ts": 7429770626214.932, "dur": 0.154,
    "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": 870, "tid": 870,
    "ts": 7429770626215.309, "dur": 0.131,
    "args": {
      "Python parent id": 3771, "Python id": 3773, "finished": true, "Ev Idx": 10603
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770626215.544, "dur": 50.015,
    "args": {
      "Python parent id": 3771, "Python id": 3774, "finished": true, "Ev Idx": 10604
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_23", "pid": 870, "tid": 870,
    "ts": 7429770626285.966, "dur": 1199.543,
    "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": 870, "tid": 870,
    "ts": 7429770626287.342, "dur": 1197.290,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770626287.642, "dur": 0.615,
    "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": 870, "tid": 870,
    "ts": 7429770626289.744, "dur": 1194.128,
    "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": 870, "tid": 870,
    "ts": 7429770626290.181, "dur": 0.657,
    "args": {
      "Python parent id": 3778, "Python id": 3779, "finished": true, "Ev Idx": 10609
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_46", "pid": 870, "tid": 870,
    "ts": 7429770626293.478, "dur": 80.017,
    "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": 870, "tid": 870,
    "ts": 7429770626294.059, "dur": 79.130,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770626294.234, "dur": 0.177,
    "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": 870, "tid": 870,
    "ts": 7429770626295.343, "dur": 77.077,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770626295.886, "dur": 1.905,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770626298.140, "dur": 13.341,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770626311.952, "dur": 15.415,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770626339.138, "dur": 11.334,
    "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": 870, "tid": 870,
    "ts": 7429770626360.504, "dur": 0.197,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770626361.196, "dur": 1.065,
    "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": 870, "tid": 870,
    "ts": 7429770626374.275, "dur": 0.311,
    "args": {
      "Python parent id": 3778, "Python id": 3790, "finished": true, "Ev Idx": 10620
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_23", "pid": 870, "tid": 870,
    "ts": 7429770626377.373, "dur": 778.758,
    "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": 870, "tid": 870,
    "ts": 7429770626378.998, "dur": 776.257,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770626379.254, "dur": 0.298,
    "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": 870, "tid": 870,
    "ts": 7429770626381.829, "dur": 772.366,
    "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": 870, "tid": 870,
    "ts": 7429770626385.061, "dur": 0.319,
    "args": {
      "Python parent id": 3794, "Python id": 3795, "finished": true, "Ev Idx": 10625
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_161", "pid": 870, "tid": 870,
    "ts": 7429770626387.378, "dur": 43.773,
    "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": 870, "tid": 870,
    "ts": 7429770626387.904, "dur": 43.019,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770626388.043, "dur": 0.164,
    "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": 870, "tid": 870,
    "ts": 7429770626388.922, "dur": 41.893,
    "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": 870, "tid": 870,
    "ts": 7429770626389.286, "dur": 0.131,
    "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": 870, "tid": 870,
    "ts": 7429770626389.618, "dur": 0.145,
    "args": {
      "Python parent id": 3799, "Python id": 3801, "finished": true, "Ev Idx": 10631
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770626389.875, "dur": 40.871,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770626432.004, "dur": 3.541,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770626436.143, "dur": 3.759,
    "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": 870, "tid": 870,
    "ts": 7429770626441.035, "dur": 0.327,
    "args": {
      "Python parent id": 3794, "Python id": 3805, "finished": true, "Ev Idx": 10635
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_162", "pid": 870, "tid": 870,
    "ts": 7429770626443.914, "dur": 36.528,
    "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": 870, "tid": 870,
    "ts": 7429770626444.391, "dur": 35.807,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770626444.623, "dur": 0.242,
    "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": 870, "tid": 870,
    "ts": 7429770626445.638, "dur": 34.421,
    "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": 870, "tid": 870,
    "ts": 7429770626446.000, "dur": 0.125,
    "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": 870, "tid": 870,
    "ts": 7429770626446.354, "dur": 0.136,
    "args": {
      "Python parent id": 3809, "Python id": 3811, "finished": true, "Ev Idx": 10641
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770626446.598, "dur": 33.414,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770626480.940, "dur": 2.463,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770626483.913, "dur": 2.717,
    "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": 870, "tid": 870,
    "ts": 7429770626487.487, "dur": 0.338,
    "args": {
      "Python parent id": 3794, "Python id": 3815, "finished": true, "Ev Idx": 10645
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_163", "pid": 870, "tid": 870,
    "ts": 7429770626490.231, "dur": 36.523,
    "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": 870, "tid": 870,
    "ts": 7429770626490.669, "dur": 35.764,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770626490.957, "dur": 0.238,
    "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": 870, "tid": 870,
    "ts": 7429770626491.751, "dur": 34.586,
    "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": 870, "tid": 870,
    "ts": 7429770626491.972, "dur": 0.146,
    "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": 870, "tid": 870,
    "ts": 7429770626492.309, "dur": 0.138,
    "args": {
      "Python parent id": 3819, "Python id": 3821, "finished": true, "Ev Idx": 10651
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770626492.545, "dur": 33.752,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770626527.348, "dur": 2.710,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770626530.474, "dur": 2.402,
    "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": 870, "tid": 870,
    "ts": 7429770626534.070, "dur": 151.634,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770626534.355, "dur": 3.649,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770626538.303, "dur": 1.638,
    "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": 870, "tid": 870,
    "ts": 7429770626551.244, "dur": 52.273,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770626579.559, "dur": 22.824,
    "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": 870, "tid": 870,
    "ts": 7429770626636.350, "dur": 28.661,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770626653.091, "dur": 10.952,
    "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": 870, "tid": 870,
    "ts": 7429770626689.983, "dur": 3.800,
    "args": {
      "Python parent id": 3794, "Python id": 3832, "finished": true, "Ev Idx": 10662
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770626690.715, "dur": 0.245,
    "args": {
      "Python parent id": 3832, "Python id": 3833, "finished": true, "Ev Idx": 10663
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770626691.319, "dur": 0.083,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770626692.764, "dur": 0.221,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770626693.394, "dur": 0.114,
    "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": 870, "tid": 870,
    "ts": 7429770626695.172, "dur": 6.808,
    "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": 870, "tid": 870,
    "ts": 7429770626698.664, "dur": 3.198,
    "args": {
      "Python parent id": 3837, "Python id": 3838, "finished": true, "Ev Idx": 10668
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770626698.967, "dur": 1.377,
    "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": 870, "tid": 870,
    "ts": 7429770626699.477, "dur": 0.705,
    "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": 870, "tid": 870,
    "ts": 7429770626700.629, "dur": 0.477,
    "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": 870, "tid": 870,
    "ts": 7429770626701.408, "dur": 0.379,
    "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": 870, "tid": 870,
    "ts": 7429770626704.723, "dur": 383.392,
    "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": 870, "tid": 870,
    "ts": 7429770626705.288, "dur": 47.299,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770626725.115, "dur": 5.041,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770626731.008, "dur": 21.436,
    "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": 870, "tid": 870,
    "ts": 7429770626753.481, "dur": 33.618,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770626767.480, "dur": 4.485,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770626772.443, "dur": 14.580,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770626787.910, "dur": 2.989,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770626791.194, "dur": 154.637,
    "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": 870, "tid": 870,
    "ts": 7429770626985.801, "dur": 19.131,
    "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": 870, "tid": 870,
    "ts": 7429770626986.022, "dur": 0.170,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770626986.951, "dur": 17.844,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770627005.705, "dur": 1.815,
    "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": 870, "tid": 870,
    "ts": 7429770627009.098, "dur": 10.762,
    "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": 870, "tid": 870,
    "ts": 7429770627009.261, "dur": 0.090,
    "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": 870, "tid": 870,
    "ts": 7429770627015.352, "dur": 1.018,
    "args": {
      "Python parent id": 3856, "Python id": 3858, "finished": true, "Ev Idx": 10688
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770627015.835, "dur": 0.456,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770627016.775, "dur": 3.010,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770627020.467, "dur": 46.153,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770627067.296, "dur": 3.536,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770627071.369, "dur": 15.764,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770627091.296, "dur": 4.028,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770627096.073, "dur": 0.289,
    "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": 870, "tid": 870,
    "ts": 7429770627097.635, "dur": 0.967,
    "args": {
      "Python parent id": 3794, "Python id": 3866, "finished": true, "Ev Idx": 10696
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_164", "pid": 870, "tid": 870,
    "ts": 7429770627103.455, "dur": 50.039,
    "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": 870, "tid": 870,
    "ts": 7429770627104.725, "dur": 48.355,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770627105.318, "dur": 1.326,
    "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": 870, "tid": 870,
    "ts": 7429770627108.237, "dur": 44.590,
    "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": 870, "tid": 870,
    "ts": 7429770627108.670, "dur": 0.196,
    "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": 870, "tid": 870,
    "ts": 7429770627109.142, "dur": 0.178,
    "args": {
      "Python parent id": 3870, "Python id": 3872, "finished": true, "Ev Idx": 10702
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770627109.491, "dur": 43.258,
    "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": 870, "tid": 870,
    "ts": 7429770627170.125, "dur": 0.649,
    "args": {
      "Python parent id": 3778, "Python id": 3874, "finished": true, "Ev Idx": 10704
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_47", "pid": 870, "tid": 870,
    "ts": 7429770627173.294, "dur": 85.429,
    "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": 870, "tid": 870,
    "ts": 7429770627174.015, "dur": 84.438,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770627174.226, "dur": 0.420,
    "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": 870, "tid": 870,
    "ts": 7429770627175.737, "dur": 81.847,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770627176.321, "dur": 1.773,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770627178.413, "dur": 14.422,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770627193.144, "dur": 17.143,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770627222.121, "dur": 12.366,
    "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": 870, "tid": 870,
    "ts": 7429770627245.464, "dur": 0.204,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770627246.031, "dur": 1.252,
    "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": 870, "tid": 870,
    "ts": 7429770627259.472, "dur": 0.516,
    "args": {
      "Python parent id": 3778, "Python id": 3885, "finished": true, "Ev Idx": 10715
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_23", "pid": 870, "tid": 870,
    "ts": 7429770627262.115, "dur": 210.657,
    "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": 870, "tid": 870,
    "ts": 7429770627262.566, "dur": 210.024,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770627262.878, "dur": 0.315,
    "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": 870, "tid": 870,
    "ts": 7429770627264.165, "dur": 208.306,
    "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": 870, "tid": 870,
    "ts": 7429770627264.665, "dur": 0.320,
    "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": 870, "tid": 870,
    "ts": 7429770627265.432, "dur": 0.286,
    "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": 870, "tid": 870,
    "ts": 7429770627265.970, "dur": 0.212,
    "args": {
      "Python parent id": 3889, "Python id": 3892, "finished": true, "Ev Idx": 10722
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_165", "pid": 870, "tid": 870,
    "ts": 7429770627268.071, "dur": 45.772,
    "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": 870, "tid": 870,
    "ts": 7429770627268.883, "dur": 44.727,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770627269.020, "dur": 0.192,
    "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": 870, "tid": 870,
    "ts": 7429770627270.301, "dur": 43.159,
    "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": 870, "tid": 870,
    "ts": 7429770627270.658, "dur": 0.270,
    "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": 870, "tid": 870,
    "ts": 7429770627271.152, "dur": 0.185,
    "args": {
      "Python parent id": 3896, "Python id": 3898, "finished": true, "Ev Idx": 10728
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770627271.456, "dur": 41.958,
    "args": {
      "Python parent id": 3896, "Python id": 3899, "finished": true, "Ev Idx": 10729
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_23", "pid": 870, "tid": 870,
    "ts": 7429770627315.419, "dur": 17.235,
    "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": 870, "tid": 870,
    "ts": 7429770627315.990, "dur": 16.424,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770627316.242, "dur": 0.304,
    "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": 870, "tid": 870,
    "ts": 7429770627317.826, "dur": 14.497,
    "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": 870, "tid": 870,
    "ts": 7429770627318.677, "dur": 13.481,
    "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": 870, "tid": 870,
    "ts": 7429770627318.823, "dur": 0.147,
    "args": {
      "Python parent id": 3904, "Python id": 3905, "finished": true, "Ev Idx": 10735
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770627319.845, "dur": 12.209,
    "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": 870, "tid": 870,
    "ts": 7429770627333.261, "dur": 0.400,
    "args": {
      "Python parent id": 3889, "Python id": 3907, "finished": true, "Ev Idx": 10737
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_166", "pid": 870, "tid": 870,
    "ts": 7429770627335.520, "dur": 76.667,
    "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": 870, "tid": 870,
    "ts": 7429770627336.060, "dur": 75.860,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770627336.385, "dur": 0.190,
    "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": 870, "tid": 870,
    "ts": 7429770627337.168, "dur": 74.653,
    "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": 870, "tid": 870,
    "ts": 7429770627337.442, "dur": 0.118,
    "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": 870, "tid": 870,
    "ts": 7429770627337.759, "dur": 0.281,
    "args": {
      "Python parent id": 3911, "Python id": 3913, "finished": true, "Ev Idx": 10743
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770627338.297, "dur": 73.468,
    "args": {
      "Python parent id": 3911, "Python id": 3914, "finished": true, "Ev Idx": 10744
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_167", "pid": 870, "tid": 870,
    "ts": 7429770627424.847, "dur": 47.153,
    "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": 870, "tid": 870,
    "ts": 7429770627425.378, "dur": 46.241,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770627425.590, "dur": 0.361,
    "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": 870, "tid": 870,
    "ts": 7429770627426.775, "dur": 44.708,
    "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": 870, "tid": 870,
    "ts": 7429770627427.061, "dur": 0.167,
    "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": 870, "tid": 870,
    "ts": 7429770627427.468, "dur": 0.167,
    "args": {
      "Python parent id": 3918, "Python id": 3920, "finished": true, "Ev Idx": 10750
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770627427.754, "dur": 43.655,
    "args": {
      "Python parent id": 3918, "Python id": 3921, "finished": true, "Ev Idx": 10751
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_24", "pid": 870, "tid": 870,
    "ts": 7429770627491.197, "dur": 1100.298,
    "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": 870, "tid": 870,
    "ts": 7429770627492.464, "dur": 1098.232,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770627492.678, "dur": 0.272,
    "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": 870, "tid": 870,
    "ts": 7429770627494.267, "dur": 1095.459,
    "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": 870, "tid": 870,
    "ts": 7429770627494.620, "dur": 0.489,
    "args": {
      "Python parent id": 3925, "Python id": 3926, "finished": true, "Ev Idx": 10756
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_48", "pid": 870, "tid": 870,
    "ts": 7429770627496.895, "dur": 73.872,
    "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": 870, "tid": 870,
    "ts": 7429770627497.375, "dur": 73.133,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770627497.531, "dur": 0.166,
    "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": 870, "tid": 870,
    "ts": 7429770627498.468, "dur": 71.228,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770627498.981, "dur": 1.528,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770627500.795, "dur": 11.631,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770627512.848, "dur": 14.582,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770627538.175, "dur": 10.362,
    "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": 870, "tid": 870,
    "ts": 7429770627557.851, "dur": 0.214,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770627558.700, "dur": 0.986,
    "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": 870, "tid": 870,
    "ts": 7429770627571.488, "dur": 0.328,
    "args": {
      "Python parent id": 3925, "Python id": 3937, "finished": true, "Ev Idx": 10767
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_24", "pid": 870, "tid": 870,
    "ts": 7429770627574.396, "dur": 719.129,
    "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": 870, "tid": 870,
    "ts": 7429770627575.861, "dur": 716.878,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770627576.115, "dur": 0.343,
    "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": 870, "tid": 870,
    "ts": 7429770627578.449, "dur": 713.127,
    "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": 870, "tid": 870,
    "ts": 7429770627581.431, "dur": 0.347,
    "args": {
      "Python parent id": 3941, "Python id": 3942, "finished": true, "Ev Idx": 10772
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_168", "pid": 870, "tid": 870,
    "ts": 7429770627583.754, "dur": 41.338,
    "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": 870, "tid": 870,
    "ts": 7429770627584.261, "dur": 40.624,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770627584.416, "dur": 0.174,
    "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": 870, "tid": 870,
    "ts": 7429770627585.541, "dur": 39.152,
    "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": 870, "tid": 870,
    "ts": 7429770627585.810, "dur": 0.136,
    "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": 870, "tid": 870,
    "ts": 7429770627586.148, "dur": 0.150,
    "args": {
      "Python parent id": 3946, "Python id": 3948, "finished": true, "Ev Idx": 10778
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770627586.407, "dur": 38.223,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770627625.884, "dur": 3.292,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770627629.574, "dur": 3.279,
    "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": 870, "tid": 870,
    "ts": 7429770627633.715, "dur": 0.345,
    "args": {
      "Python parent id": 3941, "Python id": 3952, "finished": true, "Ev Idx": 10782
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_169", "pid": 870, "tid": 870,
    "ts": 7429770627636.567, "dur": 35.746,
    "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": 870, "tid": 870,
    "ts": 7429770627637.130, "dur": 34.924,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770627637.377, "dur": 0.240,
    "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": 870, "tid": 870,
    "ts": 7429770627638.601, "dur": 33.334,
    "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": 870, "tid": 870,
    "ts": 7429770627638.858, "dur": 0.161,
    "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": 870, "tid": 870,
    "ts": 7429770627639.248, "dur": 0.136,
    "args": {
      "Python parent id": 3956, "Python id": 3958, "finished": true, "Ev Idx": 10788
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770627639.486, "dur": 32.409,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770627672.911, "dur": 2.297,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770627675.733, "dur": 2.735,
    "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": 870, "tid": 870,
    "ts": 7429770627679.199, "dur": 0.326,
    "args": {
      "Python parent id": 3941, "Python id": 3962, "finished": true, "Ev Idx": 10792
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_170", "pid": 870, "tid": 870,
    "ts": 7429770627681.452, "dur": 34.296,
    "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": 870, "tid": 870,
    "ts": 7429770627681.901, "dur": 33.625,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770627682.116, "dur": 0.329,
    "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": 870, "tid": 870,
    "ts": 7429770627683.275, "dur": 32.132,
    "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": 870, "tid": 870,
    "ts": 7429770627683.511, "dur": 0.126,
    "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": 870, "tid": 870,
    "ts": 7429770627683.830, "dur": 0.160,
    "args": {
      "Python parent id": 3966, "Python id": 3968, "finished": true, "Ev Idx": 10798
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770627684.098, "dur": 31.248,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770627716.337, "dur": 2.549,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770627719.332, "dur": 2.374,
    "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": 870, "tid": 870,
    "ts": 7429770627722.850, "dur": 147.684,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770627723.180, "dur": 3.603,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770627727.071, "dur": 1.701,
    "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": 870, "tid": 870,
    "ts": 7429770627740.587, "dur": 49.881,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770627769.216, "dur": 20.096,
    "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": 870, "tid": 870,
    "ts": 7429770627821.334, "dur": 28.736,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770627838.056, "dur": 11.153,
    "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": 870, "tid": 870,
    "ts": 7429770627874.783, "dur": 4.151,
    "args": {
      "Python parent id": 3941, "Python id": 3979, "finished": true, "Ev Idx": 10809
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770627875.582, "dur": 0.237,
    "args": {
      "Python parent id": 3979, "Python id": 3980, "finished": true, "Ev Idx": 10810
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770627876.155, "dur": 0.091,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770627877.618, "dur": 0.447,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770627878.526, "dur": 0.145,
    "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": 870, "tid": 870,
    "ts": 7429770627880.186, "dur": 6.767,
    "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": 870, "tid": 870,
    "ts": 7429770627883.272, "dur": 3.570,
    "args": {
      "Python parent id": 3984, "Python id": 3985, "finished": true, "Ev Idx": 10815
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770627883.598, "dur": 1.331,
    "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": 870, "tid": 870,
    "ts": 7429770627884.090, "dur": 0.670,
    "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": 870, "tid": 870,
    "ts": 7429770627885.264, "dur": 0.486,
    "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": 870, "tid": 870,
    "ts": 7429770627886.104, "dur": 0.664,
    "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": 870, "tid": 870,
    "ts": 7429770627889.597, "dur": 337.853,
    "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": 870, "tid": 870,
    "ts": 7429770627890.111, "dur": 42.638,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770627908.374, "dur": 4.673,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770627913.907, "dur": 18.707,
    "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": 870, "tid": 870,
    "ts": 7429770627933.733, "dur": 31.933,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770627947.653, "dur": 2.589,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770627950.746, "dur": 14.855,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770627966.387, "dur": 2.646,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770627969.334, "dur": 114.267,
    "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": 870, "tid": 870,
    "ts": 7429770628125.710, "dur": 18.836,
    "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": 870, "tid": 870,
    "ts": 7429770628126.073, "dur": 0.165,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770628126.943, "dur": 17.314,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770628145.414, "dur": 1.882,
    "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": 870, "tid": 870,
    "ts": 7429770628148.786, "dur": 10.079,
    "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": 870, "tid": 870,
    "ts": 7429770628148.918, "dur": 0.124,
    "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": 870, "tid": 870,
    "ts": 7429770628154.749, "dur": 0.925,
    "args": {
      "Python parent id": 4003, "Python id": 4005, "finished": true, "Ev Idx": 10835
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770628155.231, "dur": 0.349,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770628156.056, "dur": 2.723,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770628159.571, "dur": 46.658,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770628206.896, "dur": 3.784,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770628211.225, "dur": 15.148,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770628231.010, "dur": 3.560,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770628235.369, "dur": 0.242,
    "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": 870, "tid": 870,
    "ts": 7429770628236.613, "dur": 1.094,
    "args": {
      "Python parent id": 3941, "Python id": 4013, "finished": true, "Ev Idx": 10843
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_171", "pid": 870, "tid": 870,
    "ts": 7429770628243.213, "dur": 47.669,
    "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": 870, "tid": 870,
    "ts": 7429770628244.725, "dur": 45.698,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770628245.356, "dur": 1.412,
    "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": 870, "tid": 870,
    "ts": 7429770628248.403, "dur": 41.737,
    "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": 870, "tid": 870,
    "ts": 7429770628248.833, "dur": 0.284,
    "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": 870, "tid": 870,
    "ts": 7429770628249.391, "dur": 0.137,
    "args": {
      "Python parent id": 4017, "Python id": 4019, "finished": true, "Ev Idx": 10849
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770628249.954, "dur": 40.078,
    "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": 870, "tid": 870,
    "ts": 7429770628306.494, "dur": 0.499,
    "args": {
      "Python parent id": 3925, "Python id": 4021, "finished": true, "Ev Idx": 10851
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_49", "pid": 870, "tid": 870,
    "ts": 7429770628310.293, "dur": 91.320,
    "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": 870, "tid": 870,
    "ts": 7429770628311.119, "dur": 90.116,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770628311.401, "dur": 0.463,
    "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": 870, "tid": 870,
    "ts": 7429770628313.326, "dur": 87.096,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770628314.079, "dur": 1.597,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770628316.046, "dur": 14.802,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770628331.197, "dur": 18.796,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770628364.162, "dur": 11.442,
    "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": 870, "tid": 870,
    "ts": 7429770628388.899, "dur": 0.183,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770628389.553, "dur": 1.200,
    "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": 870, "tid": 870,
    "ts": 7429770628402.491, "dur": 0.355,
    "args": {
      "Python parent id": 3925, "Python id": 4032, "finished": true, "Ev Idx": 10862
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_24", "pid": 870, "tid": 870,
    "ts": 7429770628405.226, "dur": 173.295,
    "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": 870, "tid": 870,
    "ts": 7429770628405.734, "dur": 172.599,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770628405.963, "dur": 0.375,
    "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": 870, "tid": 870,
    "ts": 7429770628407.374, "dur": 170.853,
    "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": 870, "tid": 870,
    "ts": 7429770628407.802, "dur": 0.351,
    "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": 870, "tid": 870,
    "ts": 7429770628408.504, "dur": 0.310,
    "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": 870, "tid": 870,
    "ts": 7429770628409.085, "dur": 0.222,
    "args": {
      "Python parent id": 4036, "Python id": 4039, "finished": true, "Ev Idx": 10869
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_172", "pid": 870, "tid": 870,
    "ts": 7429770628412.065, "dur": 43.189,
    "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": 870, "tid": 870,
    "ts": 7429770628412.683, "dur": 42.338,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770628412.839, "dur": 0.157,
    "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": 870, "tid": 870,
    "ts": 7429770628413.742, "dur": 41.147,
    "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": 870, "tid": 870,
    "ts": 7429770628414.031, "dur": 0.117,
    "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": 870, "tid": 870,
    "ts": 7429770628414.531, "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": 870, "tid": 870,
    "ts": 7429770628414.981, "dur": 39.833,
    "args": {
      "Python parent id": 4043, "Python id": 4046, "finished": true, "Ev Idx": 10876
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_24", "pid": 870, "tid": 870,
    "ts": 7429770628457.126, "dur": 18.399,
    "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": 870, "tid": 870,
    "ts": 7429770628457.677, "dur": 17.654,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770628457.867, "dur": 0.336,
    "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": 870, "tid": 870,
    "ts": 7429770628459.422, "dur": 15.817,
    "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": 870, "tid": 870,
    "ts": 7429770628460.402, "dur": 14.701,
    "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": 870, "tid": 870,
    "ts": 7429770628460.587, "dur": 0.185,
    "args": {
      "Python parent id": 4051, "Python id": 4052, "finished": true, "Ev Idx": 10882
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770628461.566, "dur": 13.449,
    "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": 870, "tid": 870,
    "ts": 7429770628476.131, "dur": 0.386,
    "args": {
      "Python parent id": 4036, "Python id": 4054, "finished": true, "Ev Idx": 10884
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_173", "pid": 870, "tid": 870,
    "ts": 7429770628478.591, "dur": 39.380,
    "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": 870, "tid": 870,
    "ts": 7429770628479.003, "dur": 38.717,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770628479.186, "dur": 0.200,
    "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": 870, "tid": 870,
    "ts": 7429770628480.287, "dur": 37.321,
    "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": 870, "tid": 870,
    "ts": 7429770628480.473, "dur": 0.119,
    "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": 870, "tid": 870,
    "ts": 7429770628480.815, "dur": 0.116,
    "args": {
      "Python parent id": 4058, "Python id": 4060, "finished": true, "Ev Idx": 10890
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770628481.058, "dur": 36.504,
    "args": {
      "Python parent id": 4058, "Python id": 4061, "finished": true, "Ev Idx": 10891
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_174", "pid": 870, "tid": 870,
    "ts": 7429770628529.747, "dur": 48.134,
    "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": 870, "tid": 870,
    "ts": 7429770628530.344, "dur": 47.210,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770628530.692, "dur": 0.233,
    "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": 870, "tid": 870,
    "ts": 7429770628531.720, "dur": 45.705,
    "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": 870, "tid": 870,
    "ts": 7429770628532.106, "dur": 0.121,
    "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": 870, "tid": 870,
    "ts": 7429770628532.430, "dur": 0.114,
    "args": {
      "Python parent id": 4065, "Python id": 4067, "finished": true, "Ev Idx": 10897
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770628532.831, "dur": 44.544,
    "args": {
      "Python parent id": 4065, "Python id": 4068, "finished": true, "Ev Idx": 10898
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_25", "pid": 870, "tid": 870,
    "ts": 7429770628597.625, "dur": 1141.679,
    "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": 870, "tid": 870,
    "ts": 7429770628599.063, "dur": 1139.435,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770628599.330, "dur": 0.247,
    "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": 870, "tid": 870,
    "ts": 7429770628601.171, "dur": 1136.299,
    "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": 870, "tid": 870,
    "ts": 7429770628601.551, "dur": 0.561,
    "args": {
      "Python parent id": 4072, "Python id": 4073, "finished": true, "Ev Idx": 10903
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_50", "pid": 870, "tid": 870,
    "ts": 7429770628604.210, "dur": 71.671,
    "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": 870, "tid": 870,
    "ts": 7429770628604.770, "dur": 70.832,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770628604.930, "dur": 0.153,
    "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": 870, "tid": 870,
    "ts": 7429770628605.922, "dur": 68.972,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770628606.428, "dur": 1.723,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770628608.420, "dur": 11.926,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770628620.693, "dur": 12.526,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770628644.803, "dur": 8.458,
    "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": 870, "tid": 870,
    "ts": 7429770628663.939, "dur": 0.162,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770628664.525, "dur": 1.216,
    "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": 870, "tid": 870,
    "ts": 7429770628676.501, "dur": 0.393,
    "args": {
      "Python parent id": 4072, "Python id": 4084, "finished": true, "Ev Idx": 10914
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_25", "pid": 870, "tid": 870,
    "ts": 7429770628679.709, "dur": 717.304,
    "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": 870, "tid": 870,
    "ts": 7429770628681.183, "dur": 714.920,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770628681.411, "dur": 0.376,
    "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": 870, "tid": 870,
    "ts": 7429770628683.770, "dur": 711.480,
    "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": 870, "tid": 870,
    "ts": 7429770628686.704, "dur": 0.390,
    "args": {
      "Python parent id": 4088, "Python id": 4089, "finished": true, "Ev Idx": 10919
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_175", "pid": 870, "tid": 870,
    "ts": 7429770628689.221, "dur": 41.943,
    "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": 870, "tid": 870,
    "ts": 7429770628689.643, "dur": 41.257,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770628689.807, "dur": 0.152,
    "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": 870, "tid": 870,
    "ts": 7429770628690.911, "dur": 39.800,
    "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": 870, "tid": 870,
    "ts": 7429770628691.174, "dur": 0.114,
    "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": 870, "tid": 870,
    "ts": 7429770628691.494, "dur": 0.137,
    "args": {
      "Python parent id": 4093, "Python id": 4095, "finished": true, "Ev Idx": 10925
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770628691.744, "dur": 38.908,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770628731.901, "dur": 3.320,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770628735.669, "dur": 3.372,
    "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": 870, "tid": 870,
    "ts": 7429770628739.906, "dur": 0.329,
    "args": {
      "Python parent id": 4088, "Python id": 4099, "finished": true, "Ev Idx": 10929
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_176", "pid": 870, "tid": 870,
    "ts": 7429770628742.239, "dur": 37.586,
    "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": 870, "tid": 870,
    "ts": 7429770628742.742, "dur": 36.822,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770628742.934, "dur": 0.250,
    "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": 870, "tid": 870,
    "ts": 7429770628743.841, "dur": 35.627,
    "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": 870, "tid": 870,
    "ts": 7429770628744.075, "dur": 0.166,
    "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": 870, "tid": 870,
    "ts": 7429770628744.449, "dur": 0.132,
    "args": {
      "Python parent id": 4103, "Python id": 4105, "finished": true, "Ev Idx": 10935
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770628744.689, "dur": 34.731,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770628780.297, "dur": 2.569,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770628783.215, "dur": 3.874,
    "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": 870, "tid": 870,
    "ts": 7429770628787.951, "dur": 0.317,
    "args": {
      "Python parent id": 4088, "Python id": 4109, "finished": true, "Ev Idx": 10939
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_177", "pid": 870, "tid": 870,
    "ts": 7429770628790.273, "dur": 37.804,
    "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": 870, "tid": 870,
    "ts": 7429770628790.791, "dur": 37.053,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770628790.979, "dur": 0.297,
    "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": 870, "tid": 870,
    "ts": 7429770628791.899, "dur": 35.792,
    "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": 870, "tid": 870,
    "ts": 7429770628792.104, "dur": 0.122,
    "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": 870, "tid": 870,
    "ts": 7429770628792.446, "dur": 0.116,
    "args": {
      "Python parent id": 4113, "Python id": 4115, "finished": true, "Ev Idx": 10945
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770628792.666, "dur": 34.976,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770628828.602, "dur": 2.318,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770628831.413, "dur": 2.465,
    "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": 870, "tid": 870,
    "ts": 7429770628834.884, "dur": 154.385,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770628835.178, "dur": 3.298,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770628838.778, "dur": 1.640,
    "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": 870, "tid": 870,
    "ts": 7429770628851.601, "dur": 49.336,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770628880.132, "dur": 19.454,
    "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": 870, "tid": 870,
    "ts": 7429770628933.906, "dur": 34.037,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770628951.943, "dur": 14.965,
    "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": 870, "tid": 870,
    "ts": 7429770628993.331, "dur": 3.847,
    "args": {
      "Python parent id": 4088, "Python id": 4126, "finished": true, "Ev Idx": 10956
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770628994.052, "dur": 0.290,
    "args": {
      "Python parent id": 4126, "Python id": 4127, "finished": true, "Ev Idx": 10957
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770628994.741, "dur": 0.052,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770628996.146, "dur": 0.224,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770628996.786, "dur": 0.105,
    "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": 870, "tid": 870,
    "ts": 7429770628998.321, "dur": 6.511,
    "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": 870, "tid": 870,
    "ts": 7429770629001.662, "dur": 3.076,
    "args": {
      "Python parent id": 4131, "Python id": 4132, "finished": true, "Ev Idx": 10962
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770629001.901, "dur": 1.264,
    "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": 870, "tid": 870,
    "ts": 7429770629002.283, "dur": 0.699,
    "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": 870, "tid": 870,
    "ts": 7429770629003.503, "dur": 0.487,
    "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": 870, "tid": 870,
    "ts": 7429770629004.257, "dur": 0.404,
    "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": 870, "tid": 870,
    "ts": 7429770629007.484, "dur": 327.985,
    "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": 870, "tid": 870,
    "ts": 7429770629008.029, "dur": 44.220,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770629025.871, "dur": 6.056,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770629032.807, "dur": 19.324,
    "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": 870, "tid": 870,
    "ts": 7429770629053.177, "dur": 32.745,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770629067.644, "dur": 3.177,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770629071.311, "dur": 14.547,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770629086.822, "dur": 3.397,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770629090.544, "dur": 110.177,
    "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": 870, "tid": 870,
    "ts": 7429770629238.082, "dur": 17.868,
    "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": 870, "tid": 870,
    "ts": 7429770629238.234, "dur": 0.221,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770629239.064, "dur": 16.669,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770629256.946, "dur": 1.633,
    "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": 870, "tid": 870,
    "ts": 7429770629260.129, "dur": 9.658,
    "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": 870, "tid": 870,
    "ts": 7429770629260.285, "dur": 0.121,
    "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": 870, "tid": 870,
    "ts": 7429770629265.891, "dur": 0.934,
    "args": {
      "Python parent id": 4150, "Python id": 4152, "finished": true, "Ev Idx": 10982
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770629266.350, "dur": 0.364,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770629267.155, "dur": 2.536,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770629270.447, "dur": 45.532,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770629316.619, "dur": 3.171,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770629320.285, "dur": 14.265,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770629338.304, "dur": 3.185,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770629342.217, "dur": 0.270,
    "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": 870, "tid": 870,
    "ts": 7429770629343.442, "dur": 1.002,
    "args": {
      "Python parent id": 4088, "Python id": 4160, "finished": true, "Ev Idx": 10990
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_178", "pid": 870, "tid": 870,
    "ts": 7429770629349.140, "dur": 45.515,
    "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": 870, "tid": 870,
    "ts": 7429770629350.243, "dur": 43.942,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770629350.762, "dur": 1.167,
    "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": 870, "tid": 870,
    "ts": 7429770629353.129, "dur": 40.819,
    "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": 870, "tid": 870,
    "ts": 7429770629353.579, "dur": 0.214,
    "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": 870, "tid": 870,
    "ts": 7429770629354.102, "dur": 0.203,
    "args": {
      "Python parent id": 4164, "Python id": 4166, "finished": true, "Ev Idx": 10996
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770629354.496, "dur": 39.367,
    "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": 870, "tid": 870,
    "ts": 7429770629409.002, "dur": 0.559,
    "args": {
      "Python parent id": 4072, "Python id": 4168, "finished": true, "Ev Idx": 10998
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_51", "pid": 870, "tid": 870,
    "ts": 7429770629411.883, "dur": 83.303,
    "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": 870, "tid": 870,
    "ts": 7429770629412.438, "dur": 82.324,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770629412.715, "dur": 0.450,
    "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": 870, "tid": 870,
    "ts": 7429770629414.847, "dur": 79.096,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770629415.354, "dur": 1.622,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770629417.256, "dur": 13.576,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770629431.136, "dur": 16.637,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770629459.915, "dur": 12.104,
    "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": 870, "tid": 870,
    "ts": 7429770629482.687, "dur": 0.176,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770629483.298, "dur": 1.242,
    "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": 870, "tid": 870,
    "ts": 7429770629495.927, "dur": 0.388,
    "args": {
      "Python parent id": 4072, "Python id": 4179, "finished": true, "Ev Idx": 11009
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_25", "pid": 870, "tid": 870,
    "ts": 7429770629498.028, "dur": 229.372,
    "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": 870, "tid": 870,
    "ts": 7429770629498.563, "dur": 228.596,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770629498.806, "dur": 0.336,
    "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": 870, "tid": 870,
    "ts": 7429770629500.162, "dur": 226.873,
    "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": 870, "tid": 870,
    "ts": 7429770629500.690, "dur": 0.347,
    "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": 870, "tid": 870,
    "ts": 7429770629501.398, "dur": 0.289,
    "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": 870, "tid": 870,
    "ts": 7429770629501.930, "dur": 0.231,
    "args": {
      "Python parent id": 4183, "Python id": 4186, "finished": true, "Ev Idx": 11016
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_179", "pid": 870, "tid": 870,
    "ts": 7429770629504.059, "dur": 41.835,
    "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": 870, "tid": 870,
    "ts": 7429770629504.748, "dur": 40.882,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770629504.871, "dur": 0.173,
    "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": 870, "tid": 870,
    "ts": 7429770629505.980, "dur": 39.549,
    "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": 870, "tid": 870,
    "ts": 7429770629506.251, "dur": 0.112,
    "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": 870, "tid": 870,
    "ts": 7429770629506.572, "dur": 0.151,
    "args": {
      "Python parent id": 4190, "Python id": 4192, "finished": true, "Ev Idx": 11022
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770629506.825, "dur": 38.649,
    "args": {
      "Python parent id": 4190, "Python id": 4193, "finished": true, "Ev Idx": 11023
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_25", "pid": 870, "tid": 870,
    "ts": 7429770629547.747, "dur": 17.166,
    "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": 870, "tid": 870,
    "ts": 7429770629548.199, "dur": 16.532,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770629548.414, "dur": 0.218,
    "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": 870, "tid": 870,
    "ts": 7429770629549.524, "dur": 15.098,
    "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": 870, "tid": 870,
    "ts": 7429770629550.570, "dur": 13.917,
    "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": 870, "tid": 870,
    "ts": 7429770629550.704, "dur": 0.172,
    "args": {
      "Python parent id": 4198, "Python id": 4199, "finished": true, "Ev Idx": 11029
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770629551.507, "dur": 12.893,
    "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": 870, "tid": 870,
    "ts": 7429770629565.515, "dur": 0.382,
    "args": {
      "Python parent id": 4183, "Python id": 4201, "finished": true, "Ev Idx": 11031
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_180", "pid": 870, "tid": 870,
    "ts": 7429770629567.892, "dur": 90.370,
    "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": 870, "tid": 870,
    "ts": 7429770629568.324, "dur": 89.581,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770629568.507, "dur": 0.191,
    "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": 870, "tid": 870,
    "ts": 7429770629569.602, "dur": 88.128,
    "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": 870, "tid": 870,
    "ts": 7429770629569.802, "dur": 0.123,
    "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": 870, "tid": 870,
    "ts": 7429770629570.145, "dur": 0.118,
    "args": {
      "Python parent id": 4205, "Python id": 4207, "finished": true, "Ev Idx": 11037
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770629570.367, "dur": 87.270,
    "args": {
      "Python parent id": 4205, "Python id": 4208, "finished": true, "Ev Idx": 11038
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_181", "pid": 870, "tid": 870,
    "ts": 7429770629671.139, "dur": 55.555,
    "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": 870, "tid": 870,
    "ts": 7429770629671.728, "dur": 54.675,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770629672.009, "dur": 0.496,
    "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": 870, "tid": 870,
    "ts": 7429770629673.334, "dur": 52.976,
    "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": 870, "tid": 870,
    "ts": 7429770629673.681, "dur": 0.185,
    "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": 870, "tid": 870,
    "ts": 7429770629674.118, "dur": 0.143,
    "args": {
      "Python parent id": 4212, "Python id": 4214, "finished": true, "Ev Idx": 11044
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770629674.383, "dur": 51.870,
    "args": {
      "Python parent id": 4212, "Python id": 4215, "finished": true, "Ev Idx": 11045
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_26", "pid": 870, "tid": 870,
    "ts": 7429770629745.310, "dur": 1171.588,
    "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": 870, "tid": 870,
    "ts": 7429770629746.543, "dur": 1169.560,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770629746.805, "dur": 0.260,
    "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": 870, "tid": 870,
    "ts": 7429770629748.588, "dur": 1166.513,
    "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": 870, "tid": 870,
    "ts": 7429770629749.009, "dur": 0.609,
    "args": {
      "Python parent id": 4219, "Python id": 4220, "finished": true, "Ev Idx": 11050
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_52", "pid": 870, "tid": 870,
    "ts": 7429770629751.571, "dur": 74.432,
    "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": 870, "tid": 870,
    "ts": 7429770629752.026, "dur": 73.733,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770629752.161, "dur": 0.304,
    "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": 870, "tid": 870,
    "ts": 7429770629753.226, "dur": 71.808,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770629753.770, "dur": 1.868,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770629755.951, "dur": 12.906,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770629769.187, "dur": 13.364,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770629793.560, "dur": 10.627,
    "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": 870, "tid": 870,
    "ts": 7429770629813.620, "dur": 0.207,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770629814.223, "dur": 1.210,
    "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": 870, "tid": 870,
    "ts": 7429770629826.528, "dur": 0.352,
    "args": {
      "Python parent id": 4219, "Python id": 4231, "finished": true, "Ev Idx": 11061
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_26", "pid": 870, "tid": 870,
    "ts": 7429770629830.037, "dur": 789.303,
    "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": 870, "tid": 870,
    "ts": 7429770629831.511, "dur": 787.055,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770629831.725, "dur": 0.263,
    "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": 870, "tid": 870,
    "ts": 7429770629833.778, "dur": 783.651,
    "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": 870, "tid": 870,
    "ts": 7429770629836.815, "dur": 0.341,
    "args": {
      "Python parent id": 4235, "Python id": 4236, "finished": true, "Ev Idx": 11066
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_182", "pid": 870, "tid": 870,
    "ts": 7429770629839.231, "dur": 41.507,
    "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": 870, "tid": 870,
    "ts": 7429770629839.658, "dur": 40.845,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770629839.791, "dur": 0.318,
    "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": 870, "tid": 870,
    "ts": 7429770629840.870, "dur": 39.531,
    "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": 870, "tid": 870,
    "ts": 7429770629841.109, "dur": 0.131,
    "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": 870, "tid": 870,
    "ts": 7429770629841.449, "dur": 0.146,
    "args": {
      "Python parent id": 4240, "Python id": 4242, "finished": true, "Ev Idx": 11072
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770629841.711, "dur": 38.618,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770629881.363, "dur": 3.309,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770629885.108, "dur": 3.538,
    "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": 870, "tid": 870,
    "ts": 7429770629889.484, "dur": 0.383,
    "args": {
      "Python parent id": 4235, "Python id": 4246, "finished": true, "Ev Idx": 11076
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_183", "pid": 870, "tid": 870,
    "ts": 7429770629892.125, "dur": 35.685,
    "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": 870, "tid": 870,
    "ts": 7429770629892.925, "dur": 34.603,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770629893.159, "dur": 0.226,
    "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": 870, "tid": 870,
    "ts": 7429770629894.152, "dur": 33.269,
    "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": 870, "tid": 870,
    "ts": 7429770629894.421, "dur": 0.465,
    "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": 870, "tid": 870,
    "ts": 7429770629895.148, "dur": 0.196,
    "args": {
      "Python parent id": 4250, "Python id": 4252, "finished": true, "Ev Idx": 11082
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770629895.462, "dur": 31.917,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770629928.271, "dur": 2.323,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770629931.100, "dur": 2.517,
    "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": 870, "tid": 870,
    "ts": 7429770629934.395, "dur": 0.324,
    "args": {
      "Python parent id": 4235, "Python id": 4256, "finished": true, "Ev Idx": 11086
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_184", "pid": 870, "tid": 870,
    "ts": 7429770629937.055, "dur": 36.787,
    "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": 870, "tid": 870,
    "ts": 7429770629937.694, "dur": 35.930,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770629937.885, "dur": 0.226,
    "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": 870, "tid": 870,
    "ts": 7429770629938.852, "dur": 34.676,
    "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": 870, "tid": 870,
    "ts": 7429770629939.110, "dur": 0.121,
    "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": 870, "tid": 870,
    "ts": 7429770629939.668, "dur": 0.116,
    "args": {
      "Python parent id": 4260, "Python id": 4262, "finished": true, "Ev Idx": 11092
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770629939.888, "dur": 33.579,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770629974.391, "dur": 2.290,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770629977.178, "dur": 2.611,
    "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": 870, "tid": 870,
    "ts": 7429770629980.894, "dur": 177.561,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770629981.113, "dur": 3.888,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770629985.283, "dur": 1.942,
    "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": 870, "tid": 870,
    "ts": 7429770629997.965, "dur": 50.256,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770630025.105, "dur": 21.807,
    "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": 870, "tid": 870,
    "ts": 7429770630103.947, "dur": 33.630,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770630123.717, "dur": 12.944,
    "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": 870, "tid": 870,
    "ts": 7429770630163.421, "dur": 4.089,
    "args": {
      "Python parent id": 4235, "Python id": 4273, "finished": true, "Ev Idx": 11103
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770630164.112, "dur": 0.274,
    "args": {
      "Python parent id": 4273, "Python id": 4274, "finished": true, "Ev Idx": 11104
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770630164.789, "dur": 0.079,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770630166.197, "dur": 0.242,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770630167.099, "dur": 0.101,
    "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": 870, "tid": 870,
    "ts": 7429770630168.796, "dur": 6.673,
    "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": 870, "tid": 870,
    "ts": 7429770630172.238, "dur": 3.107,
    "args": {
      "Python parent id": 4278, "Python id": 4279, "finished": true, "Ev Idx": 11109
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770630172.562, "dur": 1.247,
    "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": 870, "tid": 870,
    "ts": 7429770630173.022, "dur": 0.656,
    "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": 870, "tid": 870,
    "ts": 7429770630174.081, "dur": 0.461,
    "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": 870, "tid": 870,
    "ts": 7429770630174.840, "dur": 0.429,
    "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": 870, "tid": 870,
    "ts": 7429770630178.427, "dur": 371.572,
    "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": 870, "tid": 870,
    "ts": 7429770630178.995, "dur": 73.425,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770630198.277, "dur": 29.874,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770630228.962, "dur": 23.342,
    "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": 870, "tid": 870,
    "ts": 7429770630253.236, "dur": 28.468,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770630266.836, "dur": 2.796,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770630270.128, "dur": 11.510,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770630282.687, "dur": 3.447,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770630286.499, "dur": 118.497,
    "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": 870, "tid": 870,
    "ts": 7429770630447.687, "dur": 19.915,
    "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": 870, "tid": 870,
    "ts": 7429770630447.901, "dur": 0.285,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770630448.964, "dur": 18.423,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770630468.775, "dur": 2.069,
    "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": 870, "tid": 870,
    "ts": 7429770630472.263, "dur": 12.066,
    "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": 870, "tid": 870,
    "ts": 7429770630472.474, "dur": 0.117,
    "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": 870, "tid": 870,
    "ts": 7429770630479.888, "dur": 1.057,
    "args": {
      "Python parent id": 4297, "Python id": 4299, "finished": true, "Ev Idx": 11129
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770630480.476, "dur": 0.385,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770630481.421, "dur": 2.828,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770630485.022, "dur": 44.835,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770630530.524, "dur": 3.434,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770630534.526, "dur": 14.490,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770630553.644, "dur": 3.232,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770630557.637, "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": 870, "tid": 870,
    "ts": 7429770630558.943, "dur": 1.343,
    "args": {
      "Python parent id": 4235, "Python id": 4307, "finished": true, "Ev Idx": 11137
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_185", "pid": 870, "tid": 870,
    "ts": 7429770630566.598, "dur": 50.116,
    "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": 870, "tid": 870,
    "ts": 7429770630568.175, "dur": 47.950,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770630568.749, "dur": 1.753,
    "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": 870, "tid": 870,
    "ts": 7429770630572.663, "dur": 43.171,
    "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": 870, "tid": 870,
    "ts": 7429770630573.165, "dur": 0.348,
    "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": 870, "tid": 870,
    "ts": 7429770630573.889, "dur": 0.196,
    "args": {
      "Python parent id": 4311, "Python id": 4313, "finished": true, "Ev Idx": 11143
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770630574.337, "dur": 41.387,
    "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": 870, "tid": 870,
    "ts": 7429770630634.692, "dur": 0.745,
    "args": {
      "Python parent id": 4219, "Python id": 4315, "finished": true, "Ev Idx": 11145
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_53", "pid": 870, "tid": 870,
    "ts": 7429770630638.551, "dur": 93.412,
    "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": 870, "tid": 870,
    "ts": 7429770630639.536, "dur": 92.017,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770630639.880, "dur": 0.536,
    "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": 870, "tid": 870,
    "ts": 7429770630641.676, "dur": 88.936,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770630642.427, "dur": 1.776,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770630644.597, "dur": 17.392,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770630662.427, "dur": 18.602,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770630694.447, "dur": 11.695,
    "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": 870, "tid": 870,
    "ts": 7429770630717.785, "dur": 0.351,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770630718.649, "dur": 1.152,
    "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": 870, "tid": 870,
    "ts": 7429770630732.817, "dur": 0.476,
    "args": {
      "Python parent id": 4219, "Python id": 4326, "finished": true, "Ev Idx": 11156
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_26", "pid": 870, "tid": 870,
    "ts": 7429770630735.312, "dur": 168.780,
    "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": 870, "tid": 870,
    "ts": 7429770630736.162, "dur": 167.702,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770630736.383, "dur": 0.397,
    "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": 870, "tid": 870,
    "ts": 7429770630737.905, "dur": 165.837,
    "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": 870, "tid": 870,
    "ts": 7429770630738.526, "dur": 0.379,
    "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": 870, "tid": 870,
    "ts": 7429770630739.339, "dur": 0.358,
    "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": 870, "tid": 870,
    "ts": 7429770630739.955, "dur": 0.300,
    "args": {
      "Python parent id": 4330, "Python id": 4333, "finished": true, "Ev Idx": 11163
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_186", "pid": 870, "tid": 870,
    "ts": 7429770630742.574, "dur": 42.097,
    "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": 870, "tid": 870,
    "ts": 7429770630743.186, "dur": 41.122,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770630743.349, "dur": 0.159,
    "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": 870, "tid": 870,
    "ts": 7429770630744.278, "dur": 39.847,
    "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": 870, "tid": 870,
    "ts": 7429770630744.662, "dur": 0.144,
    "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": 870, "tid": 870,
    "ts": 7429770630745.072, "dur": 0.151,
    "args": {
      "Python parent id": 4337, "Python id": 4339, "finished": true, "Ev Idx": 11169
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770630745.367, "dur": 38.671,
    "args": {
      "Python parent id": 4337, "Python id": 4340, "finished": true, "Ev Idx": 11170
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_26", "pid": 870, "tid": 870,
    "ts": 7429770630786.391, "dur": 18.407,
    "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": 870, "tid": 870,
    "ts": 7429770630786.905, "dur": 17.579,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770630787.104, "dur": 0.358,
    "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": 870, "tid": 870,
    "ts": 7429770630788.802, "dur": 15.557,
    "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": 870, "tid": 870,
    "ts": 7429770630789.791, "dur": 14.390,
    "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": 870, "tid": 870,
    "ts": 7429770630789.920, "dur": 0.201,
    "args": {
      "Python parent id": 4345, "Python id": 4346, "finished": true, "Ev Idx": 11176
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770630790.954, "dur": 13.152,
    "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": 870, "tid": 870,
    "ts": 7429770630805.479, "dur": 0.386,
    "args": {
      "Python parent id": 4330, "Python id": 4348, "finished": true, "Ev Idx": 11178
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_187", "pid": 870, "tid": 870,
    "ts": 7429770630807.741, "dur": 37.005,
    "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": 870, "tid": 870,
    "ts": 7429770630808.241, "dur": 36.201,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770630808.396, "dur": 0.216,
    "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": 870, "tid": 870,
    "ts": 7429770630809.231, "dur": 35.091,
    "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": 870, "tid": 870,
    "ts": 7429770630809.498, "dur": 0.136,
    "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": 870, "tid": 870,
    "ts": 7429770630809.836, "dur": 0.124,
    "args": {
      "Python parent id": 4352, "Python id": 4354, "finished": true, "Ev Idx": 11184
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770630810.065, "dur": 34.203,
    "args": {
      "Python parent id": 4352, "Python id": 4355, "finished": true, "Ev Idx": 11185
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_188", "pid": 870, "tid": 870,
    "ts": 7429770630856.000, "dur": 47.390,
    "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": 870, "tid": 870,
    "ts": 7429770630856.509, "dur": 46.586,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770630856.721, "dur": 0.246,
    "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": 870, "tid": 870,
    "ts": 7429770630857.753, "dur": 45.184,
    "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": 870, "tid": 870,
    "ts": 7429770630857.991, "dur": 0.138,
    "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": 870, "tid": 870,
    "ts": 7429770630858.349, "dur": 0.123,
    "args": {
      "Python parent id": 4359, "Python id": 4361, "finished": true, "Ev Idx": 11191
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770630858.575, "dur": 44.294,
    "args": {
      "Python parent id": 4359, "Python id": 4362, "finished": true, "Ev Idx": 11192
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_27", "pid": 870, "tid": 870,
    "ts": 7429770630923.487, "dur": 1107.010,
    "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": 870, "tid": 870,
    "ts": 7429770630925.083, "dur": 1104.653,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770630925.266, "dur": 0.273,
    "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": 870, "tid": 870,
    "ts": 7429770630926.873, "dur": 1101.833,
    "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": 870, "tid": 870,
    "ts": 7429770630927.285, "dur": 0.553,
    "args": {
      "Python parent id": 4366, "Python id": 4367, "finished": true, "Ev Idx": 11197
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_54", "pid": 870, "tid": 870,
    "ts": 7429770630929.562, "dur": 72.107,
    "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": 870, "tid": 870,
    "ts": 7429770630930.079, "dur": 71.259,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770630930.229, "dur": 0.152,
    "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": 870, "tid": 870,
    "ts": 7429770630931.482, "dur": 69.179,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770630931.918, "dur": 1.582,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770630933.814, "dur": 12.402,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770630946.565, "dur": 13.086,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770630970.847, "dur": 9.679,
    "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": 870, "tid": 870,
    "ts": 7429770630990.184, "dur": 0.252,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770630991.012, "dur": 1.262,
    "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": 870, "tid": 870,
    "ts": 7429770631002.240, "dur": 0.327,
    "args": {
      "Python parent id": 4366, "Python id": 4378, "finished": true, "Ev Idx": 11208
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_27", "pid": 870, "tid": 870,
    "ts": 7429770631005.369, "dur": 686.114,
    "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": 870, "tid": 870,
    "ts": 7429770631006.940, "dur": 683.906,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770631007.131, "dur": 0.250,
    "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": 870, "tid": 870,
    "ts": 7429770631009.162, "dur": 680.801,
    "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": 870, "tid": 870,
    "ts": 7429770631012.103, "dur": 0.352,
    "args": {
      "Python parent id": 4382, "Python id": 4383, "finished": true, "Ev Idx": 11213
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_189", "pid": 870, "tid": 870,
    "ts": 7429770631014.386, "dur": 41.463,
    "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": 870, "tid": 870,
    "ts": 7429770631014.906, "dur": 40.627,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770631015.044, "dur": 0.164,
    "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": 870, "tid": 870,
    "ts": 7429770631016.203, "dur": 39.187,
    "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": 870, "tid": 870,
    "ts": 7429770631016.508, "dur": 0.144,
    "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": 870, "tid": 870,
    "ts": 7429770631016.880, "dur": 0.134,
    "args": {
      "Python parent id": 4387, "Python id": 4389, "finished": true, "Ev Idx": 11219
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770631017.120, "dur": 38.198,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770631056.544, "dur": 3.439,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770631060.420, "dur": 3.340,
    "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": 870, "tid": 870,
    "ts": 7429770631064.471, "dur": 0.393,
    "args": {
      "Python parent id": 4382, "Python id": 4393, "finished": true, "Ev Idx": 11223
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_190", "pid": 870, "tid": 870,
    "ts": 7429770631066.809, "dur": 34.368,
    "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": 870, "tid": 870,
    "ts": 7429770631067.377, "dur": 33.484,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770631067.587, "dur": 0.253,
    "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": 870, "tid": 870,
    "ts": 7429770631068.514, "dur": 32.253,
    "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": 870, "tid": 870,
    "ts": 7429770631068.756, "dur": 0.131,
    "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": 870, "tid": 870,
    "ts": 7429770631069.090, "dur": 0.154,
    "args": {
      "Python parent id": 4397, "Python id": 4399, "finished": true, "Ev Idx": 11229
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770631069.368, "dur": 31.330,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770631101.712, "dur": 2.502,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770631104.680, "dur": 2.444,
    "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": 870, "tid": 870,
    "ts": 7429770631107.835, "dur": 0.453,
    "args": {
      "Python parent id": 4382, "Python id": 4403, "finished": true, "Ev Idx": 11233
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_191", "pid": 870, "tid": 870,
    "ts": 7429770631110.508, "dur": 36.083,
    "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": 870, "tid": 870,
    "ts": 7429770631110.955, "dur": 35.375,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770631111.172, "dur": 0.262,
    "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": 870, "tid": 870,
    "ts": 7429770631112.128, "dur": 34.084,
    "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": 870, "tid": 870,
    "ts": 7429770631112.347, "dur": 0.159,
    "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": 870, "tid": 870,
    "ts": 7429770631112.715, "dur": 0.132,
    "args": {
      "Python parent id": 4407, "Python id": 4409, "finished": true, "Ev Idx": 11239
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770631112.954, "dur": 33.197,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770631147.205, "dur": 2.305,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770631149.905, "dur": 2.413,
    "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": 870, "tid": 870,
    "ts": 7429770631153.439, "dur": 142.262,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770631153.731, "dur": 3.308,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770631157.329, "dur": 1.644,
    "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": 870, "tid": 870,
    "ts": 7429770631171.131, "dur": 46.527,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770631196.113, "dur": 20.463,
    "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": 870, "tid": 870,
    "ts": 7429770631249.005, "dur": 27.867,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770631265.026, "dur": 11.041,
    "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": 870, "tid": 870,
    "ts": 7429770631299.793, "dur": 3.668,
    "args": {
      "Python parent id": 4382, "Python id": 4420, "finished": true, "Ev Idx": 11250
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770631300.509, "dur": 0.238,
    "args": {
      "Python parent id": 4420, "Python id": 4421, "finished": true, "Ev Idx": 11251
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770631301.184, "dur": 0.080,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770631302.417, "dur": 0.226,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770631303.137, "dur": 0.089,
    "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": 870, "tid": 870,
    "ts": 7429770631304.699, "dur": 6.327,
    "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": 870, "tid": 870,
    "ts": 7429770631307.863, "dur": 3.052,
    "args": {
      "Python parent id": 4425, "Python id": 4426, "finished": true, "Ev Idx": 11256
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770631308.175, "dur": 1.196,
    "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": 870, "tid": 870,
    "ts": 7429770631308.566, "dur": 0.699,
    "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": 870, "tid": 870,
    "ts": 7429770631309.651, "dur": 0.439,
    "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": 870, "tid": 870,
    "ts": 7429770631310.419, "dur": 0.419,
    "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": 870, "tid": 870,
    "ts": 7429770631313.625, "dur": 318.454,
    "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": 870, "tid": 870,
    "ts": 7429770631314.041, "dur": 42.988,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770631331.727, "dur": 5.088,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770631337.652, "dur": 19.265,
    "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": 870, "tid": 870,
    "ts": 7429770631357.788, "dur": 31.892,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770631372.327, "dur": 2.973,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770631375.820, "dur": 13.797,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770631390.516, "dur": 2.807,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770631393.609, "dur": 107.871,
    "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": 870, "tid": 870,
    "ts": 7429770631538.598, "dur": 17.911,
    "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": 870, "tid": 870,
    "ts": 7429770631538.851, "dur": 0.153,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770631539.618, "dur": 16.716,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770631557.364, "dur": 1.750,
    "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": 870, "tid": 870,
    "ts": 7429770631560.521, "dur": 10.161,
    "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": 870, "tid": 870,
    "ts": 7429770631560.694, "dur": 0.117,
    "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": 870, "tid": 870,
    "ts": 7429770631566.613, "dur": 0.945,
    "args": {
      "Python parent id": 4444, "Python id": 4446, "finished": true, "Ev Idx": 11276
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770631567.092, "dur": 0.379,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770631567.987, "dur": 2.627,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770631571.481, "dur": 41.254,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770631613.273, "dur": 3.284,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770631617.085, "dur": 13.915,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770631634.976, "dur": 3.139,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770631638.818, "dur": 0.227,
    "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": 870, "tid": 870,
    "ts": 7429770631639.942, "dur": 1.063,
    "args": {
      "Python parent id": 4382, "Python id": 4454, "finished": true, "Ev Idx": 11284
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_192", "pid": 870, "tid": 870,
    "ts": 7429770631645.285, "dur": 44.012,
    "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": 870, "tid": 870,
    "ts": 7429770631646.456, "dur": 42.425,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770631646.991, "dur": 1.324,
    "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": 870, "tid": 870,
    "ts": 7429770631649.422, "dur": 39.179,
    "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": 870, "tid": 870,
    "ts": 7429770631649.831, "dur": 0.255,
    "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": 870, "tid": 870,
    "ts": 7429770631650.371, "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": 870, "tid": 870,
    "ts": 7429770631650.704, "dur": 37.804,
    "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": 870, "tid": 870,
    "ts": 7429770631703.399, "dur": 0.634,
    "args": {
      "Python parent id": 4366, "Python id": 4462, "finished": true, "Ev Idx": 11292
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_55", "pid": 870, "tid": 870,
    "ts": 7429770631706.644, "dur": 83.610,
    "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": 870, "tid": 870,
    "ts": 7429770631707.370, "dur": 82.510,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770631707.630, "dur": 0.458,
    "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": 870, "tid": 870,
    "ts": 7429770631709.184, "dur": 79.866,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770631709.874, "dur": 1.480,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770631711.727, "dur": 14.172,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770631726.245, "dur": 16.246,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770631754.045, "dur": 12.359,
    "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": 870, "tid": 870,
    "ts": 7429770631778.278, "dur": 0.312,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770631778.996, "dur": 1.040,
    "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": 870, "tid": 870,
    "ts": 7429770631790.889, "dur": 0.424,
    "args": {
      "Python parent id": 4366, "Python id": 4473, "finished": true, "Ev Idx": 11303
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_27", "pid": 870, "tid": 870,
    "ts": 7429770631793.275, "dur": 223.277,
    "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": 870, "tid": 870,
    "ts": 7429770631793.804, "dur": 222.581,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770631794.061, "dur": 0.386,
    "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": 870, "tid": 870,
    "ts": 7429770631795.461, "dur": 220.804,
    "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": 870, "tid": 870,
    "ts": 7429770631796.057, "dur": 0.297,
    "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": 870, "tid": 870,
    "ts": 7429770631796.812, "dur": 0.299,
    "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": 870, "tid": 870,
    "ts": 7429770631797.372, "dur": 0.243,
    "args": {
      "Python parent id": 4477, "Python id": 4480, "finished": true, "Ev Idx": 11310
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_193", "pid": 870, "tid": 870,
    "ts": 7429770631799.936, "dur": 40.097,
    "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": 870, "tid": 870,
    "ts": 7429770631800.531, "dur": 39.292,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770631800.666, "dur": 0.177,
    "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": 870, "tid": 870,
    "ts": 7429770631801.754, "dur": 37.914,
    "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": 870, "tid": 870,
    "ts": 7429770631801.963, "dur": 0.121,
    "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": 870, "tid": 870,
    "ts": 7429770631802.285, "dur": 0.149,
    "args": {
      "Python parent id": 4484, "Python id": 4486, "finished": true, "Ev Idx": 11316
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770631802.555, "dur": 37.051,
    "args": {
      "Python parent id": 4484, "Python id": 4487, "finished": true, "Ev Idx": 11317
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_27", "pid": 870, "tid": 870,
    "ts": 7429770631842.022, "dur": 17.263,
    "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": 870, "tid": 870,
    "ts": 7429770631842.617, "dur": 16.479,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770631842.847, "dur": 0.257,
    "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": 870, "tid": 870,
    "ts": 7429770631843.980, "dur": 15.033,
    "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": 870, "tid": 870,
    "ts": 7429770631844.861, "dur": 13.980,
    "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": 870, "tid": 870,
    "ts": 7429770631844.957, "dur": 0.136,
    "args": {
      "Python parent id": 4492, "Python id": 4493, "finished": true, "Ev Idx": 11323
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770631845.883, "dur": 12.877,
    "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": 870, "tid": 870,
    "ts": 7429770631859.923, "dur": 0.367,
    "args": {
      "Python parent id": 4477, "Python id": 4495, "finished": true, "Ev Idx": 11325
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_194", "pid": 870, "tid": 870,
    "ts": 7429770631861.958, "dur": 36.889,
    "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": 870, "tid": 870,
    "ts": 7429770631862.386, "dur": 36.233,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770631862.551, "dur": 0.198,
    "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": 870, "tid": 870,
    "ts": 7429770631863.467, "dur": 35.041,
    "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": 870, "tid": 870,
    "ts": 7429770631863.662, "dur": 0.127,
    "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": 870, "tid": 870,
    "ts": 7429770631864.027, "dur": 0.131,
    "args": {
      "Python parent id": 4499, "Python id": 4501, "finished": true, "Ev Idx": 11331
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770631864.279, "dur": 34.176,
    "args": {
      "Python parent id": 4499, "Python id": 4502, "finished": true, "Ev Idx": 11332
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_195", "pid": 870, "tid": 870,
    "ts": 7429770631910.308, "dur": 105.614,
    "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": 870, "tid": 870,
    "ts": 7429770631910.881, "dur": 104.651,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770631911.055, "dur": 0.244,
    "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": 870, "tid": 870,
    "ts": 7429770631911.956, "dur": 103.386,
    "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": 870, "tid": 870,
    "ts": 7429770631912.159, "dur": 0.146,
    "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": 870, "tid": 870,
    "ts": 7429770631912.497, "dur": 0.115,
    "args": {
      "Python parent id": 4506, "Python id": 4508, "finished": true, "Ev Idx": 11338
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770631912.717, "dur": 102.459,
    "args": {
      "Python parent id": 4506, "Python id": 4509, "finished": true, "Ev Idx": 11339
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_28", "pid": 870, "tid": 870,
    "ts": 7429770632036.906, "dur": 1142.226,
    "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": 870, "tid": 870,
    "ts": 7429770632038.217, "dur": 1139.918,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770632038.492, "dur": 0.445,
    "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": 870, "tid": 870,
    "ts": 7429770632040.420, "dur": 1136.847,
    "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": 870, "tid": 870,
    "ts": 7429770632040.920, "dur": 0.641,
    "args": {
      "Python parent id": 4513, "Python id": 4514, "finished": true, "Ev Idx": 11344
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_56", "pid": 870, "tid": 870,
    "ts": 7429770632043.814, "dur": 75.683,
    "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": 870, "tid": 870,
    "ts": 7429770632044.422, "dur": 74.797,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770632044.570, "dur": 0.195,
    "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": 870, "tid": 870,
    "ts": 7429770632045.670, "dur": 72.885,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770632046.157, "dur": 1.895,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770632048.352, "dur": 12.644,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770632061.327, "dur": 14.068,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770632088.518, "dur": 10.013,
    "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": 870, "tid": 870,
    "ts": 7429770632107.929, "dur": 0.212,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770632108.513, "dur": 1.127,
    "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": 870, "tid": 870,
    "ts": 7429770632120.090, "dur": 0.347,
    "args": {
      "Python parent id": 4513, "Python id": 4525, "finished": true, "Ev Idx": 11355
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_28", "pid": 870, "tid": 870,
    "ts": 7429770632123.267, "dur": 743.793,
    "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": 870, "tid": 870,
    "ts": 7429770632124.637, "dur": 741.736,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770632124.863, "dur": 0.248,
    "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": 870, "tid": 870,
    "ts": 7429770632126.732, "dur": 738.708,
    "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": 870, "tid": 870,
    "ts": 7429770632130.075, "dur": 0.506,
    "args": {
      "Python parent id": 4529, "Python id": 4530, "finished": true, "Ev Idx": 11360
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_196", "pid": 870, "tid": 870,
    "ts": 7429770632132.533, "dur": 42.547,
    "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": 870, "tid": 870,
    "ts": 7429770632133.100, "dur": 41.698,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770632133.241, "dur": 0.157,
    "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": 870, "tid": 870,
    "ts": 7429770632134.168, "dur": 40.513,
    "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": 870, "tid": 870,
    "ts": 7429770632134.474, "dur": 0.171,
    "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": 870, "tid": 870,
    "ts": 7429770632134.858, "dur": 0.162,
    "args": {
      "Python parent id": 4534, "Python id": 4536, "finished": true, "Ev Idx": 11366
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770632135.130, "dur": 39.483,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770632176.386, "dur": 3.609,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770632180.456, "dur": 3.830,
    "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": 870, "tid": 870,
    "ts": 7429770632185.109, "dur": 0.464,
    "args": {
      "Python parent id": 4529, "Python id": 4540, "finished": true, "Ev Idx": 11370
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_197", "pid": 870, "tid": 870,
    "ts": 7429770632187.626, "dur": 37.756,
    "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": 870, "tid": 870,
    "ts": 7429770632188.090, "dur": 37.022,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770632188.302, "dur": 0.234,
    "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": 870, "tid": 870,
    "ts": 7429770632189.255, "dur": 35.734,
    "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": 870, "tid": 870,
    "ts": 7429770632189.541, "dur": 0.121,
    "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": 870, "tid": 870,
    "ts": 7429770632189.844, "dur": 0.119,
    "args": {
      "Python parent id": 4544, "Python id": 4546, "finished": true, "Ev Idx": 11376
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770632190.080, "dur": 34.846,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770632225.897, "dur": 2.429,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770632228.721, "dur": 2.503,
    "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": 870, "tid": 870,
    "ts": 7429770632232.028, "dur": 0.409,
    "args": {
      "Python parent id": 4529, "Python id": 4550, "finished": true, "Ev Idx": 11380
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_198", "pid": 870, "tid": 870,
    "ts": 7429770632234.281, "dur": 35.834,
    "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": 870, "tid": 870,
    "ts": 7429770632234.849, "dur": 34.974,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770632235.070, "dur": 0.216,
    "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": 870, "tid": 870,
    "ts": 7429770632235.917, "dur": 33.806,
    "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": 870, "tid": 870,
    "ts": 7429770632236.201, "dur": 0.185,
    "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": 870, "tid": 870,
    "ts": 7429770632236.564, "dur": 0.136,
    "args": {
      "Python parent id": 4554, "Python id": 4556, "finished": true, "Ev Idx": 11386
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770632236.809, "dur": 32.867,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770632270.664, "dur": 2.335,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770632273.430, "dur": 2.393,
    "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": 870, "tid": 870,
    "ts": 7429770632276.887, "dur": 161.963,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770632277.167, "dur": 3.609,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770632281.025, "dur": 1.922,
    "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": 870, "tid": 870,
    "ts": 7429770632293.691, "dur": 63.382,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770632333.508, "dur": 22.126,
    "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": 870, "tid": 870,
    "ts": 7429770632390.276, "dur": 28.986,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770632407.520, "dur": 10.818,
    "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": 870, "tid": 870,
    "ts": 7429770632443.617, "dur": 3.917,
    "args": {
      "Python parent id": 4529, "Python id": 4567, "finished": true, "Ev Idx": 11397
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770632444.421, "dur": 0.284,
    "args": {
      "Python parent id": 4567, "Python id": 4568, "finished": true, "Ev Idx": 11398
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770632445.020, "dur": 0.087,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770632446.481, "dur": 0.267,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770632447.161, "dur": 0.098,
    "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": 870, "tid": 870,
    "ts": 7429770632448.862, "dur": 7.179,
    "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": 870, "tid": 870,
    "ts": 7429770632452.293, "dur": 3.604,
    "args": {
      "Python parent id": 4572, "Python id": 4573, "finished": true, "Ev Idx": 11403
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770632452.584, "dur": 1.576,
    "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": 870, "tid": 870,
    "ts": 7429770632453.040, "dur": 0.940,
    "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": 870, "tid": 870,
    "ts": 7429770632454.492, "dur": 0.544,
    "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": 870, "tid": 870,
    "ts": 7429770632455.393, "dur": 0.429,
    "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": 870, "tid": 870,
    "ts": 7429770632458.868, "dur": 344.597,
    "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": 870, "tid": 870,
    "ts": 7429770632459.405, "dur": 45.453,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770632476.893, "dur": 5.132,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770632482.897, "dur": 21.859,
    "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": 870, "tid": 870,
    "ts": 7429770632505.757, "dur": 30.455,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770632518.813, "dur": 3.177,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770632522.484, "dur": 13.663,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770632537.146, "dur": 2.995,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770632540.430, "dur": 110.552,
    "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": 870, "tid": 870,
    "ts": 7429770632692.458, "dur": 18.820,
    "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": 870, "tid": 870,
    "ts": 7429770632692.785, "dur": 0.228,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770632693.773, "dur": 17.287,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770632712.242, "dur": 2.158,
    "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": 870, "tid": 870,
    "ts": 7429770632715.918, "dur": 12.113,
    "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": 870, "tid": 870,
    "ts": 7429770632716.118, "dur": 0.087,
    "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": 870, "tid": 870,
    "ts": 7429770632723.413, "dur": 1.142,
    "args": {
      "Python parent id": 4591, "Python id": 4593, "finished": true, "Ev Idx": 11423
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770632724.039, "dur": 0.429,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770632725.031, "dur": 2.927,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770632728.812, "dur": 45.462,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770632775.038, "dur": 3.432,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770632778.959, "dur": 23.403,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770632806.882, "dur": 3.330,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770632810.951, "dur": 0.283,
    "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": 870, "tid": 870,
    "ts": 7429770632812.394, "dur": 1.346,
    "args": {
      "Python parent id": 4529, "Python id": 4601, "finished": true, "Ev Idx": 11431
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_199", "pid": 870, "tid": 870,
    "ts": 7429770632819.630, "dur": 45.165,
    "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": 870, "tid": 870,
    "ts": 7429770632821.182, "dur": 43.108,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770632821.793, "dur": 1.495,
    "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": 870, "tid": 870,
    "ts": 7429770632824.861, "dur": 39.103,
    "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": 870, "tid": 870,
    "ts": 7429770632825.301, "dur": 0.343,
    "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": 870, "tid": 870,
    "ts": 7429770632826.012, "dur": 0.163,
    "args": {
      "Python parent id": 4605, "Python id": 4607, "finished": true, "Ev Idx": 11437
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770632826.359, "dur": 37.508,
    "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": 870, "tid": 870,
    "ts": 7429770632879.799, "dur": 0.622,
    "args": {
      "Python parent id": 4513, "Python id": 4609, "finished": true, "Ev Idx": 11439
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_57", "pid": 870, "tid": 870,
    "ts": 7429770632883.144, "dur": 110.897,
    "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": 870, "tid": 870,
    "ts": 7429770632883.939, "dur": 109.713,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770632884.146, "dur": 0.469,
    "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": 870, "tid": 870,
    "ts": 7429770632885.925, "dur": 106.797,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770632886.630, "dur": 1.875,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770632888.844, "dur": 15.521,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770632904.751, "dur": 16.847,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770632934.938, "dur": 11.103,
    "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": 870, "tid": 870,
    "ts": 7429770632957.855, "dur": 0.352,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770632958.661, "dur": 1.298,
    "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": 870, "tid": 870,
    "ts": 7429770632994.972, "dur": 0.828,
    "args": {
      "Python parent id": 4513, "Python id": 4620, "finished": true, "Ev Idx": 11450
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_28", "pid": 870, "tid": 870,
    "ts": 7429770632997.803, "dur": 167.820,
    "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": 870, "tid": 870,
    "ts": 7429770632998.416, "dur": 166.983,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770632998.670, "dur": 0.382,
    "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": 870, "tid": 870,
    "ts": 7429770633000.104, "dur": 165.132,
    "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": 870, "tid": 870,
    "ts": 7429770633000.599, "dur": 0.356,
    "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": 870, "tid": 870,
    "ts": 7429770633001.331, "dur": 0.297,
    "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": 870, "tid": 870,
    "ts": 7429770633001.898, "dur": 0.269,
    "args": {
      "Python parent id": 4624, "Python id": 4627, "finished": true, "Ev Idx": 11457
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_200", "pid": 870, "tid": 870,
    "ts": 7429770633004.308, "dur": 41.148,
    "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": 870, "tid": 870,
    "ts": 7429770633004.911, "dur": 40.183,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770633005.034, "dur": 0.177,
    "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": 870, "tid": 870,
    "ts": 7429770633005.877, "dur": 39.080,
    "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": 870, "tid": 870,
    "ts": 7429770633006.157, "dur": 0.111,
    "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": 870, "tid": 870,
    "ts": 7429770633006.520, "dur": 0.151,
    "args": {
      "Python parent id": 4631, "Python id": 4633, "finished": true, "Ev Idx": 11463
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770633006.827, "dur": 38.043,
    "args": {
      "Python parent id": 4631, "Python id": 4634, "finished": true, "Ev Idx": 11464
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_28", "pid": 870, "tid": 870,
    "ts": 7429770633047.628, "dur": 17.328,
    "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": 870, "tid": 870,
    "ts": 7429770633048.212, "dur": 16.452,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770633048.413, "dur": 0.251,
    "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": 870, "tid": 870,
    "ts": 7429770633049.698, "dur": 14.841,
    "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": 870, "tid": 870,
    "ts": 7429770633050.666, "dur": 13.703,
    "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": 870, "tid": 870,
    "ts": 7429770633050.842, "dur": 0.177,
    "args": {
      "Python parent id": 4639, "Python id": 4640, "finished": true, "Ev Idx": 11470
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770633051.892, "dur": 12.397,
    "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": 870, "tid": 870,
    "ts": 7429770633065.717, "dur": 0.365,
    "args": {
      "Python parent id": 4624, "Python id": 4642, "finished": true, "Ev Idx": 11472
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_201", "pid": 870, "tid": 870,
    "ts": 7429770633067.737, "dur": 36.729,
    "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": 870, "tid": 870,
    "ts": 7429770633068.201, "dur": 35.978,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770633068.358, "dur": 0.216,
    "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": 870, "tid": 870,
    "ts": 7429770633069.471, "dur": 34.606,
    "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": 870, "tid": 870,
    "ts": 7429770633069.758, "dur": 0.123,
    "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": 870, "tid": 870,
    "ts": 7429770633070.103, "dur": 0.118,
    "args": {
      "Python parent id": 4646, "Python id": 4648, "finished": true, "Ev Idx": 11478
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770633070.326, "dur": 33.688,
    "args": {
      "Python parent id": 4646, "Python id": 4649, "finished": true, "Ev Idx": 11479
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_202", "pid": 870, "tid": 870,
    "ts": 7429770633115.986, "dur": 48.922,
    "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": 870, "tid": 870,
    "ts": 7429770633116.531, "dur": 48.161,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770633116.761, "dur": 0.231,
    "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": 870, "tid": 870,
    "ts": 7429770633117.897, "dur": 46.691,
    "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": 870, "tid": 870,
    "ts": 7429770633118.148, "dur": 0.179,
    "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": 870, "tid": 870,
    "ts": 7429770633118.524, "dur": 0.132,
    "args": {
      "Python parent id": 4653, "Python id": 4655, "finished": true, "Ev Idx": 11485
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770633118.776, "dur": 45.741,
    "args": {
      "Python parent id": 4653, "Python id": 4656, "finished": true, "Ev Idx": 11486
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_29", "pid": 870, "tid": 870,
    "ts": 7429770633185.542, "dur": 1085.196,
    "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": 870, "tid": 870,
    "ts": 7429770633186.947, "dur": 1083.007,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770633187.181, "dur": 0.293,
    "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": 870, "tid": 870,
    "ts": 7429770633188.851, "dur": 1080.161,
    "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": 870, "tid": 870,
    "ts": 7429770633189.227, "dur": 0.628,
    "args": {
      "Python parent id": 4660, "Python id": 4661, "finished": true, "Ev Idx": 11491
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_58", "pid": 870, "tid": 870,
    "ts": 7429770633192.663, "dur": 69.370,
    "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": 870, "tid": 870,
    "ts": 7429770633193.169, "dur": 68.494,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770633193.296, "dur": 0.161,
    "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": 870, "tid": 870,
    "ts": 7429770633194.399, "dur": 66.433,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770633194.869, "dur": 1.446,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770633196.579, "dur": 10.995,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770633207.951, "dur": 13.116,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770633232.015, "dur": 9.015,
    "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": 870, "tid": 870,
    "ts": 7429770633250.296, "dur": 0.239,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770633250.934, "dur": 1.065,
    "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": 870, "tid": 870,
    "ts": 7429770633262.624, "dur": 0.340,
    "args": {
      "Python parent id": 4660, "Python id": 4672, "finished": true, "Ev Idx": 11502
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_29", "pid": 870, "tid": 870,
    "ts": 7429770633265.672, "dur": 693.158,
    "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": 870, "tid": 870,
    "ts": 7429770633267.252, "dur": 690.909,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770633267.469, "dur": 0.313,
    "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": 870, "tid": 870,
    "ts": 7429770633269.634, "dur": 687.753,
    "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": 870, "tid": 870,
    "ts": 7429770633272.620, "dur": 0.307,
    "args": {
      "Python parent id": 4676, "Python id": 4677, "finished": true, "Ev Idx": 11507
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_203", "pid": 870, "tid": 870,
    "ts": 7429770633274.858, "dur": 39.788,
    "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": 870, "tid": 870,
    "ts": 7429770633275.394, "dur": 38.985,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770633275.546, "dur": 0.152,
    "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": 870, "tid": 870,
    "ts": 7429770633276.415, "dur": 37.797,
    "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": 870, "tid": 870,
    "ts": 7429770633276.770, "dur": 0.136,
    "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": 870, "tid": 870,
    "ts": 7429770633277.135, "dur": 0.113,
    "args": {
      "Python parent id": 4681, "Python id": 4683, "finished": true, "Ev Idx": 11513
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770633277.384, "dur": 36.761,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770633315.663, "dur": 3.252,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770633319.450, "dur": 3.242,
    "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": 870, "tid": 870,
    "ts": 7429770633323.486, "dur": 0.297,
    "args": {
      "Python parent id": 4676, "Python id": 4687, "finished": true, "Ev Idx": 11517
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_204", "pid": 870, "tid": 870,
    "ts": 7429770633325.706, "dur": 36.788,
    "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": 870, "tid": 870,
    "ts": 7429770633326.200, "dur": 35.994,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770633326.374, "dur": 0.291,
    "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": 870, "tid": 870,
    "ts": 7429770633327.555, "dur": 34.540,
    "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": 870, "tid": 870,
    "ts": 7429770633327.803, "dur": 0.152,
    "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": 870, "tid": 870,
    "ts": 7429770633328.181, "dur": 0.132,
    "args": {
      "Python parent id": 4691, "Python id": 4693, "finished": true, "Ev Idx": 11523
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770633328.415, "dur": 33.633,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770633363.060, "dur": 2.399,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770633365.904, "dur": 2.434,
    "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": 870, "tid": 870,
    "ts": 7429770633369.100, "dur": 0.327,
    "args": {
      "Python parent id": 4676, "Python id": 4697, "finished": true, "Ev Idx": 11527
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_205", "pid": 870, "tid": 870,
    "ts": 7429770633371.293, "dur": 33.629,
    "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": 870, "tid": 870,
    "ts": 7429770633371.796, "dur": 32.847,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770633371.974, "dur": 0.223,
    "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": 870, "tid": 870,
    "ts": 7429770633373.113, "dur": 31.363,
    "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": 870, "tid": 870,
    "ts": 7429770633373.349, "dur": 0.140,
    "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": 870, "tid": 870,
    "ts": 7429770633373.690, "dur": 0.117,
    "args": {
      "Python parent id": 4701, "Python id": 4703, "finished": true, "Ev Idx": 11533
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770633373.917, "dur": 30.518,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770633405.512, "dur": 2.273,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770633408.234, "dur": 2.357,
    "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": 870, "tid": 870,
    "ts": 7429770633411.676, "dur": 148.382,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770633412.248, "dur": 3.456,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770633415.994, "dur": 1.635,
    "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": 870, "tid": 870,
    "ts": 7429770633428.623, "dur": 47.511,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770633454.215, "dur": 20.837,
    "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": 870, "tid": 870,
    "ts": 7429770633510.481, "dur": 29.177,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770633526.962, "dur": 11.782,
    "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": 870, "tid": 870,
    "ts": 7429770633564.151, "dur": 3.593,
    "args": {
      "Python parent id": 4676, "Python id": 4714, "finished": true, "Ev Idx": 11544
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770633564.798, "dur": 0.212,
    "args": {
      "Python parent id": 4714, "Python id": 4715, "finished": true, "Ev Idx": 11545
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770633565.358, "dur": 0.085,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770633566.698, "dur": 0.282,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770633567.380, "dur": 0.098,
    "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": 870, "tid": 870,
    "ts": 7429770633568.932, "dur": 6.356,
    "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": 870, "tid": 870,
    "ts": 7429770633572.070, "dur": 3.093,
    "args": {
      "Python parent id": 4719, "Python id": 4720, "finished": true, "Ev Idx": 11550
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770633572.336, "dur": 1.240,
    "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": 870, "tid": 870,
    "ts": 7429770633572.752, "dur": 0.692,
    "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": 870, "tid": 870,
    "ts": 7429770633573.846, "dur": 0.457,
    "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": 870, "tid": 870,
    "ts": 7429770633574.621, "dur": 0.459,
    "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": 870, "tid": 870,
    "ts": 7429770633577.786, "dur": 319.624,
    "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": 870, "tid": 870,
    "ts": 7429770633578.235, "dur": 43.283,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770633595.812, "dur": 5.028,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770633601.636, "dur": 19.771,
    "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": 870, "tid": 870,
    "ts": 7429770633622.366, "dur": 30.035,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770633636.438, "dur": 2.608,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770633639.561, "dur": 12.751,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770633653.262, "dur": 2.833,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770633656.402, "dur": 109.248,
    "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": 870, "tid": 870,
    "ts": 7429770633801.887, "dur": 17.581,
    "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": 870, "tid": 870,
    "ts": 7429770633802.159, "dur": 0.140,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770633803.055, "dur": 16.207,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770633820.385, "dur": 1.728,
    "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": 870, "tid": 870,
    "ts": 7429770633823.506, "dur": 10.443,
    "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": 870, "tid": 870,
    "ts": 7429770633823.672, "dur": 0.114,
    "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": 870, "tid": 870,
    "ts": 7429770633829.569, "dur": 1.131,
    "args": {
      "Python parent id": 4738, "Python id": 4740, "finished": true, "Ev Idx": 11570
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770633830.153, "dur": 0.447,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770633831.177, "dur": 2.693,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770633834.597, "dur": 43.468,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770633878.790, "dur": 3.201,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770633882.503, "dur": 13.912,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770633900.658, "dur": 3.353,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770633904.808, "dur": 0.276,
    "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": 870, "tid": 870,
    "ts": 7429770633906.060, "dur": 1.145,
    "args": {
      "Python parent id": 4676, "Python id": 4748, "finished": true, "Ev Idx": 11578
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_206", "pid": 870, "tid": 870,
    "ts": 7429770633912.419, "dur": 44.391,
    "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": 870, "tid": 870,
    "ts": 7429770633913.453, "dur": 42.928,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770633913.912, "dur": 1.385,
    "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": 870, "tid": 870,
    "ts": 7429770633916.457, "dur": 39.695,
    "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": 870, "tid": 870,
    "ts": 7429770633916.882, "dur": 0.301,
    "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": 870, "tid": 870,
    "ts": 7429770633917.447, "dur": 0.142,
    "args": {
      "Python parent id": 4752, "Python id": 4754, "finished": true, "Ev Idx": 11584
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770633917.750, "dur": 38.308,
    "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": 870, "tid": 870,
    "ts": 7429770633970.732, "dur": 0.475,
    "args": {
      "Python parent id": 4660, "Python id": 4756, "finished": true, "Ev Idx": 11586
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_59", "pid": 870, "tid": 870,
    "ts": 7429770633973.773, "dur": 84.258,
    "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": 870, "tid": 870,
    "ts": 7429770633974.419, "dur": 83.230,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770633974.594, "dur": 0.372,
    "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": 870, "tid": 870,
    "ts": 7429770633976.632, "dur": 80.219,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770633977.229, "dur": 1.546,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770633979.054, "dur": 14.596,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770633994.013, "dur": 15.679,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770634022.517, "dur": 11.718,
    "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": 870, "tid": 870,
    "ts": 7429770634045.380, "dur": 0.246,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770634046.026, "dur": 1.063,
    "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": 870, "tid": 870,
    "ts": 7429770634058.807, "dur": 0.397,
    "args": {
      "Python parent id": 4660, "Python id": 4767, "finished": true, "Ev Idx": 11597
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_29", "pid": 870, "tid": 870,
    "ts": 7429770634060.996, "dur": 195.845,
    "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": 870, "tid": 870,
    "ts": 7429770634061.465, "dur": 195.176,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770634061.699, "dur": 0.312,
    "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": 870, "tid": 870,
    "ts": 7429770634062.867, "dur": 193.637,
    "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": 870, "tid": 870,
    "ts": 7429770634063.315, "dur": 0.640,
    "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": 870, "tid": 870,
    "ts": 7429770634064.256, "dur": 0.304,
    "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": 870, "tid": 870,
    "ts": 7429770634064.827, "dur": 0.254,
    "args": {
      "Python parent id": 4771, "Python id": 4774, "finished": true, "Ev Idx": 11604
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_207", "pid": 870, "tid": 870,
    "ts": 7429770634067.148, "dur": 66.312,
    "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": 870, "tid": 870,
    "ts": 7429770634067.683, "dur": 65.407,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770634067.955, "dur": 0.152,
    "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": 870, "tid": 870,
    "ts": 7429770634068.903, "dur": 63.914,
    "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": 870, "tid": 870,
    "ts": 7429770634069.160, "dur": 17.243,
    "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": 870, "tid": 870,
    "ts": 7429770634086.877, "dur": 0.182,
    "args": {
      "Python parent id": 4778, "Python id": 4780, "finished": true, "Ev Idx": 11610
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770634087.394, "dur": 45.358,
    "args": {
      "Python parent id": 4778, "Python id": 4781, "finished": true, "Ev Idx": 11611
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_29", "pid": 870, "tid": 870,
    "ts": 7429770634135.539, "dur": 18.397,
    "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": 870, "tid": 870,
    "ts": 7429770634136.216, "dur": 17.440,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770634136.457, "dur": 0.457,
    "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": 870, "tid": 870,
    "ts": 7429770634138.107, "dur": 15.417,
    "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": 870, "tid": 870,
    "ts": 7429770634138.979, "dur": 14.342,
    "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": 870, "tid": 870,
    "ts": 7429770634139.168, "dur": 0.192,
    "args": {
      "Python parent id": 4786, "Python id": 4787, "finished": true, "Ev Idx": 11617
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770634140.303, "dur": 12.921,
    "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": 870, "tid": 870,
    "ts": 7429770634154.674, "dur": 0.390,
    "args": {
      "Python parent id": 4771, "Python id": 4789, "finished": true, "Ev Idx": 11619
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_208", "pid": 870, "tid": 870,
    "ts": 7429770634157.504, "dur": 36.947,
    "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": 870, "tid": 870,
    "ts": 7429770634157.996, "dur": 36.143,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770634158.153, "dur": 0.360,
    "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": 870, "tid": 870,
    "ts": 7429770634159.395, "dur": 34.649,
    "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": 870, "tid": 870,
    "ts": 7429770634159.635, "dur": 0.165,
    "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": 870, "tid": 870,
    "ts": 7429770634159.965, "dur": 0.127,
    "args": {
      "Python parent id": 4793, "Python id": 4795, "finished": true, "Ev Idx": 11625
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770634160.197, "dur": 33.790,
    "args": {
      "Python parent id": 4793, "Python id": 4796, "finished": true, "Ev Idx": 11626
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_209", "pid": 870, "tid": 870,
    "ts": 7429770634206.945, "dur": 49.239,
    "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": 870, "tid": 870,
    "ts": 7429770634207.732, "dur": 48.143,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770634207.916, "dur": 0.230,
    "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": 870, "tid": 870,
    "ts": 7429770634208.870, "dur": 46.884,
    "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": 870, "tid": 870,
    "ts": 7429770634209.073, "dur": 0.190,
    "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": 870, "tid": 870,
    "ts": 7429770634209.467, "dur": 0.130,
    "args": {
      "Python parent id": 4800, "Python id": 4802, "finished": true, "Ev Idx": 11632
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770634209.701, "dur": 46.005,
    "args": {
      "Python parent id": 4800, "Python id": 4803, "finished": true, "Ev Idx": 11633
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_30", "pid": 870, "tid": 870,
    "ts": 7429770634276.858, "dur": 1175.512,
    "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": 870, "tid": 870,
    "ts": 7429770634278.292, "dur": 1173.201,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770634278.486, "dur": 0.288,
    "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": 870, "tid": 870,
    "ts": 7429770634280.162, "dur": 1170.198,
    "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": 870, "tid": 870,
    "ts": 7429770634280.574, "dur": 0.629,
    "args": {
      "Python parent id": 4807, "Python id": 4808, "finished": true, "Ev Idx": 11638
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_60", "pid": 870, "tid": 870,
    "ts": 7429770634283.159, "dur": 75.258,
    "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": 870, "tid": 870,
    "ts": 7429770634283.629, "dur": 74.468,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770634283.764, "dur": 0.191,
    "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": 870, "tid": 870,
    "ts": 7429770634284.787, "dur": 72.497,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770634285.342, "dur": 1.634,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770634287.220, "dur": 12.391,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770634299.964, "dur": 14.623,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770634326.314, "dur": 9.777,
    "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": 870, "tid": 870,
    "ts": 7429770634346.286, "dur": 0.243,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770634346.931, "dur": 1.024,
    "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": 870, "tid": 870,
    "ts": 7429770634359.089, "dur": 0.326,
    "args": {
      "Python parent id": 4807, "Python id": 4819, "finished": true, "Ev Idx": 11649
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_30", "pid": 870, "tid": 870,
    "ts": 7429770634362.324, "dur": 696.054,
    "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": 870, "tid": 870,
    "ts": 7429770634363.605, "dur": 694.224,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770634363.835, "dur": 0.277,
    "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": 870, "tid": 870,
    "ts": 7429770634366.047, "dur": 690.923,
    "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": 870, "tid": 870,
    "ts": 7429770634369.164, "dur": 0.411,
    "args": {
      "Python parent id": 4823, "Python id": 4824, "finished": true, "Ev Idx": 11654
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_210", "pid": 870, "tid": 870,
    "ts": 7429770634371.518, "dur": 40.289,
    "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": 870, "tid": 870,
    "ts": 7429770634371.964, "dur": 39.525,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770634372.120, "dur": 0.176,
    "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": 870, "tid": 870,
    "ts": 7429770634373.390, "dur": 37.976,
    "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": 870, "tid": 870,
    "ts": 7429770634373.735, "dur": 0.149,
    "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": 870, "tid": 870,
    "ts": 7429770634374.144, "dur": 0.120,
    "args": {
      "Python parent id": 4828, "Python id": 4830, "finished": true, "Ev Idx": 11660
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770634374.387, "dur": 36.900,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770634412.702, "dur": 3.440,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770634416.521, "dur": 3.231,
    "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": 870, "tid": 870,
    "ts": 7429770634420.420, "dur": 0.282,
    "args": {
      "Python parent id": 4823, "Python id": 4834, "finished": true, "Ev Idx": 11664
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_211", "pid": 870, "tid": 870,
    "ts": 7429770634422.486, "dur": 35.178,
    "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": 870, "tid": 870,
    "ts": 7429770634422.967, "dur": 34.401,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770634423.324, "dur": 0.282,
    "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": 870, "tid": 870,
    "ts": 7429770634424.498, "dur": 32.748,
    "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": 870, "tid": 870,
    "ts": 7429770634424.723, "dur": 0.122,
    "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": 870, "tid": 870,
    "ts": 7429770634425.044, "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": 870, "tid": 870,
    "ts": 7429770634425.295, "dur": 31.911,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770634458.233, "dur": 2.272,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770634460.878, "dur": 2.488,
    "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": 870, "tid": 870,
    "ts": 7429770634464.047, "dur": 0.335,
    "args": {
      "Python parent id": 4823, "Python id": 4844, "finished": true, "Ev Idx": 11674
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_212", "pid": 870, "tid": 870,
    "ts": 7429770634466.038, "dur": 33.655,
    "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": 870, "tid": 870,
    "ts": 7429770634466.543, "dur": 32.822,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770634466.718, "dur": 0.211,
    "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": 870, "tid": 870,
    "ts": 7429770634467.737, "dur": 31.512,
    "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": 870, "tid": 870,
    "ts": 7429770634467.953, "dur": 0.156,
    "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": 870, "tid": 870,
    "ts": 7429770634468.277, "dur": 0.119,
    "args": {
      "Python parent id": 4848, "Python id": 4850, "finished": true, "Ev Idx": 11680
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770634468.503, "dur": 30.698,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770634500.413, "dur": 2.372,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770634503.189, "dur": 2.411,
    "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": 870, "tid": 870,
    "ts": 7429770634506.595, "dur": 140.605,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770634507.144, "dur": 3.783,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770634511.168, "dur": 2.014,
    "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": 870, "tid": 870,
    "ts": 7429770634523.653, "dur": 48.730,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770634549.030, "dur": 22.101,
    "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": 870, "tid": 870,
    "ts": 7429770634601.621, "dur": 27.800,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770634617.545, "dur": 10.945,
    "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": 870, "tid": 870,
    "ts": 7429770634651.202, "dur": 3.799,
    "args": {
      "Python parent id": 4823, "Python id": 4861, "finished": true, "Ev Idx": 11691
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770634652.002, "dur": 0.263,
    "args": {
      "Python parent id": 4861, "Python id": 4862, "finished": true, "Ev Idx": 11692
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770634652.564, "dur": 0.077,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770634653.932, "dur": 0.263,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770634654.602, "dur": 0.104,
    "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": 870, "tid": 870,
    "ts": 7429770634656.172, "dur": 6.381,
    "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": 870, "tid": 870,
    "ts": 7429770634659.235, "dur": 3.155,
    "args": {
      "Python parent id": 4866, "Python id": 4867, "finished": true, "Ev Idx": 11697
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770634659.567, "dur": 1.381,
    "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": 870, "tid": 870,
    "ts": 7429770634660.020, "dur": 0.802,
    "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": 870, "tid": 870,
    "ts": 7429770634661.197, "dur": 0.431,
    "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": 870, "tid": 870,
    "ts": 7429770634661.887, "dur": 0.432,
    "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": 870, "tid": 870,
    "ts": 7429770634665.063, "dur": 327.168,
    "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": 870, "tid": 870,
    "ts": 7429770634665.511, "dur": 44.843,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770634683.214, "dur": 4.697,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770634688.702, "dur": 21.522,
    "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": 870, "tid": 870,
    "ts": 7429770634711.119, "dur": 29.017,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770634724.921, "dur": 2.830,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770634728.238, "dur": 11.829,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770634740.892, "dur": 3.416,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770634744.647, "dur": 111.880,
    "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": 870, "tid": 870,
    "ts": 7429770634892.897, "dur": 18.417,
    "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": 870, "tid": 870,
    "ts": 7429770634893.221, "dur": 0.179,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770634894.044, "dur": 17.038,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770634912.404, "dur": 1.943,
    "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": 870, "tid": 870,
    "ts": 7429770634915.779, "dur": 10.830,
    "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": 870, "tid": 870,
    "ts": 7429770634915.890, "dur": 0.125,
    "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": 870, "tid": 870,
    "ts": 7429770634922.144, "dur": 1.087,
    "args": {
      "Python parent id": 4885, "Python id": 4887, "finished": true, "Ev Idx": 11717
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770634922.633, "dur": 0.490,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770634923.741, "dur": 2.773,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770634927.408, "dur": 44.009,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770634972.151, "dur": 3.599,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770634976.243, "dur": 14.965,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770634995.694, "dur": 3.403,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770634999.844, "dur": 0.263,
    "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": 870, "tid": 870,
    "ts": 7429770635001.156, "dur": 1.219,
    "args": {
      "Python parent id": 4823, "Python id": 4895, "finished": true, "Ev Idx": 11725
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_213", "pid": 870, "tid": 870,
    "ts": 7429770635007.361, "dur": 48.858,
    "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": 870, "tid": 870,
    "ts": 7429770635008.715, "dur": 46.988,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770635009.270, "dur": 1.351,
    "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": 870, "tid": 870,
    "ts": 7429770635012.165, "dur": 43.245,
    "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": 870, "tid": 870,
    "ts": 7429770635012.590, "dur": 0.283,
    "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": 870, "tid": 870,
    "ts": 7429770635013.199, "dur": 0.202,
    "args": {
      "Python parent id": 4899, "Python id": 4901, "finished": true, "Ev Idx": 11731
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770635013.641, "dur": 41.644,
    "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": 870, "tid": 870,
    "ts": 7429770635070.845, "dur": 0.521,
    "args": {
      "Python parent id": 4807, "Python id": 4903, "finished": true, "Ev Idx": 11733
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_61", "pid": 870, "tid": 870,
    "ts": 7429770635074.344, "dur": 85.827,
    "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": 870, "tid": 870,
    "ts": 7429770635075.203, "dur": 84.530,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770635075.493, "dur": 0.454,
    "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": 870, "tid": 870,
    "ts": 7429770635076.962, "dur": 81.933,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770635077.569, "dur": 1.657,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770635079.626, "dur": 14.843,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770635094.820, "dur": 15.720,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770635124.211, "dur": 10.820,
    "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": 870, "tid": 870,
    "ts": 7429770635147.449, "dur": 0.316,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770635148.172, "dur": 1.199,
    "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": 870, "tid": 870,
    "ts": 7429770635161.111, "dur": 0.385,
    "args": {
      "Python parent id": 4807, "Python id": 4914, "finished": true, "Ev Idx": 11744
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_30", "pid": 870, "tid": 870,
    "ts": 7429770635163.507, "dur": 276.599,
    "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": 870, "tid": 870,
    "ts": 7429770635164.143, "dur": 275.683,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770635164.425, "dur": 0.369,
    "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": 870, "tid": 870,
    "ts": 7429770635165.644, "dur": 274.026,
    "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": 870, "tid": 870,
    "ts": 7429770635166.076, "dur": 0.353,
    "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": 870, "tid": 870,
    "ts": 7429770635166.817, "dur": 0.342,
    "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": 870, "tid": 870,
    "ts": 7429770635167.449, "dur": 0.224,
    "args": {
      "Python parent id": 4918, "Python id": 4921, "finished": true, "Ev Idx": 11751
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_214", "pid": 870, "tid": 870,
    "ts": 7429770635169.500, "dur": 142.549,
    "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": 870, "tid": 870,
    "ts": 7429770635170.463, "dur": 141.228,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770635170.613, "dur": 0.166,
    "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": 870, "tid": 870,
    "ts": 7429770635171.422, "dur": 140.033,
    "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": 870, "tid": 870,
    "ts": 7429770635171.748, "dur": 0.118,
    "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": 870, "tid": 870,
    "ts": 7429770635172.111, "dur": 0.130,
    "args": {
      "Python parent id": 4925, "Python id": 4927, "finished": true, "Ev Idx": 11757
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770635172.422, "dur": 138.914,
    "args": {
      "Python parent id": 4925, "Python id": 4928, "finished": true, "Ev Idx": 11758
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_30", "pid": 870, "tid": 870,
    "ts": 7429770635314.692, "dur": 18.271,
    "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": 870, "tid": 870,
    "ts": 7429770635315.296, "dur": 17.459,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770635315.666, "dur": 0.570,
    "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": 870, "tid": 870,
    "ts": 7429770635317.236, "dur": 15.426,
    "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": 870, "tid": 870,
    "ts": 7429770635318.185, "dur": 14.276,
    "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": 870, "tid": 870,
    "ts": 7429770635318.320, "dur": 0.246,
    "args": {
      "Python parent id": 4933, "Python id": 4934, "finished": true, "Ev Idx": 11764
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770635319.387, "dur": 12.986,
    "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": 870, "tid": 870,
    "ts": 7429770635333.868, "dur": 0.462,
    "args": {
      "Python parent id": 4918, "Python id": 4936, "finished": true, "Ev Idx": 11766
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_215", "pid": 870, "tid": 870,
    "ts": 7429770635336.265, "dur": 41.191,
    "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": 870, "tid": 870,
    "ts": 7429770635336.791, "dur": 40.425,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770635336.951, "dur": 0.191,
    "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": 870, "tid": 870,
    "ts": 7429770635337.726, "dur": 39.400,
    "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": 870, "tid": 870,
    "ts": 7429770635337.970, "dur": 0.114,
    "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": 870, "tid": 870,
    "ts": 7429770635338.304, "dur": 0.129,
    "args": {
      "Python parent id": 4940, "Python id": 4942, "finished": true, "Ev Idx": 11772
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770635338.583, "dur": 38.450,
    "args": {
      "Python parent id": 4940, "Python id": 4943, "finished": true, "Ev Idx": 11773
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_216", "pid": 870, "tid": 870,
    "ts": 7429770635389.828, "dur": 49.572,
    "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": 870, "tid": 870,
    "ts": 7429770635390.424, "dur": 48.727,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770635390.627, "dur": 0.261,
    "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": 870, "tid": 870,
    "ts": 7429770635391.510, "dur": 47.500,
    "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": 870, "tid": 870,
    "ts": 7429770635391.766, "dur": 0.140,
    "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": 870, "tid": 870,
    "ts": 7429770635392.118, "dur": 0.144,
    "args": {
      "Python parent id": 4947, "Python id": 4949, "finished": true, "Ev Idx": 11779
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770635392.399, "dur": 46.521,
    "args": {
      "Python parent id": 4947, "Python id": 4950, "finished": true, "Ev Idx": 11780
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaDecoderLayer_31", "pid": 870, "tid": 870,
    "ts": 7429770635459.192, "dur": 1096.496,
    "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": 870, "tid": 870,
    "ts": 7429770635460.631, "dur": 1094.310,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770635460.904, "dur": 0.324,
    "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": 870, "tid": 870,
    "ts": 7429770635462.631, "dur": 1091.396,
    "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": 870, "tid": 870,
    "ts": 7429770635462.970, "dur": 0.566,
    "args": {
      "Python parent id": 4954, "Python id": 4955, "finished": true, "Ev Idx": 11785
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_62", "pid": 870, "tid": 870,
    "ts": 7429770635465.347, "dur": 76.256,
    "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": 870, "tid": 870,
    "ts": 7429770635465.863, "dur": 75.478,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770635465.996, "dur": 0.159,
    "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": 870, "tid": 870,
    "ts": 7429770635467.001, "dur": 73.520,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770635467.580, "dur": 1.928,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770635469.803, "dur": 12.267,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770635482.378, "dur": 15.555,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770635509.737, "dur": 9.587,
    "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": 870, "tid": 870,
    "ts": 7429770635529.634, "dur": 0.283,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770635530.287, "dur": 1.168,
    "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": 870, "tid": 870,
    "ts": 7429770635542.289, "dur": 0.329,
    "args": {
      "Python parent id": 4954, "Python id": 4966, "finished": true, "Ev Idx": 11796
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaAttention_31", "pid": 870, "tid": 870,
    "ts": 7429770635545.513, "dur": 724.154,
    "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": 870, "tid": 870,
    "ts": 7429770635546.897, "dur": 722.026,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770635547.107, "dur": 0.248,
    "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": 870, "tid": 870,
    "ts": 7429770635549.523, "dur": 718.413,
    "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": 870, "tid": 870,
    "ts": 7429770635552.713, "dur": 0.384,
    "args": {
      "Python parent id": 4970, "Python id": 4971, "finished": true, "Ev Idx": 11801
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_217", "pid": 870, "tid": 870,
    "ts": 7429770635555.194, "dur": 41.342,
    "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": 870, "tid": 870,
    "ts": 7429770635555.731, "dur": 40.517,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770635555.872, "dur": 0.190,
    "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": 870, "tid": 870,
    "ts": 7429770635556.841, "dur": 39.277,
    "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": 870, "tid": 870,
    "ts": 7429770635557.147, "dur": 0.130,
    "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": 870, "tid": 870,
    "ts": 7429770635557.484, "dur": 0.128,
    "args": {
      "Python parent id": 4975, "Python id": 4977, "finished": true, "Ev Idx": 11807
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770635557.736, "dur": 38.335,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770635597.341, "dur": 3.432,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770635601.310, "dur": 3.637,
    "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": 870, "tid": 870,
    "ts": 7429770635605.835, "dur": 0.345,
    "args": {
      "Python parent id": 4970, "Python id": 4981, "finished": true, "Ev Idx": 11811
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_218", "pid": 870, "tid": 870,
    "ts": 7429770635608.039, "dur": 35.753,
    "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": 870, "tid": 870,
    "ts": 7429770635608.534, "dur": 35.052,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770635608.702, "dur": 0.373,
    "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": 870, "tid": 870,
    "ts": 7429770635609.962, "dur": 33.492,
    "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": 870, "tid": 870,
    "ts": 7429770635610.177, "dur": 0.122,
    "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": 870, "tid": 870,
    "ts": 7429770635610.478, "dur": 0.133,
    "args": {
      "Python parent id": 4985, "Python id": 4987, "finished": true, "Ev Idx": 11817
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770635610.714, "dur": 32.692,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770635644.243, "dur": 2.528,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770635647.256, "dur": 2.467,
    "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": 870, "tid": 870,
    "ts": 7429770635650.572, "dur": 0.338,
    "args": {
      "Python parent id": 4970, "Python id": 4991, "finished": true, "Ev Idx": 11821
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_219", "pid": 870, "tid": 870,
    "ts": 7429770635653.436, "dur": 34.489,
    "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": 870, "tid": 870,
    "ts": 7429770635654.002, "dur": 33.674,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770635654.159, "dur": 0.224,
    "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": 870, "tid": 870,
    "ts": 7429770635655.060, "dur": 32.502,
    "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": 870, "tid": 870,
    "ts": 7429770635655.248, "dur": 0.120,
    "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": 870, "tid": 870,
    "ts": 7429770635655.535, "dur": 0.119,
    "args": {
      "Python parent id": 4995, "Python id": 4997, "finished": true, "Ev Idx": 11827
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770635655.969, "dur": 31.531,
    "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 0x7f8f8b4d3e30>", "pid": 870, "tid": 870,
    "ts": 7429770635688.570, "dur": 2.347,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770635691.320, "dur": 2.609,
    "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": 870, "tid": 870,
    "ts": 7429770635695.093, "dur": 174.884,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770635695.712, "dur": 3.787,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770635699.825, "dur": 2.217,
    "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": 870, "tid": 870,
    "ts": 7429770635714.258, "dur": 52.850,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770635741.915, "dur": 24.098,
    "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": 870, "tid": 870,
    "ts": 7429770635824.024, "dur": 27.796,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770635839.956, "dur": 11.006,
    "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": 870, "tid": 870,
    "ts": 7429770635874.311, "dur": 4.094,
    "args": {
      "Python parent id": 4970, "Python id": 5008, "finished": true, "Ev Idx": 11838
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770635875.092, "dur": 0.410,
    "args": {
      "Python parent id": 5008, "Python id": 5009, "finished": true, "Ev Idx": 11839
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770635875.927, "dur": 0.051,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770635877.336, "dur": 0.283,
    "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 0x7f9848dd4d00>", "pid": 870, "tid": 870,
    "ts": 7429770635877.999, "dur": 0.097,
    "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": 870, "tid": 870,
    "ts": 7429770635879.790, "dur": 7.155,
    "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": 870, "tid": 870,
    "ts": 7429770635883.472, "dur": 3.364,
    "args": {
      "Python parent id": 5013, "Python id": 5014, "finished": true, "Ev Idx": 11844
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770635883.795, "dur": 1.443,
    "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": 870, "tid": 870,
    "ts": 7429770635884.263, "dur": 0.776,
    "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": 870, "tid": 870,
    "ts": 7429770635885.559, "dur": 0.465,
    "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": 870, "tid": 870,
    "ts": 7429770635886.379, "dur": 0.382,
    "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": 870, "tid": 870,
    "ts": 7429770635889.619, "dur": 319.897,
    "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": 870, "tid": 870,
    "ts": 7429770635890.091, "dur": 46.280,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770635908.692, "dur": 5.349,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770635914.937, "dur": 21.308,
    "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": 870, "tid": 870,
    "ts": 7429770635937.145, "dur": 30.236,
    "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 0x7f8f8b5c33e0>", "pid": 870, "tid": 870,
    "ts": 7429770635950.677, "dur": 2.806,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770635953.990, "dur": 13.329,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770635968.277, "dur": 2.966,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770635971.508, "dur": 105.898,
    "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": 870, "tid": 870,
    "ts": 7429770636113.026, "dur": 17.866,
    "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": 870, "tid": 870,
    "ts": 7429770636113.197, "dur": 0.252,
    "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 0x7f8f8b4d38e0>", "pid": 870, "tid": 870,
    "ts": 7429770636114.098, "dur": 16.556,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770636131.823, "dur": 1.915,
    "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": 870, "tid": 870,
    "ts": 7429770636135.174, "dur": 10.198,
    "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": 870, "tid": 870,
    "ts": 7429770636135.327, "dur": 0.120,
    "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": 870, "tid": 870,
    "ts": 7429770636141.104, "dur": 1.091,
    "args": {
      "Python parent id": 5032, "Python id": 5034, "finished": true, "Ev Idx": 11864
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770636141.671, "dur": 0.435,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770636142.676, "dur": 2.618,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770636146.037, "dur": 43.691,
    "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 0x7f8f8b4d30c0>", "pid": 870, "tid": 870,
    "ts": 7429770636190.410, "dur": 3.267,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770636194.363, "dur": 14.010,
    "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 0x7f8f8b4f6b70>", "pid": 870, "tid": 870,
    "ts": 7429770636212.541, "dur": 3.143,
    "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 0x7f8f8b4d2530>", "pid": 870, "tid": 870,
    "ts": 7429770636216.372, "dur": 0.239,
    "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": 870, "tid": 870,
    "ts": 7429770636217.678, "dur": 1.048,
    "args": {
      "Python parent id": 4970, "Python id": 5042, "finished": true, "Ev Idx": 11872
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_220", "pid": 870, "tid": 870,
    "ts": 7429770636223.384, "dur": 43.904,
    "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": 870, "tid": 870,
    "ts": 7429770636224.902, "dur": 41.953,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770636225.577, "dur": 1.205,
    "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": 870, "tid": 870,
    "ts": 7429770636228.396, "dur": 38.151,
    "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": 870, "tid": 870,
    "ts": 7429770636228.719, "dur": 0.215,
    "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": 870, "tid": 870,
    "ts": 7429770636229.192, "dur": 0.143,
    "args": {
      "Python parent id": 5046, "Python id": 5048, "finished": true, "Ev Idx": 11878
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770636229.611, "dur": 36.860,
    "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": 870, "tid": 870,
    "ts": 7429770636282.682, "dur": 0.488,
    "args": {
      "Python parent id": 4954, "Python id": 5050, "finished": true, "Ev Idx": 11880
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_63", "pid": 870, "tid": 870,
    "ts": 7429770636286.048, "dur": 84.677,
    "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": 870, "tid": 870,
    "ts": 7429770636286.769, "dur": 83.509,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770636287.029, "dur": 0.362,
    "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": 870, "tid": 870,
    "ts": 7429770636288.520, "dur": 80.912,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770636289.092, "dur": 1.628,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770636291.093, "dur": 13.931,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770636305.427, "dur": 15.914,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770636335.104, "dur": 11.325,
    "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": 870, "tid": 870,
    "ts": 7429770636358.624, "dur": 0.317,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770636359.339, "dur": 1.090,
    "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": 870, "tid": 870,
    "ts": 7429770636371.422, "dur": 0.321,
    "args": {
      "Python parent id": 4954, "Python id": 5061, "finished": true, "Ev Idx": 11891
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaMLP_31", "pid": 870, "tid": 870,
    "ts": 7429770636373.668, "dur": 170.529,
    "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": 870, "tid": 870,
    "ts": 7429770636374.268, "dur": 169.644,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770636374.453, "dur": 0.292,
    "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": 870, "tid": 870,
    "ts": 7429770636375.652, "dur": 168.159,
    "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": 870, "tid": 870,
    "ts": 7429770636376.138, "dur": 0.367,
    "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": 870, "tid": 870,
    "ts": 7429770636376.883, "dur": 0.323,
    "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": 870, "tid": 870,
    "ts": 7429770636377.482, "dur": 0.256,
    "args": {
      "Python parent id": 5065, "Python id": 5068, "finished": true, "Ev Idx": 11898
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_221", "pid": 870, "tid": 870,
    "ts": 7429770636379.818, "dur": 40.318,
    "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": 870, "tid": 870,
    "ts": 7429770636380.403, "dur": 39.452,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770636380.545, "dur": 0.174,
    "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": 870, "tid": 870,
    "ts": 7429770636381.387, "dur": 38.348,
    "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": 870, "tid": 870,
    "ts": 7429770636381.678, "dur": 0.168,
    "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": 870, "tid": 870,
    "ts": 7429770636382.128, "dur": 0.135,
    "args": {
      "Python parent id": 5072, "Python id": 5074, "finished": true, "Ev Idx": 11904
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770636382.407, "dur": 37.265,
    "args": {
      "Python parent id": 5072, "Python id": 5075, "finished": true, "Ev Idx": 11905
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: SiLU_31", "pid": 870, "tid": 870,
    "ts": 7429770636422.019, "dur": 16.290,
    "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": 870, "tid": 870,
    "ts": 7429770636422.532, "dur": 15.505,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770636422.709, "dur": 0.208,
    "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": 870, "tid": 870,
    "ts": 7429770636423.800, "dur": 14.163,
    "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": 870, "tid": 870,
    "ts": 7429770636424.740, "dur": 13.030,
    "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": 870, "tid": 870,
    "ts": 7429770636424.864, "dur": 0.195,
    "args": {
      "Python parent id": 5080, "Python id": 5081, "finished": true, "Ev Idx": 11911
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function silu>", "pid": 870, "tid": 870,
    "ts": 7429770636425.707, "dur": 11.980,
    "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": 870, "tid": 870,
    "ts": 7429770636438.997, "dur": 0.366,
    "args": {
      "Python parent id": 5065, "Python id": 5083, "finished": true, "Ev Idx": 11913
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_222", "pid": 870, "tid": 870,
    "ts": 7429770636441.377, "dur": 39.351,
    "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": 870, "tid": 870,
    "ts": 7429770636441.786, "dur": 38.687,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770636441.942, "dur": 0.193,
    "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": 870, "tid": 870,
    "ts": 7429770636442.913, "dur": 37.443,
    "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": 870, "tid": 870,
    "ts": 7429770636443.142, "dur": 0.143,
    "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": 870, "tid": 870,
    "ts": 7429770636443.477, "dur": 0.150,
    "args": {
      "Python parent id": 5087, "Python id": 5089, "finished": true, "Ev Idx": 11919
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770636443.891, "dur": 36.409,
    "args": {
      "Python parent id": 5087, "Python id": 5090, "finished": true, "Ev Idx": 11920
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_223", "pid": 870, "tid": 870,
    "ts": 7429770636492.297, "dur": 51.182,
    "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": 870, "tid": 870,
    "ts": 7429770636492.817, "dur": 50.386,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770636492.985, "dur": 0.230,
    "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": 870, "tid": 870,
    "ts": 7429770636493.808, "dur": 49.268,
    "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": 870, "tid": 870,
    "ts": 7429770636494.037, "dur": 0.152,
    "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": 870, "tid": 870,
    "ts": 7429770636494.389, "dur": 0.142,
    "args": {
      "Python parent id": 5094, "Python id": 5096, "finished": true, "Ev Idx": 11926
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770636494.638, "dur": 48.323,
    "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": 870, "tid": 870,
    "ts": 7429770636560.444, "dur": 1.302,
    "args": {
      "Python parent id": 64, "Python id": 5098, "finished": true, "Ev Idx": 11928
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: LlamaRMSNorm_64", "pid": 870, "tid": 870,
    "ts": 7429770636565.179, "dur": 68.676,
    "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": 870, "tid": 870,
    "ts": 7429770636565.817, "dur": 67.715,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770636566.095, "dur": 0.278,
    "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": 870, "tid": 870,
    "ts": 7429770636567.375, "dur": 65.313,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770636567.864, "dur": 1.551,
    "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 0x7f8f8b5dc6e0>", "pid": 870, "tid": 870,
    "ts": 7429770636569.707, "dur": 10.452,
    "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 0x7f8f8b4d1b30>", "pid": 870, "tid": 870,
    "ts": 7429770636580.479, "dur": 12.943,
    "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 0x7f9a8535bf40>", "pid": 870, "tid": 870,
    "ts": 7429770636604.321, "dur": 8.505,
    "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": 870, "tid": 870,
    "ts": 7429770636622.344, "dur": 0.270,
    "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 0x7f8f8b5c34d0>", "pid": 870, "tid": 870,
    "ts": 7429770636623.090, "dur": 1.128,
    "args": {
      "Python parent id": 5102, "Python id": 5108, "finished": true, "Ev Idx": 11938
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<string>(2): __init__", "pid": 870, "tid": 870,
    "ts": 7429770636639.635, "dur": 69.292,
    "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": 870, "tid": 870,
    "ts": 7429770636641.472, "dur": 25.144,
    "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 0x7f8f8b4c2d40>", "pid": 870, "tid": 870,
    "ts": 7429770636663.480, "dur": 0.487,
    "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": 870, "tid": 870,
    "ts": 7429770636667.158, "dur": 1.165,
    "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 0x7f8f8b4c2d40>", "pid": 870, "tid": 870,
    "ts": 7429770636667.414, "dur": 0.134,
    "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": 870, "tid": 870,
    "ts": 7429770636668.628, "dur": 0.638,
    "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 0x7f8f8b4c2d40>", "pid": 870, "tid": 870,
    "ts": 7429770636668.728, "dur": 0.069,
    "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": 870, "tid": 870,
    "ts": 7429770636669.633, "dur": 0.753,
    "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 0x7f8f8b4c2d40>", "pid": 870, "tid": 870,
    "ts": 7429770636669.843, "dur": 0.060,
    "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": 870, "tid": 870,
    "ts": 7429770636674.492, "dur": 34.113,
    "args": {
      "Python parent id": 5109, "Python id": 5118, "finished": true, "Ev Idx": 11948
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "dataclasses.py(1278): fields", "pid": 870, "tid": 870,
    "ts": 7429770636676.394, "dur": 8.430,
    "args": {
      "Python parent id": 5118, "Python id": 5119, "finished": true, "Ev Idx": 11949
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770636676.885, "dur": 0.525,
    "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 0x7f8f8b4ab740>", "pid": 870, "tid": 870,
    "ts": 7429770636678.892, "dur": 0.207,
    "args": {
      "Python parent id": 5119, "Python id": 5121, "finished": true, "Ev Idx": 11951
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "dataclasses.py(1293): <genexpr>", "pid": 870, "tid": 870,
    "ts": 7429770636682.045, "dur": 0.517,
    "args": {
      "Python parent id": 5119, "Python id": 5122, "finished": true, "Ev Idx": 11952
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "dataclasses.py(1293): <genexpr>", "pid": 870, "tid": 870,
    "ts": 7429770636682.975, "dur": 0.227,
    "args": {
      "Python parent id": 5119, "Python id": 5123, "finished": true, "Ev Idx": 11953
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "dataclasses.py(1293): <genexpr>", "pid": 870, "tid": 870,
    "ts": 7429770636683.363, "dur": 0.128,
    "args": {
      "Python parent id": 5119, "Python id": 5124, "finished": true, "Ev Idx": 11954
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "dataclasses.py(1293): <genexpr>", "pid": 870, "tid": 870,
    "ts": 7429770636683.663, "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": 870, "tid": 870,
    "ts": 7429770636683.878, "dur": 0.113,
    "args": {
      "Python parent id": 5119, "Python id": 5126, "finished": true, "Ev Idx": 11956
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770636685.191, "dur": 0.133,
    "args": {
      "Python parent id": 5118, "Python id": 5127, "finished": true, "Ev Idx": 11957
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function all>", "pid": 870, "tid": 870,
    "ts": 7429770636687.881, "dur": 3.086,
    "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": 870, "tid": 870,
    "ts": 7429770636689.124, "dur": 0.241,
    "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": 870, "tid": 870,
    "ts": 7429770636689.662, "dur": 0.087,
    "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": 870, "tid": 870,
    "ts": 7429770636689.909, "dur": 0.065,
    "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": 870, "tid": 870,
    "ts": 7429770636690.316, "dur": 0.437,
    "args": {
      "Python parent id": 5128, "Python id": 5132, "finished": true, "Ev Idx": 11962
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770636691.593, "dur": 0.242,
    "args": {
      "Python parent id": 5118, "Python id": 5133, "finished": true, "Ev Idx": 11963
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function all>", "pid": 870, "tid": 870,
    "ts": 7429770636692.942, "dur": 2.445,
    "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": 870, "tid": 870,
    "ts": 7429770636694.548, "dur": 0.758,
    "args": {
      "Python parent id": 5134, "Python id": 5135, "finished": true, "Ev Idx": 11965
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770636695.052, "dur": 0.132,
    "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": 870, "tid": 870,
    "ts": 7429770636696.806, "dur": 0.723,
    "args": {
      "Python parent id": 5118, "Python id": 5137, "finished": true, "Ev Idx": 11967
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770636700.463, "dur": 0.094,
    "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": 870, "tid": 870,
    "ts": 7429770636703.430, "dur": 1.564,
    "args": {
      "Python parent id": 5118, "Python id": 5139, "finished": true, "Ev Idx": 11969
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770636705.898, "dur": 0.116,
    "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": 870, "tid": 870,
    "ts": 7429770636706.661, "dur": 0.643,
    "args": {
      "Python parent id": 5118, "Python id": 5141, "finished": true, "Ev Idx": 11971
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770636707.815, "dur": 0.078,
    "args": {
      "Python parent id": 5118, "Python id": 5142, "finished": true, "Ev Idx": 11972
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770636708.380, "dur": 0.087,
    "args": {
      "Python parent id": 5118, "Python id": 5143, "finished": true, "Ev Idx": 11973
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770636717.655, "dur": 0.195,
    "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": 870, "tid": 870,
    "ts": 7429770636719.359, "dur": 0.981,
    "args": {
      "Python parent id": 48, "Python id": 5145, "finished": true, "Ev Idx": 11975
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "nn.Module: Linear_224", "pid": 870, "tid": 870,
    "ts": 7429770636739.251, "dur": 114.214,
    "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": 870, "tid": 870,
    "ts": 7429770636740.244, "dur": 112.716,
    "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 0x7f99b7dede90>", "pid": 870, "tid": 870,
    "ts": 7429770636740.570, "dur": 0.776,
    "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": 870, "tid": 870,
    "ts": 7429770636742.378, "dur": 110.222,
    "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": 870, "tid": 870,
    "ts": 7429770636742.808, "dur": 0.498,
    "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": 870, "tid": 870,
    "ts": 7429770636743.613, "dur": 0.199,
    "args": {
      "Python parent id": 5149, "Python id": 5151, "finished": true, "Ev Idx": 11981
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function linear>", "pid": 870, "tid": 870,
    "ts": 7429770636744.015, "dur": 108.432,
    "args": {
      "Python parent id": 5149, "Python id": 5152, "finished": true, "Ev Idx": 11982
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<string>(2): __init__", "pid": 870, "tid": 870,
    "ts": 7429770636857.749, "dur": 24.403,
    "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": 870, "tid": 870,
    "ts": 7429770636858.164, "dur": 1.962,
    "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 0x7f8f8b4c2d40>", "pid": 870, "tid": 870,
    "ts": 7429770636858.490, "dur": 0.288,
    "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": 870, "tid": 870,
    "ts": 7429770636860.530, "dur": 0.871,
    "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 0x7f8f8b4c2d40>", "pid": 870, "tid": 870,
    "ts": 7429770636860.798, "dur": 0.079,
    "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": 870, "tid": 870,
    "ts": 7429770636861.672, "dur": 0.544,
    "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 0x7f8f8b4c2d40>", "pid": 870, "tid": 870,
    "ts": 7429770636861.830, "dur": 0.086,
    "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": 870, "tid": 870,
    "ts": 7429770636862.434, "dur": 0.714,
    "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 0x7f8f8b4c2d40>", "pid": 870, "tid": 870,
    "ts": 7429770636862.563, "dur": 0.071,
    "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": 870, "tid": 870,
    "ts": 7429770636863.546, "dur": 0.516,
    "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 0x7f8f8b4c2d40>", "pid": 870, "tid": 870,
    "ts": 7429770636863.649, "dur": 0.068,
    "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": 870, "tid": 870,
    "ts": 7429770636864.894, "dur": 16.993,
    "args": {
      "Python parent id": 5153, "Python id": 5164, "finished": true, "Ev Idx": 11994
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "dataclasses.py(1278): fields", "pid": 870, "tid": 870,
    "ts": 7429770636865.289, "dur": 3.945,
    "args": {
      "Python parent id": 5164, "Python id": 5165, "finished": true, "Ev Idx": 11995
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770636865.546, "dur": 0.241,
    "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 0x7f8f8b4ab740>", "pid": 870, "tid": 870,
    "ts": 7429770636866.299, "dur": 0.164,
    "args": {
      "Python parent id": 5165, "Python id": 5167, "finished": true, "Ev Idx": 11997
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "dataclasses.py(1293): <genexpr>", "pid": 870, "tid": 870,
    "ts": 7429770636867.238, "dur": 0.281,
    "args": {
      "Python parent id": 5165, "Python id": 5168, "finished": true, "Ev Idx": 11998
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "dataclasses.py(1293): <genexpr>", "pid": 870, "tid": 870,
    "ts": 7429770636867.763, "dur": 0.123,
    "args": {
      "Python parent id": 5165, "Python id": 5169, "finished": true, "Ev Idx": 11999
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "dataclasses.py(1293): <genexpr>", "pid": 870, "tid": 870,
    "ts": 7429770636868.049, "dur": 0.050,
    "args": {
      "Python parent id": 5165, "Python id": 5170, "finished": true, "Ev Idx": 12000
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "dataclasses.py(1293): <genexpr>", "pid": 870, "tid": 870,
    "ts": 7429770636868.295, "dur": 0.048,
    "args": {
      "Python parent id": 5165, "Python id": 5171, "finished": true, "Ev Idx": 12001
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "dataclasses.py(1293): <genexpr>", "pid": 870, "tid": 870,
    "ts": 7429770636868.505, "dur": 0.042,
    "args": {
      "Python parent id": 5165, "Python id": 5172, "finished": true, "Ev Idx": 12002
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "dataclasses.py(1293): <genexpr>", "pid": 870, "tid": 870,
    "ts": 7429770636868.706, "dur": 0.090,
    "args": {
      "Python parent id": 5165, "Python id": 5173, "finished": true, "Ev Idx": 12003
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function len>", "pid": 870, "tid": 870,
    "ts": 7429770636869.463, "dur": 0.144,
    "args": {
      "Python parent id": 5164, "Python id": 5174, "finished": true, "Ev Idx": 12004
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function all>", "pid": 870, "tid": 870,
    "ts": 7429770636870.186, "dur": 1.802,
    "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": 870, "tid": 870,
    "ts": 7429770636870.465, "dur": 0.169,
    "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": 870, "tid": 870,
    "ts": 7429770636870.862, "dur": 0.075,
    "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": 870, "tid": 870,
    "ts": 7429770636871.099, "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": 870, "tid": 870,
    "ts": 7429770636871.503, "dur": 0.033,
    "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": 870, "tid": 870,
    "ts": 7429770636871.689, "dur": 0.159,
    "args": {
      "Python parent id": 5175, "Python id": 5180, "finished": true, "Ev Idx": 12010
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770636872.344, "dur": 0.137,
    "args": {
      "Python parent id": 5164, "Python id": 5181, "finished": true, "Ev Idx": 12011
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function all>", "pid": 870, "tid": 870,
    "ts": 7429770636873.242, "dur": 0.828,
    "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": 870, "tid": 870,
    "ts": 7429770636873.560, "dur": 0.450,
    "args": {
      "Python parent id": 5182, "Python id": 5183, "finished": true, "Ev Idx": 12013
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770636873.784, "dur": 0.100,
    "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": 870, "tid": 870,
    "ts": 7429770636874.908, "dur": 0.460,
    "args": {
      "Python parent id": 5164, "Python id": 5185, "finished": true, "Ev Idx": 12015
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770636876.808, "dur": 0.106,
    "args": {
      "Python parent id": 5164, "Python id": 5186, "finished": true, "Ev Idx": 12016
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770636877.408, "dur": 0.089,
    "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": 870, "tid": 870,
    "ts": 7429770636878.190, "dur": 0.909,
    "args": {
      "Python parent id": 5164, "Python id": 5188, "finished": true, "Ev Idx": 12018
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770636879.589, "dur": 0.057,
    "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": 870, "tid": 870,
    "ts": 7429770636880.213, "dur": 0.569,
    "args": {
      "Python parent id": 5164, "Python id": 5190, "finished": true, "Ev Idx": 12020
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770636881.203, "dur": 0.073,
    "args": {
      "Python parent id": 5164, "Python id": 5191, "finished": true, "Ev Idx": 12021
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770636881.707, "dur": 0.077,
    "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": 870, "tid": 870,
    "ts": 7429770636965.176, "dur": 39247.045,
    "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": 870, "tid": 870,
    "ts": 7429770636968.853, "dur": 39243.368,
    "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": 870, "tid": 870,
    "ts": 7429770636982.841, "dur": 39229.380,
    "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 0x7f8f8b4c4f80>", "pid": 870, "tid": 870,
    "ts": 7429770636987.136, "dur": 6.404,
    "args": {
      "Python parent id": 5195, "Python id": 5196, "finished": true, "Ev Idx": 12026
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "enum.py(1261): __hash__", "pid": 870, "tid": 870,
    "ts": 7429770636989.964, "dur": 2.592,
    "args": {
      "Python parent id": 5196, "Python id": 5197, "finished": true, "Ev Idx": 12027
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hash>", "pid": 870, "tid": 870,
    "ts": 7429770636992.144, "dur": 0.335,
    "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": 870, "tid": 870,
    "ts": 7429770636995.625, "dur": 39216.596,
    "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": 870, "tid": 870,
    "ts": 7429770637002.351, "dur": 39209.870,
    "args": {
      "Python parent id": 5199, "Python id": 5200, "finished": true, "Ev Idx": 12030
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770637004.489, "dur": 0.482,
    "args": {
      "Python parent id": 5200, "Python id": 5201, "finished": true, "Ev Idx": 12031
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function getattr>", "pid": 870, "tid": 870,
    "ts": 7429770637011.873, "dur": 0.231,
    "args": {
      "Python parent id": 5200, "Python id": 5202, "finished": true, "Ev Idx": 12032
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770637013.117, "dur": 2.111,
    "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": 870, "tid": 870,
    "ts": 7429770637017.910, "dur": 39194.311,
    "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": 870, "tid": 870,
    "ts": 7429770637023.709, "dur": 7.697,
    "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": 870, "tid": 870,
    "ts": 7429770637026.020, "dur": 4.875,
    "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": 870, "tid": 870,
    "ts": 7429770637029.872, "dur": 0.881,
    "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": 870, "tid": 870,
    "ts": 7429770637037.216, "dur": 72.522,
    "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": 870, "tid": 870,
    "ts": 7429770637041.725, "dur": 66.973,
    "args": {
      "Python parent id": 5208, "Python id": 5209, "finished": true, "Ev Idx": 12039
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770637043.417, "dur": 0.188,
    "args": {
      "Python parent id": 5209, "Python id": 5210, "finished": true, "Ev Idx": 12040
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770637044.338, "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": 870, "tid": 870,
    "ts": 7429770637045.329, "dur": 0.045,
    "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": 870, "tid": 870,
    "ts": 7429770637048.765, "dur": 0.093,
    "args": {
      "Python parent id": 5209, "Python id": 5213, "finished": true, "Ev Idx": 12043
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770637049.924, "dur": 0.136,
    "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": 870, "tid": 870,
    "ts": 7429770637054.544, "dur": 54.018,
    "args": {
      "Python parent id": 5209, "Python id": 5215, "finished": true, "Ev Idx": 12045
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770637056.020, "dur": 0.082,
    "args": {
      "Python parent id": 5215, "Python id": 5216, "finished": true, "Ev Idx": 12046
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770637056.817, "dur": 0.045,
    "args": {
      "Python parent id": 5215, "Python id": 5217, "finished": true, "Ev Idx": 12047
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770637058.126, "dur": 0.062,
    "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": 870, "tid": 870,
    "ts": 7429770637058.994, "dur": 0.040,
    "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": 870, "tid": 870,
    "ts": 7429770637061.171, "dur": 47.233,
    "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": 870, "tid": 870,
    "ts": 7429770637065.162, "dur": 43.069,
    "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": 870, "tid": 870,
    "ts": 7429770637069.079, "dur": 28.369,
    "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": 870, "tid": 870,
    "ts": 7429770637071.682, "dur": 25.293,
    "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": 870, "tid": 870,
    "ts": 7429770637073.421, "dur": 2.007,
    "args": {
      "Python parent id": 5223, "Python id": 5224, "finished": true, "Ev Idx": 12054
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function hasattr>", "pid": 870, "tid": 870,
    "ts": 7429770637074.158, "dur": 0.823,
    "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": 870, "tid": 870,
    "ts": 7429770637077.042, "dur": 17.793,
    "args": {
      "Python parent id": 5223, "Python id": 5226, "finished": true, "Ev Idx": 12056
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(808): getenv", "pid": 870, "tid": 870,
    "ts": 7429770637080.427, "dur": 14.022,
    "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": 870, "tid": 870,
    "ts": 7429770637082.781, "dur": 11.536,
    "args": {
      "Python parent id": 5227, "Python id": 5228, "finished": true, "Ev Idx": 12058
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(709): __getitem__", "pid": 870, "tid": 870,
    "ts": 7429770637084.501, "dur": 8.825,
    "args": {
      "Python parent id": 5228, "Python id": 5229, "finished": true, "Ev Idx": 12059
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<frozen os>(791): encode", "pid": 870, "tid": 870,
    "ts": 7429770637087.158, "dur": 3.597,
    "args": {
      "Python parent id": 5229, "Python id": 5230, "finished": true, "Ev Idx": 12060
    }
  },
  {
    "ph": "X", "cat": "python_function", "name": "<built-in function isinstance>", "pid": 870, "tid": 870,
    "ts": 7429770637087.452, "dur": 0.059,
    "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 0x7f98edfc58e0>", "pid": 870, "tid": 870,
    "ts": 7429770637089.729, "dur": 0.903,
    "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": 870, "tid": 870,
    "ts": 7429770637096.388, "dur": 0.459,
    "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 0x7f9a860afab0>", "pid": 870, "tid": 870,
    "ts": 7429770637099.405, "dur": 0.573,
    "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": 870, "tid": 870,
    "ts": 7429770637101.418, "dur": 6.699,
    "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": 870, "tid": 870,
    "ts": 7429770637103.374, "dur": 4.666,
    "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": 870, "tid": 870,
    "ts": 7429770637104.215, "dur": 1.170,
    "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": 870, "tid": 870,
    "ts": 7429770637104.834, "dur": 0.468,
    "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": 870, "tid": 870,
    "ts": 7429770637105.157, "dur": 0.085,
    "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": 870, "tid": 870,
    "ts": 7429770637106.821, "dur": 1.153,
    "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": 870, "tid": 870,
    "ts": 7429770637112.077, "dur": 3.083,
    "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": 870, "tid": 870,
    "ts": 7429770637113.741, "dur": 1.241,
    "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": 870, "tid": 870,
    "ts": 7429770637117.359, "dur": 39075.107,
    "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": 870, "tid": 870,
    "ts": 7429770676194.277, "dur": 17.944,
    "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": 870, "tid": 870,
    "ts": 7429770676196.303, "dur": 0.712,
    "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": 870, "tid": 870,
    "ts": 7429770676200.645, "dur": 0.842,
    "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 0x7f984a877090>", "pid": 870, "tid": 870,
    "ts": 7429770676204.215, "dur": 8.006,
    "args": {
      "Python parent id": 5244, "Python id": 5247, "finished": true, "Ev Idx": 12077
    }
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770599315.766, "dur": 15.633,
    "args": {
      "External id": 4,          "kernel": "void at::native::(anonymous namespace)::indexSelectLargeIndex<float, long, unsigned int, 2, 2, -2, true>(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, unsigned int, long)", "cid": 107, "correlation": 1,      "grid": [672, 1, 1],      "block": [512, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1, "pid": 870, "tid": 870, "ts": 7429770599315.766,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770599379.445, "dur": 7.682,
    "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": 870, "tid": 870, "ts": 7429770599379.445,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770599463.274, "dur": 6.541,
    "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": [7, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 3, "pid": 870, "tid": 870, "ts": 7429770599463.274,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770599480.378, "dur": 6.825,
    "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": [28, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 4, "pid": 870, "tid": 870, "ts": 7429770599480.378,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770599497.789, "dur": 3.612,
    "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": 870, "tid": 870, "ts": 7429770599497.789,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770599519.890, "dur": 5.271,
    "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": [14, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 6, "pid": 870, "tid": 870, "ts": 7429770599519.890,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770599536.967, "dur": 5.711,
    "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": [14, 1, 1],      "block": [512, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 7, "pid": 870, "tid": 870, "ts": 7429770599536.967,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 870, "tid": 870,
    "ts": 7429770599603.448, "dur": 11.910,
    "args": {
      "External id": 38,      "cid": 134, "correlation": 8, "src": "0x7f84521f8000", "dst": "0x7f8451df8400", "bytes": "28224", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 8, "pid": 870, "tid": 870, "ts": 7429770599603.448,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770599650.956, "dur": 5.859,
    "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": [14, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 9, "pid": 870, "tid": 870, "ts": 7429770599650.956,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770599670.864, "dur": 5.983,
    "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": [2, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 10, "pid": 870, "tid": 870, "ts": 7429770599670.864,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 870, "tid": 870,
    "ts": 7429770599704.281, "dur": 5.751,
    "args": {
      "External id": 66,      "cid": 134, "correlation": 11, "src": "0x7f8451df8400", "dst": "0x7f84521f8000", "bytes": "28224", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 11, "pid": 870, "tid": 870, "ts": 7429770599704.281,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770599720.551, "dur": 4.229,
    "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": [7, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 12, "pid": 870, "tid": 870, "ts": 7429770599720.551,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMemcpyAsync", "pid": 870, "tid": 870,
    "ts": 7429770599736.864, "dur": 3.588,
    "args": {
      "External id": 74,      "cid": 134, "correlation": 13, "src": "0x7f84521f8000", "dst": "0x7f8451df8400", "bytes": "28224", "kind": "3"
    }
  },
  {
    "ph": "s", "id": 13, "pid": 870, "tid": 870, "ts": 7429770599736.864,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770599826.885, "dur": 5.949,
    "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": 870, "tid": 870, "ts": 7429770599826.885,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770599929.637, "dur": 0.454,
    "args": {
      "External id": 106,      "cid": 370, "correlation": 15
    }
  },
  {
    "ph": "f", "id": 15, "pid": 870, "tid": 870, "ts": 7429770599929.637,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770599940.719, "dur": 0.242,
    "args": {
      "External id": 106,      "cid": 370, "correlation": 16
    }
  },
  {
    "ph": "f", "id": 16, "pid": 870, "tid": 870, "ts": 7429770599940.719,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770599946.247, "dur": 0.257,
    "args": {
      "External id": 106,      "cid": 370, "correlation": 17
    }
  },
  {
    "ph": "f", "id": 17, "pid": 870, "tid": 870, "ts": 7429770599946.247,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770599954.120, "dur": 6.487,
    "args": {
      "External id": 106,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 18,      "grid": [1536, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 18, "pid": 870, "tid": 870, "ts": 7429770599954.120,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770599991.574, "dur": 7.649,
    "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": [3, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 19, "pid": 870, "tid": 870, "ts": 7429770599991.574,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770600010.306, "dur": 5.339,
    "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": [11, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 20, "pid": 870, "tid": 870, "ts": 7429770600010.306,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770600026.842, "dur": 5.041,
    "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": [11, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 21, "pid": 870, "tid": 870, "ts": 7429770600026.842,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770600041.007, "dur": 4.744,
    "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": [11, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 22, "pid": 870, "tid": 870, "ts": 7429770600041.007,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770600050.818, "dur": 2.850,
    "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": [11, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 23, "pid": 870, "tid": 870, "ts": 7429770600050.818,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770600273.975, "dur": 5.470,
    "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": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 24, "pid": 870, "tid": 870, "ts": 7429770600273.975,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770600305.970, "dur": 6.529,
    "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": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 25, "pid": 870, "tid": 870, "ts": 7429770600305.970,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770600325.718, "dur": 5.061,
    "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": 870, "tid": 870, "ts": 7429770600325.718,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770600339.475, "dur": 5.684,
    "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": 870, "tid": 870, "ts": 7429770600339.475,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770600356.604, "dur": 4.555,
    "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": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 28, "pid": 870, "tid": 870, "ts": 7429770600356.604,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770600371.154, "dur": 3.577,
    "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": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 29, "pid": 870, "tid": 870, "ts": 7429770600371.154,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770600457.733, "dur": 0.396,
    "args": {
      "External id": 135,      "cid": 370, "correlation": 30
    }
  },
  {
    "ph": "f", "id": 30, "pid": 870, "tid": 870, "ts": 7429770600457.733,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770600464.570, "dur": 0.144,
    "args": {
      "External id": 135,      "cid": 370, "correlation": 31
    }
  },
  {
    "ph": "f", "id": 31, "pid": 870, "tid": 870, "ts": 7429770600464.570,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770600468.648, "dur": 0.381,
    "args": {
      "External id": 135,      "cid": 370, "correlation": 32
    }
  },
  {
    "ph": "f", "id": 32, "pid": 870, "tid": 870, "ts": 7429770600468.648,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770600473.906, "dur": 5.982,
    "args": {
      "External id": 135,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 33,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 33, "pid": 870, "tid": 870, "ts": 7429770600473.906,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770600531.300, "dur": 0.188,
    "args": {
      "External id": 148,      "cid": 370, "correlation": 34
    }
  },
  {
    "ph": "f", "id": 34, "pid": 870, "tid": 870, "ts": 7429770600531.300,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770600534.566, "dur": 0.140,
    "args": {
      "External id": 148,      "cid": 370, "correlation": 35
    }
  },
  {
    "ph": "f", "id": 35, "pid": 870, "tid": 870, "ts": 7429770600534.566,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770600537.499, "dur": 0.156,
    "args": {
      "External id": 148,      "cid": 370, "correlation": 36
    }
  },
  {
    "ph": "f", "id": 36, "pid": 870, "tid": 870, "ts": 7429770600537.499,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770600539.795, "dur": 5.446,
    "args": {
      "External id": 148,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 37,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 37, "pid": 870, "tid": 870, "ts": 7429770600539.795,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770600583.675, "dur": 0.169,
    "args": {
      "External id": 161,      "cid": 370, "correlation": 38
    }
  },
  {
    "ph": "f", "id": 38, "pid": 870, "tid": 870, "ts": 7429770600583.675,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770600585.979, "dur": 0.139,
    "args": {
      "External id": 161,      "cid": 370, "correlation": 39
    }
  },
  {
    "ph": "f", "id": 39, "pid": 870, "tid": 870, "ts": 7429770600585.979,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770600588.052, "dur": 0.149,
    "args": {
      "External id": 161,      "cid": 370, "correlation": 40
    }
  },
  {
    "ph": "f", "id": 40, "pid": 870, "tid": 870, "ts": 7429770600588.052,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770600589.835, "dur": 4.763,
    "args": {
      "External id": 161,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 41,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 41, "pid": 870, "tid": 870, "ts": 7429770600589.835,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770600623.224, "dur": 4.115,
    "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": 42,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 42, "pid": 870, "tid": 870, "ts": 7429770600623.224,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770600652.783, "dur": 4.683,
    "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": 43,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 43, "pid": 870, "tid": 870, "ts": 7429770600652.783,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770600672.478, "dur": 7.324,
    "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": 44,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 44, "pid": 870, "tid": 870, "ts": 7429770600672.478,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770600688.892, "dur": 3.940,
    "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": 45,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 45, "pid": 870, "tid": 870, "ts": 7429770600688.892,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770600699.953, "dur": 4.822,
    "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": 46,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 46, "pid": 870, "tid": 870, "ts": 7429770600699.953,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770600711.518, "dur": 3.648,
    "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": 47,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 47, "pid": 870, "tid": 870, "ts": 7429770600711.518,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770600729.115, "dur": 3.224,
    "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": 48,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 48, "pid": 870, "tid": 870, "ts": 7429770600729.115,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770600740.057, "dur": 4.592,
    "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": 49,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 49, "pid": 870, "tid": 870, "ts": 7429770600740.057,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770600751.408, "dur": 3.603,
    "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": 50,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 50, "pid": 870, "tid": 870, "ts": 7429770600751.408,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770600759.438, "dur": 2.964,
    "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": 51,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 51, "pid": 870, "tid": 870, "ts": 7429770600759.438,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770600835.817, "dur": 5.518,
    "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": 52,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 52, "pid": 870, "tid": 870, "ts": 7429770600835.817,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770600871.125, "dur": 3.101,
    "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": 53,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 53, "pid": 870, "tid": 870, "ts": 7429770600871.125,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770600921.474, "dur": 0.375,
    "args": {
      "External id": 236,      "cid": 370, "correlation": 54
    }
  },
  {
    "ph": "f", "id": 54, "pid": 870, "tid": 870, "ts": 7429770600921.474,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770600926.324, "dur": 0.153,
    "args": {
      "External id": 236,      "cid": 370, "correlation": 55
    }
  },
  {
    "ph": "f", "id": 55, "pid": 870, "tid": 870, "ts": 7429770600926.324,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770600929.659, "dur": 0.289,
    "args": {
      "External id": 236,      "cid": 370, "correlation": 56
    }
  },
  {
    "ph": "f", "id": 56, "pid": 870, "tid": 870, "ts": 7429770600929.659,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770600932.666, "dur": 5.135,
    "args": {
      "External id": 236,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 57,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 57, "pid": 870, "tid": 870, "ts": 7429770600932.666,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770600950.775, "dur": 4.016,
    "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": 58,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 58, "pid": 870, "tid": 870, "ts": 7429770600950.775,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770600970.588, "dur": 3.650,
    "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": 59,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 59, "pid": 870, "tid": 870, "ts": 7429770600970.588,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770600992.865, "dur": 5.145,
    "args": {
      "External id": 249,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 60,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 60, "pid": 870, "tid": 870, "ts": 7429770600992.865,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770601048.123, "dur": 0.295,
    "args": {
      "External id": 261,      "cid": 370, "correlation": 61
    }
  },
  {
    "ph": "f", "id": 61, "pid": 870, "tid": 870, "ts": 7429770601048.123,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770601051.894, "dur": 0.140,
    "args": {
      "External id": 261,      "cid": 370, "correlation": 62
    }
  },
  {
    "ph": "f", "id": 62, "pid": 870, "tid": 870, "ts": 7429770601051.894,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770601054.882, "dur": 0.150,
    "args": {
      "External id": 261,      "cid": 370, "correlation": 63
    }
  },
  {
    "ph": "f", "id": 63, "pid": 870, "tid": 870, "ts": 7429770601054.882,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601057.673, "dur": 6.041,
    "args": {
      "External id": 261,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 64,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 64, "pid": 870, "tid": 870, "ts": 7429770601057.673,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601083.645, "dur": 3.270,
    "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": 65,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 65, "pid": 870, "tid": 870, "ts": 7429770601083.645,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770601138.291, "dur": 0.256,
    "args": {
      "External id": 280,      "cid": 370, "correlation": 66
    }
  },
  {
    "ph": "f", "id": 66, "pid": 870, "tid": 870, "ts": 7429770601138.291,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770601141.194, "dur": 0.143,
    "args": {
      "External id": 280,      "cid": 370, "correlation": 67
    }
  },
  {
    "ph": "f", "id": 67, "pid": 870, "tid": 870, "ts": 7429770601141.194,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770601143.842, "dur": 0.518,
    "args": {
      "External id": 280,      "cid": 370, "correlation": 68
    }
  },
  {
    "ph": "f", "id": 68, "pid": 870, "tid": 870, "ts": 7429770601143.842,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601146.325, "dur": 4.094,
    "args": {
      "External id": 280,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 69,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 69, "pid": 870, "tid": 870, "ts": 7429770601146.325,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601164.739, "dur": 3.593,
    "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": 70,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 70, "pid": 870, "tid": 870, "ts": 7429770601164.739,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601189.233, "dur": 4.496,
    "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": 71,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 71, "pid": 870, "tid": 870, "ts": 7429770601189.233,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601205.390, "dur": 4.595,
    "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": 72,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 72, "pid": 870, "tid": 870, "ts": 7429770601205.390,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601218.739, "dur": 4.025,
    "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": 73,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 73, "pid": 870, "tid": 870, "ts": 7429770601218.739,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601229.441, "dur": 4.373,
    "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": 74,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 74, "pid": 870, "tid": 870, "ts": 7429770601229.441,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601241.708, "dur": 2.914,
    "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": 75,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 75, "pid": 870, "tid": 870, "ts": 7429770601241.708,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601252.899, "dur": 4.095,
    "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": 76,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 76, "pid": 870, "tid": 870, "ts": 7429770601252.899,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770601302.738, "dur": 0.220,
    "args": {
      "External id": 301,      "cid": 370, "correlation": 77
    }
  },
  {
    "ph": "f", "id": 77, "pid": 870, "tid": 870, "ts": 7429770601302.738,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770601306.098, "dur": 0.169,
    "args": {
      "External id": 301,      "cid": 370, "correlation": 78
    }
  },
  {
    "ph": "f", "id": 78, "pid": 870, "tid": 870, "ts": 7429770601306.098,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770601308.917, "dur": 0.156,
    "args": {
      "External id": 301,      "cid": 370, "correlation": 79
    }
  },
  {
    "ph": "f", "id": 79, "pid": 870, "tid": 870, "ts": 7429770601308.917,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601311.004, "dur": 5.454,
    "args": {
      "External id": 301,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 80,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 80, "pid": 870, "tid": 870, "ts": 7429770601311.004,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601340.681, "dur": 3.842,
    "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": 81,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 81, "pid": 870, "tid": 870, "ts": 7429770601340.681,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770601372.290, "dur": 0.193,
    "args": {
      "External id": 312,      "cid": 370, "correlation": 82
    }
  },
  {
    "ph": "f", "id": 82, "pid": 870, "tid": 870, "ts": 7429770601372.290,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770601374.795, "dur": 0.136,
    "args": {
      "External id": 312,      "cid": 370, "correlation": 83
    }
  },
  {
    "ph": "f", "id": 83, "pid": 870, "tid": 870, "ts": 7429770601374.795,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770601376.916, "dur": 0.139,
    "args": {
      "External id": 312,      "cid": 370, "correlation": 84
    }
  },
  {
    "ph": "f", "id": 84, "pid": 870, "tid": 870, "ts": 7429770601376.916,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601378.670, "dur": 4.497,
    "args": {
      "External id": 312,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 85,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 85, "pid": 870, "tid": 870, "ts": 7429770601378.670,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601391.833, "dur": 4.018,
    "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": 86,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 86, "pid": 870, "tid": 870, "ts": 7429770601391.833,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770601424.357, "dur": 0.176,
    "args": {
      "External id": 323,      "cid": 370, "correlation": 87
    }
  },
  {
    "ph": "f", "id": 87, "pid": 870, "tid": 870, "ts": 7429770601424.357,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770601428.022, "dur": 0.140,
    "args": {
      "External id": 323,      "cid": 370, "correlation": 88
    }
  },
  {
    "ph": "f", "id": 88, "pid": 870, "tid": 870, "ts": 7429770601428.022,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770601430.535, "dur": 0.141,
    "args": {
      "External id": 323,      "cid": 370, "correlation": 89
    }
  },
  {
    "ph": "f", "id": 89, "pid": 870, "tid": 870, "ts": 7429770601430.535,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601439.685, "dur": 3.953,
    "args": {
      "External id": 323,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 90,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 90, "pid": 870, "tid": 870, "ts": 7429770601439.685,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601443.987, "dur": 4.288,
    "args": {
      "External id": 323,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 91,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 91, "pid": 870, "tid": 870, "ts": 7429770601443.987,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601458.293, "dur": 3.508,
    "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": 92,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 92, "pid": 870, "tid": 870, "ts": 7429770601458.293,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601490.185, "dur": 3.634,
    "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": 93,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 93, "pid": 870, "tid": 870, "ts": 7429770601490.185,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601502.303, "dur": 6.120,
    "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": 94,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 94, "pid": 870, "tid": 870, "ts": 7429770601502.303,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601515.350, "dur": 10.491,
    "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": 95,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 95, "pid": 870, "tid": 870, "ts": 7429770601515.350,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601531.557, "dur": 3.823,
    "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": 96,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 96, "pid": 870, "tid": 870, "ts": 7429770601531.557,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601542.050, "dur": 5.496,
    "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": 97,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 97, "pid": 870, "tid": 870, "ts": 7429770601542.050,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601555.124, "dur": 4.217,
    "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": 98,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 98, "pid": 870, "tid": 870, "ts": 7429770601555.124,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770601599.497, "dur": 0.226,
    "args": {
      "External id": 344,      "cid": 370, "correlation": 99
    }
  },
  {
    "ph": "f", "id": 99, "pid": 870, "tid": 870, "ts": 7429770601599.497,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770601603.842, "dur": 0.159,
    "args": {
      "External id": 344,      "cid": 370, "correlation": 100
    }
  },
  {
    "ph": "f", "id": 100, "pid": 870, "tid": 870, "ts": 7429770601603.842,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770601606.387, "dur": 0.176,
    "args": {
      "External id": 344,      "cid": 370, "correlation": 101
    }
  },
  {
    "ph": "f", "id": 101, "pid": 870, "tid": 870, "ts": 7429770601606.387,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601609.156, "dur": 5.150,
    "args": {
      "External id": 344,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 102,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 102, "pid": 870, "tid": 870, "ts": 7429770601609.156,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770601652.476, "dur": 0.249,
    "args": {
      "External id": 357,      "cid": 370, "correlation": 103
    }
  },
  {
    "ph": "f", "id": 103, "pid": 870, "tid": 870, "ts": 7429770601652.476,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770601655.457, "dur": 0.151,
    "args": {
      "External id": 357,      "cid": 370, "correlation": 104
    }
  },
  {
    "ph": "f", "id": 104, "pid": 870, "tid": 870, "ts": 7429770601655.457,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770601658.083, "dur": 0.148,
    "args": {
      "External id": 357,      "cid": 370, "correlation": 105
    }
  },
  {
    "ph": "f", "id": 105, "pid": 870, "tid": 870, "ts": 7429770601658.083,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601659.660, "dur": 5.321,
    "args": {
      "External id": 357,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 106,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 106, "pid": 870, "tid": 870, "ts": 7429770601659.660,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770601700.487, "dur": 0.176,
    "args": {
      "External id": 370,      "cid": 370, "correlation": 107
    }
  },
  {
    "ph": "f", "id": 107, "pid": 870, "tid": 870, "ts": 7429770601700.487,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770601703.720, "dur": 0.159,
    "args": {
      "External id": 370,      "cid": 370, "correlation": 108
    }
  },
  {
    "ph": "f", "id": 108, "pid": 870, "tid": 870, "ts": 7429770601703.720,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770601705.632, "dur": 0.153,
    "args": {
      "External id": 370,      "cid": 370, "correlation": 109
    }
  },
  {
    "ph": "f", "id": 109, "pid": 870, "tid": 870, "ts": 7429770601705.632,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601707.180, "dur": 4.893,
    "args": {
      "External id": 370,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 110,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 110, "pid": 870, "tid": 870, "ts": 7429770601707.180,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601734.972, "dur": 3.831,
    "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": 111,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 111, "pid": 870, "tid": 870, "ts": 7429770601734.972,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601759.736, "dur": 4.688,
    "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": 112,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 112, "pid": 870, "tid": 870, "ts": 7429770601759.736,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601777.682, "dur": 7.163,
    "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": 113,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 113, "pid": 870, "tid": 870, "ts": 7429770601777.682,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601792.492, "dur": 5.034,
    "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": 114,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 114, "pid": 870, "tid": 870, "ts": 7429770601792.492,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601802.830, "dur": 4.942,
    "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": 115,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 115, "pid": 870, "tid": 870, "ts": 7429770601802.830,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601813.886, "dur": 4.352,
    "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": 116,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 116, "pid": 870, "tid": 870, "ts": 7429770601813.886,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601830.323, "dur": 4.627,
    "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": 117,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 117, "pid": 870, "tid": 870, "ts": 7429770601830.323,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601843.100, "dur": 4.762,
    "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": 118,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 118, "pid": 870, "tid": 870, "ts": 7429770601843.100,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601853.727, "dur": 3.557,
    "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": 119,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 119, "pid": 870, "tid": 870, "ts": 7429770601853.727,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601862.274, "dur": 3.319,
    "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": 120,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 120, "pid": 870, "tid": 870, "ts": 7429770601862.274,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601922.649, "dur": 4.773,
    "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": 121,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 121, "pid": 870, "tid": 870, "ts": 7429770601922.649,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770601956.925, "dur": 4.281,
    "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": 122,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 122, "pid": 870, "tid": 870, "ts": 7429770601956.925,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770601996.188, "dur": 0.235,
    "args": {
      "External id": 445,      "cid": 370, "correlation": 123
    }
  },
  {
    "ph": "f", "id": 123, "pid": 870, "tid": 870, "ts": 7429770601996.188,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770602000.423, "dur": 0.164,
    "args": {
      "External id": 445,      "cid": 370, "correlation": 124
    }
  },
  {
    "ph": "f", "id": 124, "pid": 870, "tid": 870, "ts": 7429770602000.423,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770602003.093, "dur": 0.145,
    "args": {
      "External id": 445,      "cid": 370, "correlation": 125
    }
  },
  {
    "ph": "f", "id": 125, "pid": 870, "tid": 870, "ts": 7429770602003.093,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602005.584, "dur": 5.376,
    "args": {
      "External id": 445,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 126,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 126, "pid": 870, "tid": 870, "ts": 7429770602005.584,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602022.463, "dur": 4.162,
    "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": 127,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 127, "pid": 870, "tid": 870, "ts": 7429770602022.463,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602040.968, "dur": 3.751,
    "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": 128,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 128, "pid": 870, "tid": 870, "ts": 7429770602040.968,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602058.120, "dur": 5.205,
    "args": {
      "External id": 458,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 129,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 129, "pid": 870, "tid": 870, "ts": 7429770602058.120,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770602135.352, "dur": 0.341,
    "args": {
      "External id": 470,      "cid": 370, "correlation": 130
    }
  },
  {
    "ph": "f", "id": 130, "pid": 870, "tid": 870, "ts": 7429770602135.352,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770602139.895, "dur": 0.147,
    "args": {
      "External id": 470,      "cid": 370, "correlation": 131
    }
  },
  {
    "ph": "f", "id": 131, "pid": 870, "tid": 870, "ts": 7429770602139.895,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770602142.740, "dur": 0.144,
    "args": {
      "External id": 470,      "cid": 370, "correlation": 132
    }
  },
  {
    "ph": "f", "id": 132, "pid": 870, "tid": 870, "ts": 7429770602142.740,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602145.644, "dur": 6.559,
    "args": {
      "External id": 470,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 133,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 133, "pid": 870, "tid": 870, "ts": 7429770602145.644,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602171.570, "dur": 4.083,
    "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": 134,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 134, "pid": 870, "tid": 870, "ts": 7429770602171.570,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770602222.223, "dur": 0.273,
    "args": {
      "External id": 489,      "cid": 370, "correlation": 135
    }
  },
  {
    "ph": "f", "id": 135, "pid": 870, "tid": 870, "ts": 7429770602222.223,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770602225.747, "dur": 0.155,
    "args": {
      "External id": 489,      "cid": 370, "correlation": 136
    }
  },
  {
    "ph": "f", "id": 136, "pid": 870, "tid": 870, "ts": 7429770602225.747,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770602228.498, "dur": 0.567,
    "args": {
      "External id": 489,      "cid": 370, "correlation": 137
    }
  },
  {
    "ph": "f", "id": 137, "pid": 870, "tid": 870, "ts": 7429770602228.498,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602230.788, "dur": 4.898,
    "args": {
      "External id": 489,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 138,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 138, "pid": 870, "tid": 870, "ts": 7429770602230.788,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602248.389, "dur": 4.549,
    "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": 139,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 139, "pid": 870, "tid": 870, "ts": 7429770602248.389,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602272.689, "dur": 4.945,
    "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": 140,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 140, "pid": 870, "tid": 870, "ts": 7429770602272.689,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602288.808, "dur": 5.934,
    "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": 141,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 141, "pid": 870, "tid": 870, "ts": 7429770602288.808,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602311.718, "dur": 5.458,
    "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": 142,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 142, "pid": 870, "tid": 870, "ts": 7429770602311.718,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602324.849, "dur": 4.207,
    "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": 143,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 143, "pid": 870, "tid": 870, "ts": 7429770602324.849,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602336.326, "dur": 7.107,
    "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": 144,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 144, "pid": 870, "tid": 870, "ts": 7429770602336.326,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602353.390, "dur": 4.285,
    "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": 145,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 145, "pid": 870, "tid": 870, "ts": 7429770602353.390,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770602398.675, "dur": 0.217,
    "args": {
      "External id": 510,      "cid": 370, "correlation": 146
    }
  },
  {
    "ph": "f", "id": 146, "pid": 870, "tid": 870, "ts": 7429770602398.675,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770602402.225, "dur": 0.159,
    "args": {
      "External id": 510,      "cid": 370, "correlation": 147
    }
  },
  {
    "ph": "f", "id": 147, "pid": 870, "tid": 870, "ts": 7429770602402.225,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770602404.890, "dur": 0.163,
    "args": {
      "External id": 510,      "cid": 370, "correlation": 148
    }
  },
  {
    "ph": "f", "id": 148, "pid": 870, "tid": 870, "ts": 7429770602404.890,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602407.327, "dur": 5.525,
    "args": {
      "External id": 510,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 149,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 149, "pid": 870, "tid": 870, "ts": 7429770602407.327,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602462.923, "dur": 5.011,
    "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": 150,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 150, "pid": 870, "tid": 870, "ts": 7429770602462.923,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770602492.496, "dur": 0.163,
    "args": {
      "External id": 521,      "cid": 370, "correlation": 151
    }
  },
  {
    "ph": "f", "id": 151, "pid": 870, "tid": 870, "ts": 7429770602492.496,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770602495.193, "dur": 0.137,
    "args": {
      "External id": 521,      "cid": 370, "correlation": 152
    }
  },
  {
    "ph": "f", "id": 152, "pid": 870, "tid": 870, "ts": 7429770602495.193,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770602497.646, "dur": 0.140,
    "args": {
      "External id": 521,      "cid": 370, "correlation": 153
    }
  },
  {
    "ph": "f", "id": 153, "pid": 870, "tid": 870, "ts": 7429770602497.646,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602499.268, "dur": 5.349,
    "args": {
      "External id": 521,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 154,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 154, "pid": 870, "tid": 870, "ts": 7429770602499.268,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602513.218, "dur": 4.763,
    "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": 155,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 155, "pid": 870, "tid": 870, "ts": 7429770602513.218,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770602544.146, "dur": 0.181,
    "args": {
      "External id": 532,      "cid": 370, "correlation": 156
    }
  },
  {
    "ph": "f", "id": 156, "pid": 870, "tid": 870, "ts": 7429770602544.146,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770602547.356, "dur": 0.141,
    "args": {
      "External id": 532,      "cid": 370, "correlation": 157
    }
  },
  {
    "ph": "f", "id": 157, "pid": 870, "tid": 870, "ts": 7429770602547.356,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770602549.684, "dur": 0.138,
    "args": {
      "External id": 532,      "cid": 370, "correlation": 158
    }
  },
  {
    "ph": "f", "id": 158, "pid": 870, "tid": 870, "ts": 7429770602549.684,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602558.933, "dur": 5.252,
    "args": {
      "External id": 532,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 159,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 159, "pid": 870, "tid": 870, "ts": 7429770602558.933,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602564.558, "dur": 5.923,
    "args": {
      "External id": 532,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 160,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 160, "pid": 870, "tid": 870, "ts": 7429770602564.558,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602579.668, "dur": 4.213,
    "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": 161,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 161, "pid": 870, "tid": 870, "ts": 7429770602579.668,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602613.368, "dur": 3.757,
    "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": 162,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 162, "pid": 870, "tid": 870, "ts": 7429770602613.368,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602625.772, "dur": 7.205,
    "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": 163,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 163, "pid": 870, "tid": 870, "ts": 7429770602625.772,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602642.370, "dur": 4.960,
    "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": 164,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 164, "pid": 870, "tid": 870, "ts": 7429770602642.370,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602653.623, "dur": 4.276,
    "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": 165,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 165, "pid": 870, "tid": 870, "ts": 7429770602653.623,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602663.197, "dur": 3.947,
    "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": 166,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 166, "pid": 870, "tid": 870, "ts": 7429770602663.197,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602675.285, "dur": 4.042,
    "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": 167,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 167, "pid": 870, "tid": 870, "ts": 7429770602675.285,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770602722.899, "dur": 0.279,
    "args": {
      "External id": 553,      "cid": 370, "correlation": 168
    }
  },
  {
    "ph": "f", "id": 168, "pid": 870, "tid": 870, "ts": 7429770602722.899,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770602726.752, "dur": 0.161,
    "args": {
      "External id": 553,      "cid": 370, "correlation": 169
    }
  },
  {
    "ph": "f", "id": 169, "pid": 870, "tid": 870, "ts": 7429770602726.752,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770602729.069, "dur": 0.158,
    "args": {
      "External id": 553,      "cid": 370, "correlation": 170
    }
  },
  {
    "ph": "f", "id": 170, "pid": 870, "tid": 870, "ts": 7429770602729.069,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602731.135, "dur": 5.321,
    "args": {
      "External id": 553,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 171,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 171, "pid": 870, "tid": 870, "ts": 7429770602731.135,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770602771.827, "dur": 0.188,
    "args": {
      "External id": 566,      "cid": 370, "correlation": 172
    }
  },
  {
    "ph": "f", "id": 172, "pid": 870, "tid": 870, "ts": 7429770602771.827,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770602774.984, "dur": 0.174,
    "args": {
      "External id": 566,      "cid": 370, "correlation": 173
    }
  },
  {
    "ph": "f", "id": 173, "pid": 870, "tid": 870, "ts": 7429770602774.984,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770602776.963, "dur": 0.149,
    "args": {
      "External id": 566,      "cid": 370, "correlation": 174
    }
  },
  {
    "ph": "f", "id": 174, "pid": 870, "tid": 870, "ts": 7429770602776.963,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602778.483, "dur": 5.380,
    "args": {
      "External id": 566,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 175,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 175, "pid": 870, "tid": 870, "ts": 7429770602778.483,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770602817.034, "dur": 0.165,
    "args": {
      "External id": 579,      "cid": 370, "correlation": 176
    }
  },
  {
    "ph": "f", "id": 176, "pid": 870, "tid": 870, "ts": 7429770602817.034,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770602820.074, "dur": 0.139,
    "args": {
      "External id": 579,      "cid": 370, "correlation": 177
    }
  },
  {
    "ph": "f", "id": 177, "pid": 870, "tid": 870, "ts": 7429770602820.074,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770602822.001, "dur": 0.151,
    "args": {
      "External id": 579,      "cid": 370, "correlation": 178
    }
  },
  {
    "ph": "f", "id": 178, "pid": 870, "tid": 870, "ts": 7429770602822.001,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602823.444, "dur": 5.165,
    "args": {
      "External id": 579,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 179,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 179, "pid": 870, "tid": 870, "ts": 7429770602823.444,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602851.387, "dur": 3.911,
    "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": 180,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 180, "pid": 870, "tid": 870, "ts": 7429770602851.387,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602878.395, "dur": 5.449,
    "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": 181,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 181, "pid": 870, "tid": 870, "ts": 7429770602878.395,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602898.121, "dur": 6.133,
    "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": 182,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 182, "pid": 870, "tid": 870, "ts": 7429770602898.121,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602911.350, "dur": 4.244,
    "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": 183,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 183, "pid": 870, "tid": 870, "ts": 7429770602911.350,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602921.131, "dur": 6.498,
    "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": 184,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 184, "pid": 870, "tid": 870, "ts": 7429770602921.131,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602933.928, "dur": 3.894,
    "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": 185,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 185, "pid": 870, "tid": 870, "ts": 7429770602933.928,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602951.096, "dur": 4.096,
    "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": 186,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 186, "pid": 870, "tid": 870, "ts": 7429770602951.096,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602961.676, "dur": 4.665,
    "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": 187,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 187, "pid": 870, "tid": 870, "ts": 7429770602961.676,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602972.414, "dur": 3.966,
    "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": 188,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 188, "pid": 870, "tid": 870, "ts": 7429770602972.414,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770602981.061, "dur": 3.552,
    "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": 189,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 189, "pid": 870, "tid": 870, "ts": 7429770602981.061,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603041.169, "dur": 6.809,
    "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": 190,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 190, "pid": 870, "tid": 870, "ts": 7429770603041.169,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603076.754, "dur": 4.327,
    "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": 191,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 191, "pid": 870, "tid": 870, "ts": 7429770603076.754,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770603115.954, "dur": 0.279,
    "args": {
      "External id": 654,      "cid": 370, "correlation": 192
    }
  },
  {
    "ph": "f", "id": 192, "pid": 870, "tid": 870, "ts": 7429770603115.954,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770603120.319, "dur": 0.170,
    "args": {
      "External id": 654,      "cid": 370, "correlation": 193
    }
  },
  {
    "ph": "f", "id": 193, "pid": 870, "tid": 870, "ts": 7429770603120.319,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770603123.481, "dur": 0.148,
    "args": {
      "External id": 654,      "cid": 370, "correlation": 194
    }
  },
  {
    "ph": "f", "id": 194, "pid": 870, "tid": 870, "ts": 7429770603123.481,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603126.044, "dur": 5.176,
    "args": {
      "External id": 654,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 195,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 195, "pid": 870, "tid": 870, "ts": 7429770603126.044,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603142.937, "dur": 4.673,
    "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": 196,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 196, "pid": 870, "tid": 870, "ts": 7429770603142.937,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603162.514, "dur": 4.542,
    "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": 197,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 197, "pid": 870, "tid": 870, "ts": 7429770603162.514,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603181.816, "dur": 6.108,
    "args": {
      "External id": 667,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 198,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 198, "pid": 870, "tid": 870, "ts": 7429770603181.816,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770603230.377, "dur": 0.232,
    "args": {
      "External id": 679,      "cid": 370, "correlation": 199
    }
  },
  {
    "ph": "f", "id": 199, "pid": 870, "tid": 870, "ts": 7429770603230.377,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770603233.496, "dur": 0.146,
    "args": {
      "External id": 679,      "cid": 370, "correlation": 200
    }
  },
  {
    "ph": "f", "id": 200, "pid": 870, "tid": 870, "ts": 7429770603233.496,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770603236.549, "dur": 0.146,
    "args": {
      "External id": 679,      "cid": 370, "correlation": 201
    }
  },
  {
    "ph": "f", "id": 201, "pid": 870, "tid": 870, "ts": 7429770603236.549,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603239.036, "dur": 7.116,
    "args": {
      "External id": 679,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 202,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 202, "pid": 870, "tid": 870, "ts": 7429770603239.036,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603262.479, "dur": 4.805,
    "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": 203,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 203, "pid": 870, "tid": 870, "ts": 7429770603262.479,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770603310.910, "dur": 0.279,
    "args": {
      "External id": 698,      "cid": 370, "correlation": 204
    }
  },
  {
    "ph": "f", "id": 204, "pid": 870, "tid": 870, "ts": 7429770603310.910,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770603314.056, "dur": 0.152,
    "args": {
      "External id": 698,      "cid": 370, "correlation": 205
    }
  },
  {
    "ph": "f", "id": 205, "pid": 870, "tid": 870, "ts": 7429770603314.056,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770603316.202, "dur": 0.366,
    "args": {
      "External id": 698,      "cid": 370, "correlation": 206
    }
  },
  {
    "ph": "f", "id": 206, "pid": 870, "tid": 870, "ts": 7429770603316.202,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603318.241, "dur": 5.663,
    "args": {
      "External id": 698,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 207,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 207, "pid": 870, "tid": 870, "ts": 7429770603318.241,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603336.023, "dur": 4.652,
    "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": 208,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 208, "pid": 870, "tid": 870, "ts": 7429770603336.023,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603358.978, "dur": 5.644,
    "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": 209,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 209, "pid": 870, "tid": 870, "ts": 7429770603358.978,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603374.236, "dur": 5.167,
    "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": 210,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 210, "pid": 870, "tid": 870, "ts": 7429770603374.236,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603388.297, "dur": 4.879,
    "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": 211,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 211, "pid": 870, "tid": 870, "ts": 7429770603388.297,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603398.615, "dur": 6.973,
    "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": 212,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 212, "pid": 870, "tid": 870, "ts": 7429770603398.615,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603412.527, "dur": 3.957,
    "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": 213,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 213, "pid": 870, "tid": 870, "ts": 7429770603412.527,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603424.918, "dur": 4.453,
    "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": 214,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 214, "pid": 870, "tid": 870, "ts": 7429770603424.918,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770603466.812, "dur": 0.184,
    "args": {
      "External id": 719,      "cid": 370, "correlation": 215
    }
  },
  {
    "ph": "f", "id": 215, "pid": 870, "tid": 870, "ts": 7429770603466.812,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770603469.971, "dur": 0.154,
    "args": {
      "External id": 719,      "cid": 370, "correlation": 216
    }
  },
  {
    "ph": "f", "id": 216, "pid": 870, "tid": 870, "ts": 7429770603469.971,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770603472.471, "dur": 0.139,
    "args": {
      "External id": 719,      "cid": 370, "correlation": 217
    }
  },
  {
    "ph": "f", "id": 217, "pid": 870, "tid": 870, "ts": 7429770603472.471,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603474.658, "dur": 5.483,
    "args": {
      "External id": 719,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 218,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 218, "pid": 870, "tid": 870, "ts": 7429770603474.658,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603497.412, "dur": 4.451,
    "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": 219,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 219, "pid": 870, "tid": 870, "ts": 7429770603497.412,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770603529.191, "dur": 0.176,
    "args": {
      "External id": 730,      "cid": 370, "correlation": 220
    }
  },
  {
    "ph": "f", "id": 220, "pid": 870, "tid": 870, "ts": 7429770603529.191,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770603532.427, "dur": 0.161,
    "args": {
      "External id": 730,      "cid": 370, "correlation": 221
    }
  },
  {
    "ph": "f", "id": 221, "pid": 870, "tid": 870, "ts": 7429770603532.427,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770603534.544, "dur": 0.140,
    "args": {
      "External id": 730,      "cid": 370, "correlation": 222
    }
  },
  {
    "ph": "f", "id": 222, "pid": 870, "tid": 870, "ts": 7429770603534.544,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603536.160, "dur": 6.364,
    "args": {
      "External id": 730,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 223,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 223, "pid": 870, "tid": 870, "ts": 7429770603536.160,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603551.914, "dur": 4.853,
    "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": 224,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 224, "pid": 870, "tid": 870, "ts": 7429770603551.914,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770603582.870, "dur": 0.177,
    "args": {
      "External id": 741,      "cid": 370, "correlation": 225
    }
  },
  {
    "ph": "f", "id": 225, "pid": 870, "tid": 870, "ts": 7429770603582.870,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770603586.083, "dur": 0.153,
    "args": {
      "External id": 741,      "cid": 370, "correlation": 226
    }
  },
  {
    "ph": "f", "id": 226, "pid": 870, "tid": 870, "ts": 7429770603586.083,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770603588.473, "dur": 0.143,
    "args": {
      "External id": 741,      "cid": 370, "correlation": 227
    }
  },
  {
    "ph": "f", "id": 227, "pid": 870, "tid": 870, "ts": 7429770603588.473,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603596.596, "dur": 5.056,
    "args": {
      "External id": 741,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 228,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 228, "pid": 870, "tid": 870, "ts": 7429770603596.596,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603602.252, "dur": 4.651,
    "args": {
      "External id": 741,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 229,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 229, "pid": 870, "tid": 870, "ts": 7429770603602.252,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603615.900, "dur": 4.049,
    "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": 230,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 230, "pid": 870, "tid": 870, "ts": 7429770603615.900,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603649.957, "dur": 3.972,
    "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": 231,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 231, "pid": 870, "tid": 870, "ts": 7429770603649.957,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603661.668, "dur": 4.794,
    "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": 232,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 232, "pid": 870, "tid": 870, "ts": 7429770603661.668,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603673.313, "dur": 4.581,
    "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": 233,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 233, "pid": 870, "tid": 870, "ts": 7429770603673.313,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603682.627, "dur": 3.791,
    "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": 234,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 234, "pid": 870, "tid": 870, "ts": 7429770603682.627,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603691.741, "dur": 6.473,
    "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": 235,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 235, "pid": 870, "tid": 870, "ts": 7429770603691.741,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603705.307, "dur": 3.677,
    "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": 236,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 236, "pid": 870, "tid": 870, "ts": 7429770603705.307,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770603750.669, "dur": 0.215,
    "args": {
      "External id": 762,      "cid": 370, "correlation": 237
    }
  },
  {
    "ph": "f", "id": 237, "pid": 870, "tid": 870, "ts": 7429770603750.669,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770603754.155, "dur": 0.138,
    "args": {
      "External id": 762,      "cid": 370, "correlation": 238
    }
  },
  {
    "ph": "f", "id": 238, "pid": 870, "tid": 870, "ts": 7429770603754.155,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770603756.118, "dur": 0.158,
    "args": {
      "External id": 762,      "cid": 370, "correlation": 239
    }
  },
  {
    "ph": "f", "id": 239, "pid": 870, "tid": 870, "ts": 7429770603756.118,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603758.216, "dur": 4.632,
    "args": {
      "External id": 762,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 240,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 240, "pid": 870, "tid": 870, "ts": 7429770603758.216,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770603801.013, "dur": 0.246,
    "args": {
      "External id": 775,      "cid": 370, "correlation": 241
    }
  },
  {
    "ph": "f", "id": 241, "pid": 870, "tid": 870, "ts": 7429770603801.013,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770603804.496, "dur": 0.141,
    "args": {
      "External id": 775,      "cid": 370, "correlation": 242
    }
  },
  {
    "ph": "f", "id": 242, "pid": 870, "tid": 870, "ts": 7429770603804.496,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770603806.404, "dur": 0.153,
    "args": {
      "External id": 775,      "cid": 370, "correlation": 243
    }
  },
  {
    "ph": "f", "id": 243, "pid": 870, "tid": 870, "ts": 7429770603806.404,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603808.022, "dur": 5.189,
    "args": {
      "External id": 775,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 244,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 244, "pid": 870, "tid": 870, "ts": 7429770603808.022,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770603849.165, "dur": 0.165,
    "args": {
      "External id": 788,      "cid": 370, "correlation": 245
    }
  },
  {
    "ph": "f", "id": 245, "pid": 870, "tid": 870, "ts": 7429770603849.165,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770603851.916, "dur": 0.141,
    "args": {
      "External id": 788,      "cid": 370, "correlation": 246
    }
  },
  {
    "ph": "f", "id": 246, "pid": 870, "tid": 870, "ts": 7429770603851.916,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770603853.854, "dur": 0.151,
    "args": {
      "External id": 788,      "cid": 370, "correlation": 247
    }
  },
  {
    "ph": "f", "id": 247, "pid": 870, "tid": 870, "ts": 7429770603853.854,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603855.285, "dur": 5.781,
    "args": {
      "External id": 788,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 248,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 248, "pid": 870, "tid": 870, "ts": 7429770603855.285,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603882.879, "dur": 4.819,
    "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": 249,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 249, "pid": 870, "tid": 870, "ts": 7429770603882.879,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603906.624, "dur": 6.551,
    "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": 250,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 250, "pid": 870, "tid": 870, "ts": 7429770603906.624,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603925.581, "dur": 6.367,
    "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": 251,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 251, "pid": 870, "tid": 870, "ts": 7429770603925.581,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603939.142, "dur": 5.701,
    "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": 252,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 252, "pid": 870, "tid": 870, "ts": 7429770603939.142,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603950.269, "dur": 4.362,
    "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": 253,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 253, "pid": 870, "tid": 870, "ts": 7429770603950.269,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603959.812, "dur": 5.184,
    "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": 254,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 254, "pid": 870, "tid": 870, "ts": 7429770603959.812,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603978.133, "dur": 3.854,
    "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": 255,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 255, "pid": 870, "tid": 870, "ts": 7429770603978.133,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770603989.720, "dur": 6.625,
    "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": 256,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 256, "pid": 870, "tid": 870, "ts": 7429770603989.720,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604001.958, "dur": 4.003,
    "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": 257,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 257, "pid": 870, "tid": 870, "ts": 7429770604001.958,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604013.668, "dur": 3.637,
    "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": 258,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 258, "pid": 870, "tid": 870, "ts": 7429770604013.668,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604069.375, "dur": 6.851,
    "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": 259,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 259, "pid": 870, "tid": 870, "ts": 7429770604069.375,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604106.589, "dur": 4.394,
    "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": 260,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 260, "pid": 870, "tid": 870, "ts": 7429770604106.589,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770604142.540, "dur": 0.231,
    "args": {
      "External id": 863,      "cid": 370, "correlation": 261
    }
  },
  {
    "ph": "f", "id": 261, "pid": 870, "tid": 870, "ts": 7429770604142.540,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770604146.847, "dur": 0.142,
    "args": {
      "External id": 863,      "cid": 370, "correlation": 262
    }
  },
  {
    "ph": "f", "id": 262, "pid": 870, "tid": 870, "ts": 7429770604146.847,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770604149.578, "dur": 0.142,
    "args": {
      "External id": 863,      "cid": 370, "correlation": 263
    }
  },
  {
    "ph": "f", "id": 263, "pid": 870, "tid": 870, "ts": 7429770604149.578,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604151.871, "dur": 4.642,
    "args": {
      "External id": 863,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 264,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 264, "pid": 870, "tid": 870, "ts": 7429770604151.871,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604167.602, "dur": 4.136,
    "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": 265,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 265, "pid": 870, "tid": 870, "ts": 7429770604167.602,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604186.221, "dur": 3.684,
    "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": 266,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 266, "pid": 870, "tid": 870, "ts": 7429770604186.221,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604200.175, "dur": 4.950,
    "args": {
      "External id": 876,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 267,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 267, "pid": 870, "tid": 870, "ts": 7429770604200.175,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770604245.919, "dur": 0.196,
    "args": {
      "External id": 888,      "cid": 370, "correlation": 268
    }
  },
  {
    "ph": "f", "id": 268, "pid": 870, "tid": 870, "ts": 7429770604245.919,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770604249.138, "dur": 0.141,
    "args": {
      "External id": 888,      "cid": 370, "correlation": 269
    }
  },
  {
    "ph": "f", "id": 269, "pid": 870, "tid": 870, "ts": 7429770604249.138,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770604251.506, "dur": 0.146,
    "args": {
      "External id": 888,      "cid": 370, "correlation": 270
    }
  },
  {
    "ph": "f", "id": 270, "pid": 870, "tid": 870, "ts": 7429770604251.506,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604253.622, "dur": 4.910,
    "args": {
      "External id": 888,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 271,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 271, "pid": 870, "tid": 870, "ts": 7429770604253.622,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604273.466, "dur": 4.515,
    "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": 272,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 272, "pid": 870, "tid": 870, "ts": 7429770604273.466,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770604320.157, "dur": 0.171,
    "args": {
      "External id": 907,      "cid": 370, "correlation": 273
    }
  },
  {
    "ph": "f", "id": 273, "pid": 870, "tid": 870, "ts": 7429770604320.157,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770604323.375, "dur": 0.153,
    "args": {
      "External id": 907,      "cid": 370, "correlation": 274
    }
  },
  {
    "ph": "f", "id": 274, "pid": 870, "tid": 870, "ts": 7429770604323.375,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770604325.541, "dur": 0.148,
    "args": {
      "External id": 907,      "cid": 370, "correlation": 275
    }
  },
  {
    "ph": "f", "id": 275, "pid": 870, "tid": 870, "ts": 7429770604325.541,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604327.394, "dur": 5.429,
    "args": {
      "External id": 907,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 276,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 276, "pid": 870, "tid": 870, "ts": 7429770604327.394,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604344.678, "dur": 4.679,
    "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": 277,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 277, "pid": 870, "tid": 870, "ts": 7429770604344.678,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604365.381, "dur": 4.433,
    "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": 278,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 278, "pid": 870, "tid": 870, "ts": 7429770604365.381,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604378.324, "dur": 6.159,
    "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": 279,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 279, "pid": 870, "tid": 870, "ts": 7429770604378.324,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604392.860, "dur": 4.010,
    "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": 280,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 280, "pid": 870, "tid": 870, "ts": 7429770604392.860,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604401.966, "dur": 4.968,
    "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": 281,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 281, "pid": 870, "tid": 870, "ts": 7429770604401.966,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604413.784, "dur": 4.156,
    "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": 282,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 282, "pid": 870, "tid": 870, "ts": 7429770604413.784,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604425.769, "dur": 5.642,
    "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": 283,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 283, "pid": 870, "tid": 870, "ts": 7429770604425.769,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770604468.442, "dur": 0.178,
    "args": {
      "External id": 928,      "cid": 370, "correlation": 284
    }
  },
  {
    "ph": "f", "id": 284, "pid": 870, "tid": 870, "ts": 7429770604468.442,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770604471.759, "dur": 0.143,
    "args": {
      "External id": 928,      "cid": 370, "correlation": 285
    }
  },
  {
    "ph": "f", "id": 285, "pid": 870, "tid": 870, "ts": 7429770604471.759,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770604474.123, "dur": 0.142,
    "args": {
      "External id": 928,      "cid": 370, "correlation": 286
    }
  },
  {
    "ph": "f", "id": 286, "pid": 870, "tid": 870, "ts": 7429770604474.123,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604476.059, "dur": 5.062,
    "args": {
      "External id": 928,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 287,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 287, "pid": 870, "tid": 870, "ts": 7429770604476.059,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604497.323, "dur": 4.121,
    "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": 288,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 288, "pid": 870, "tid": 870, "ts": 7429770604497.323,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770604528.000, "dur": 0.146,
    "args": {
      "External id": 939,      "cid": 370, "correlation": 289
    }
  },
  {
    "ph": "f", "id": 289, "pid": 870, "tid": 870, "ts": 7429770604528.000,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770604530.893, "dur": 0.140,
    "args": {
      "External id": 939,      "cid": 370, "correlation": 290
    }
  },
  {
    "ph": "f", "id": 290, "pid": 870, "tid": 870, "ts": 7429770604530.893,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770604533.153, "dur": 0.141,
    "args": {
      "External id": 939,      "cid": 370, "correlation": 291
    }
  },
  {
    "ph": "f", "id": 291, "pid": 870, "tid": 870, "ts": 7429770604533.153,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604534.746, "dur": 5.213,
    "args": {
      "External id": 939,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 292,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 292, "pid": 870, "tid": 870, "ts": 7429770604534.746,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604548.371, "dur": 4.465,
    "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": 293,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 293, "pid": 870, "tid": 870, "ts": 7429770604548.371,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770604575.122, "dur": 0.171,
    "args": {
      "External id": 950,      "cid": 370, "correlation": 294
    }
  },
  {
    "ph": "f", "id": 294, "pid": 870, "tid": 870, "ts": 7429770604575.122,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770604577.867, "dur": 0.137,
    "args": {
      "External id": 950,      "cid": 370, "correlation": 295
    }
  },
  {
    "ph": "f", "id": 295, "pid": 870, "tid": 870, "ts": 7429770604577.867,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770604580.353, "dur": 0.138,
    "args": {
      "External id": 950,      "cid": 370, "correlation": 296
    }
  },
  {
    "ph": "f", "id": 296, "pid": 870, "tid": 870, "ts": 7429770604580.353,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604587.897, "dur": 4.450,
    "args": {
      "External id": 950,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 297,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 297, "pid": 870, "tid": 870, "ts": 7429770604587.897,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604592.661, "dur": 3.564,
    "args": {
      "External id": 950,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 298,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 298, "pid": 870, "tid": 870, "ts": 7429770604592.661,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604604.810, "dur": 3.854,
    "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": 299,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 299, "pid": 870, "tid": 870, "ts": 7429770604604.810,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604632.398, "dur": 3.659,
    "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": 300,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 300, "pid": 870, "tid": 870, "ts": 7429770604632.398,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604644.468, "dur": 6.613,
    "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": 301,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 301, "pid": 870, "tid": 870, "ts": 7429770604644.468,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604657.899, "dur": 4.652,
    "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": 302,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 302, "pid": 870, "tid": 870, "ts": 7429770604657.899,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604666.981, "dur": 5.825,
    "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": 303,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 303, "pid": 870, "tid": 870, "ts": 7429770604666.981,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604679.307, "dur": 4.478,
    "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": 304,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 304, "pid": 870, "tid": 870, "ts": 7429770604679.307,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604690.482, "dur": 5.174,
    "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": 305,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 305, "pid": 870, "tid": 870, "ts": 7429770604690.482,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770604738.619, "dur": 0.319,
    "args": {
      "External id": 972,      "cid": 290, "correlation": 306
    }
  },
  {
    "ph": "f", "id": 306, "pid": 870, "tid": 870, "ts": 7429770604738.619,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770604739.657, "dur": 117.881,
    "args": {
      "External id": 972,            "bytes": 79691776, "cid": 108, "correlation": 307, "ptr": "0x100000001"
    }
  },
  {
    "ph": "s", "id": 307, "pid": 870, "tid": 870, "ts": 7429770604739.657,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770604860.966, "dur": 0.435,
    "args": {
      "External id": 971,      "cid": 370, "correlation": 308
    }
  },
  {
    "ph": "f", "id": 308, "pid": 870, "tid": 870, "ts": 7429770604860.966,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770604866.866, "dur": 0.165,
    "args": {
      "External id": 971,      "cid": 370, "correlation": 309
    }
  },
  {
    "ph": "f", "id": 309, "pid": 870, "tid": 870, "ts": 7429770604866.866,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770604869.857, "dur": 0.397,
    "args": {
      "External id": 971,      "cid": 370, "correlation": 310
    }
  },
  {
    "ph": "f", "id": 310, "pid": 870, "tid": 870, "ts": 7429770604869.857,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604874.154, "dur": 6.823,
    "args": {
      "External id": 971,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 311,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 311, "pid": 870, "tid": 870, "ts": 7429770604874.154,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770604939.301, "dur": 0.185,
    "args": {
      "External id": 984,      "cid": 370, "correlation": 312
    }
  },
  {
    "ph": "f", "id": 312, "pid": 870, "tid": 870, "ts": 7429770604939.301,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770604942.473, "dur": 0.141,
    "args": {
      "External id": 984,      "cid": 370, "correlation": 313
    }
  },
  {
    "ph": "f", "id": 313, "pid": 870, "tid": 870, "ts": 7429770604942.473,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770604944.809, "dur": 0.150,
    "args": {
      "External id": 984,      "cid": 370, "correlation": 314
    }
  },
  {
    "ph": "f", "id": 314, "pid": 870, "tid": 870, "ts": 7429770604944.809,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604946.813, "dur": 6.442,
    "args": {
      "External id": 984,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 315,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 315, "pid": 870, "tid": 870, "ts": 7429770604946.813,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770604991.196, "dur": 0.183,
    "args": {
      "External id": 997,      "cid": 370, "correlation": 316
    }
  },
  {
    "ph": "f", "id": 316, "pid": 870, "tid": 870, "ts": 7429770604991.196,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770604994.620, "dur": 0.157,
    "args": {
      "External id": 997,      "cid": 370, "correlation": 317
    }
  },
  {
    "ph": "f", "id": 317, "pid": 870, "tid": 870, "ts": 7429770604994.620,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770604996.553, "dur": 0.150,
    "args": {
      "External id": 997,      "cid": 370, "correlation": 318
    }
  },
  {
    "ph": "f", "id": 318, "pid": 870, "tid": 870, "ts": 7429770604996.553,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770604998.127, "dur": 5.481,
    "args": {
      "External id": 997,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 319,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 319, "pid": 870, "tid": 870, "ts": 7429770604998.127,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605031.029, "dur": 4.988,
    "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": 320,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 320, "pid": 870, "tid": 870, "ts": 7429770605031.029,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605061.114, "dur": 4.572,
    "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": 321,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 321, "pid": 870, "tid": 870, "ts": 7429770605061.114,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605083.227, "dur": 8.624,
    "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": 322,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 322, "pid": 870, "tid": 870, "ts": 7429770605083.227,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605099.377, "dur": 3.877,
    "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": 323,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 323, "pid": 870, "tid": 870, "ts": 7429770605099.377,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605110.901, "dur": 4.570,
    "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": 324,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 324, "pid": 870, "tid": 870, "ts": 7429770605110.901,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605122.396, "dur": 4.350,
    "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": 325,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 325, "pid": 870, "tid": 870, "ts": 7429770605122.396,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605139.296, "dur": 4.943,
    "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": 326,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 326, "pid": 870, "tid": 870, "ts": 7429770605139.296,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605150.336, "dur": 4.357,
    "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": 327,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 327, "pid": 870, "tid": 870, "ts": 7429770605150.336,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605160.751, "dur": 4.130,
    "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": 328,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 328, "pid": 870, "tid": 870, "ts": 7429770605160.751,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605169.365, "dur": 2.979,
    "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": 329,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 329, "pid": 870, "tid": 870, "ts": 7429770605169.365,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605269.022, "dur": 4.709,
    "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": 330,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 330, "pid": 870, "tid": 870, "ts": 7429770605269.022,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605303.722, "dur": 4.083,
    "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": 331,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 331, "pid": 870, "tid": 870, "ts": 7429770605303.722,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770605344.424, "dur": 0.294,
    "args": {
      "External id": 1072,      "cid": 370, "correlation": 332
    }
  },
  {
    "ph": "f", "id": 332, "pid": 870, "tid": 870, "ts": 7429770605344.424,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770605348.827, "dur": 0.191,
    "args": {
      "External id": 1072,      "cid": 370, "correlation": 333
    }
  },
  {
    "ph": "f", "id": 333, "pid": 870, "tid": 870, "ts": 7429770605348.827,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770605352.086, "dur": 0.237,
    "args": {
      "External id": 1072,      "cid": 370, "correlation": 334
    }
  },
  {
    "ph": "f", "id": 334, "pid": 870, "tid": 870, "ts": 7429770605352.086,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605354.992, "dur": 5.235,
    "args": {
      "External id": 1072,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 335,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 335, "pid": 870, "tid": 870, "ts": 7429770605354.992,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605373.601, "dur": 4.175,
    "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": 336,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 336, "pid": 870, "tid": 870, "ts": 7429770605373.601,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605392.491, "dur": 3.725,
    "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": 337,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 337, "pid": 870, "tid": 870, "ts": 7429770605392.491,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605409.449, "dur": 5.035,
    "args": {
      "External id": 1085,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 338,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 338, "pid": 870, "tid": 870, "ts": 7429770605409.449,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770605459.974, "dur": 0.191,
    "args": {
      "External id": 1097,      "cid": 370, "correlation": 339
    }
  },
  {
    "ph": "f", "id": 339, "pid": 870, "tid": 870, "ts": 7429770605459.974,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770605463.633, "dur": 0.154,
    "args": {
      "External id": 1097,      "cid": 370, "correlation": 340
    }
  },
  {
    "ph": "f", "id": 340, "pid": 870, "tid": 870, "ts": 7429770605463.633,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770605466.653, "dur": 0.178,
    "args": {
      "External id": 1097,      "cid": 370, "correlation": 341
    }
  },
  {
    "ph": "f", "id": 341, "pid": 870, "tid": 870, "ts": 7429770605466.653,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605469.046, "dur": 6.016,
    "args": {
      "External id": 1097,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 342,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 342, "pid": 870, "tid": 870, "ts": 7429770605469.046,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605491.663, "dur": 4.185,
    "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": 343,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 343, "pid": 870, "tid": 870, "ts": 7429770605491.663,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770605541.983, "dur": 0.221,
    "args": {
      "External id": 1116,      "cid": 370, "correlation": 344
    }
  },
  {
    "ph": "f", "id": 344, "pid": 870, "tid": 870, "ts": 7429770605541.983,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770605544.768, "dur": 0.163,
    "args": {
      "External id": 1116,      "cid": 370, "correlation": 345
    }
  },
  {
    "ph": "f", "id": 345, "pid": 870, "tid": 870, "ts": 7429770605544.768,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770605546.901, "dur": 0.439,
    "args": {
      "External id": 1116,      "cid": 370, "correlation": 346
    }
  },
  {
    "ph": "f", "id": 346, "pid": 870, "tid": 870, "ts": 7429770605546.901,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605548.997, "dur": 4.702,
    "args": {
      "External id": 1116,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 347,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 347, "pid": 870, "tid": 870, "ts": 7429770605548.997,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605569.163, "dur": 5.099,
    "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": 348,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 348, "pid": 870, "tid": 870, "ts": 7429770605569.163,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605595.476, "dur": 6.911,
    "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": 349,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 349, "pid": 870, "tid": 870, "ts": 7429770605595.476,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605614.294, "dur": 7.527,
    "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": 350,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 350, "pid": 870, "tid": 870, "ts": 7429770605614.294,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605630.729, "dur": 4.322,
    "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": 351,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 351, "pid": 870, "tid": 870, "ts": 7429770605630.729,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605642.015, "dur": 5.715,
    "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": 352,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 352, "pid": 870, "tid": 870, "ts": 7429770605642.015,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605655.293, "dur": 4.303,
    "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": 353,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 353, "pid": 870, "tid": 870, "ts": 7429770605655.293,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605667.853, "dur": 5.431,
    "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": 354,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 354, "pid": 870, "tid": 870, "ts": 7429770605667.853,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770605712.954, "dur": 0.212,
    "args": {
      "External id": 1137,      "cid": 370, "correlation": 355
    }
  },
  {
    "ph": "f", "id": 355, "pid": 870, "tid": 870, "ts": 7429770605712.954,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770605716.677, "dur": 0.160,
    "args": {
      "External id": 1137,      "cid": 370, "correlation": 356
    }
  },
  {
    "ph": "f", "id": 356, "pid": 870, "tid": 870, "ts": 7429770605716.677,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770605719.277, "dur": 0.150,
    "args": {
      "External id": 1137,      "cid": 370, "correlation": 357
    }
  },
  {
    "ph": "f", "id": 357, "pid": 870, "tid": 870, "ts": 7429770605719.277,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605721.392, "dur": 5.227,
    "args": {
      "External id": 1137,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 358,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 358, "pid": 870, "tid": 870, "ts": 7429770605721.392,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605743.476, "dur": 5.268,
    "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": 359,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 359, "pid": 870, "tid": 870, "ts": 7429770605743.476,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770605776.846, "dur": 0.169,
    "args": {
      "External id": 1148,      "cid": 370, "correlation": 360
    }
  },
  {
    "ph": "f", "id": 360, "pid": 870, "tid": 870, "ts": 7429770605776.846,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770605779.657, "dur": 0.141,
    "args": {
      "External id": 1148,      "cid": 370, "correlation": 361
    }
  },
  {
    "ph": "f", "id": 361, "pid": 870, "tid": 870, "ts": 7429770605779.657,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770605781.903, "dur": 0.143,
    "args": {
      "External id": 1148,      "cid": 370, "correlation": 362
    }
  },
  {
    "ph": "f", "id": 362, "pid": 870, "tid": 870, "ts": 7429770605781.903,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605783.608, "dur": 4.980,
    "args": {
      "External id": 1148,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 363,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 363, "pid": 870, "tid": 870, "ts": 7429770605783.608,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605797.206, "dur": 5.327,
    "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": 364,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 364, "pid": 870, "tid": 870, "ts": 7429770605797.206,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770605827.187, "dur": 0.161,
    "args": {
      "External id": 1159,      "cid": 370, "correlation": 365
    }
  },
  {
    "ph": "f", "id": 365, "pid": 870, "tid": 870, "ts": 7429770605827.187,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770605829.954, "dur": 0.141,
    "args": {
      "External id": 1159,      "cid": 370, "correlation": 366
    }
  },
  {
    "ph": "f", "id": 366, "pid": 870, "tid": 870, "ts": 7429770605829.954,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770605832.294, "dur": 0.142,
    "args": {
      "External id": 1159,      "cid": 370, "correlation": 367
    }
  },
  {
    "ph": "f", "id": 367, "pid": 870, "tid": 870, "ts": 7429770605832.294,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605841.435, "dur": 5.518,
    "args": {
      "External id": 1159,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 368,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 368, "pid": 870, "tid": 870, "ts": 7429770605841.435,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605847.341, "dur": 5.599,
    "args": {
      "External id": 1159,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 369,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 369, "pid": 870, "tid": 870, "ts": 7429770605847.341,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605861.799, "dur": 4.380,
    "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": 370,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 370, "pid": 870, "tid": 870, "ts": 7429770605861.799,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605893.312, "dur": 3.911,
    "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": 371,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 371, "pid": 870, "tid": 870, "ts": 7429770605893.312,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605905.127, "dur": 5.104,
    "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": 372,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 372, "pid": 870, "tid": 870, "ts": 7429770605905.127,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605917.124, "dur": 4.634,
    "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": 373,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 373, "pid": 870, "tid": 870, "ts": 7429770605917.124,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605927.778, "dur": 4.010,
    "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": 374,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 374, "pid": 870, "tid": 870, "ts": 7429770605927.778,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605937.326, "dur": 3.896,
    "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": 375,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 375, "pid": 870, "tid": 870, "ts": 7429770605937.326,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770605949.670, "dur": 4.035,
    "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": 376,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 376, "pid": 870, "tid": 870, "ts": 7429770605949.670,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770605992.900, "dur": 0.189,
    "args": {
      "External id": 1180,      "cid": 370, "correlation": 377
    }
  },
  {
    "ph": "f", "id": 377, "pid": 870, "tid": 870, "ts": 7429770605992.900,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770605996.195, "dur": 0.152,
    "args": {
      "External id": 1180,      "cid": 370, "correlation": 378
    }
  },
  {
    "ph": "f", "id": 378, "pid": 870, "tid": 870, "ts": 7429770605996.195,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770605998.306, "dur": 0.147,
    "args": {
      "External id": 1180,      "cid": 370, "correlation": 379
    }
  },
  {
    "ph": "f", "id": 379, "pid": 870, "tid": 870, "ts": 7429770605998.306,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606000.427, "dur": 5.290,
    "args": {
      "External id": 1180,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 380,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 380, "pid": 870, "tid": 870, "ts": 7429770606000.427,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770606040.000, "dur": 0.178,
    "args": {
      "External id": 1193,      "cid": 370, "correlation": 381
    }
  },
  {
    "ph": "f", "id": 381, "pid": 870, "tid": 870, "ts": 7429770606040.000,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770606042.952, "dur": 0.167,
    "args": {
      "External id": 1193,      "cid": 370, "correlation": 382
    }
  },
  {
    "ph": "f", "id": 382, "pid": 870, "tid": 870, "ts": 7429770606042.952,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770606044.943, "dur": 0.139,
    "args": {
      "External id": 1193,      "cid": 370, "correlation": 383
    }
  },
  {
    "ph": "f", "id": 383, "pid": 870, "tid": 870, "ts": 7429770606044.943,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606047.001, "dur": 5.429,
    "args": {
      "External id": 1193,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 384,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 384, "pid": 870, "tid": 870, "ts": 7429770606047.001,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770606110.510, "dur": 0.285,
    "args": {
      "External id": 1206,      "cid": 370, "correlation": 385
    }
  },
  {
    "ph": "f", "id": 385, "pid": 870, "tid": 870, "ts": 7429770606110.510,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770606114.183, "dur": 0.137,
    "args": {
      "External id": 1206,      "cid": 370, "correlation": 386
    }
  },
  {
    "ph": "f", "id": 386, "pid": 870, "tid": 870, "ts": 7429770606114.183,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770606116.430, "dur": 0.155,
    "args": {
      "External id": 1206,      "cid": 370, "correlation": 387
    }
  },
  {
    "ph": "f", "id": 387, "pid": 870, "tid": 870, "ts": 7429770606116.430,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606118.440, "dur": 6.666,
    "args": {
      "External id": 1206,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 388,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 388, "pid": 870, "tid": 870, "ts": 7429770606118.440,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606151.045, "dur": 4.292,
    "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": 389,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 389, "pid": 870, "tid": 870, "ts": 7429770606151.045,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606177.746, "dur": 5.050,
    "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": 390,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 390, "pid": 870, "tid": 870, "ts": 7429770606177.746,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606196.028, "dur": 6.732,
    "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": 391,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 391, "pid": 870, "tid": 870, "ts": 7429770606196.028,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606210.449, "dur": 5.138,
    "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": 392,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 392, "pid": 870, "tid": 870, "ts": 7429770606210.449,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606221.011, "dur": 5.557,
    "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": 393,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 393, "pid": 870, "tid": 870, "ts": 7429770606221.011,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606232.199, "dur": 4.939,
    "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": 394,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 394, "pid": 870, "tid": 870, "ts": 7429770606232.199,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606249.886, "dur": 4.428,
    "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": 395,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 395, "pid": 870, "tid": 870, "ts": 7429770606249.886,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606262.088, "dur": 5.658,
    "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": 396,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 396, "pid": 870, "tid": 870, "ts": 7429770606262.088,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606273.105, "dur": 3.617,
    "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": 397,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 397, "pid": 870, "tid": 870, "ts": 7429770606273.105,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606281.246, "dur": 3.904,
    "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": 398,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 398, "pid": 870, "tid": 870, "ts": 7429770606281.246,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606344.502, "dur": 4.095,
    "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": 399,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 399, "pid": 870, "tid": 870, "ts": 7429770606344.502,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606378.109, "dur": 6.048,
    "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": 400,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 400, "pid": 870, "tid": 870, "ts": 7429770606378.109,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770606420.593, "dur": 0.236,
    "args": {
      "External id": 1281,      "cid": 370, "correlation": 401
    }
  },
  {
    "ph": "f", "id": 401, "pid": 870, "tid": 870, "ts": 7429770606420.593,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770606424.655, "dur": 0.185,
    "args": {
      "External id": 1281,      "cid": 370, "correlation": 402
    }
  },
  {
    "ph": "f", "id": 402, "pid": 870, "tid": 870, "ts": 7429770606424.655,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770606427.467, "dur": 0.171,
    "args": {
      "External id": 1281,      "cid": 370, "correlation": 403
    }
  },
  {
    "ph": "f", "id": 403, "pid": 870, "tid": 870, "ts": 7429770606427.467,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606430.230, "dur": 5.636,
    "args": {
      "External id": 1281,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 404,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 404, "pid": 870, "tid": 870, "ts": 7429770606430.230,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606447.466, "dur": 4.608,
    "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": 405,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 405, "pid": 870, "tid": 870, "ts": 7429770606447.466,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606465.805, "dur": 3.985,
    "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": 406,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 406, "pid": 870, "tid": 870, "ts": 7429770606465.805,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606483.114, "dur": 5.446,
    "args": {
      "External id": 1294,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 407,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 407, "pid": 870, "tid": 870, "ts": 7429770606483.114,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770606531.883, "dur": 0.183,
    "args": {
      "External id": 1306,      "cid": 370, "correlation": 408
    }
  },
  {
    "ph": "f", "id": 408, "pid": 870, "tid": 870, "ts": 7429770606531.883,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770606535.014, "dur": 0.140,
    "args": {
      "External id": 1306,      "cid": 370, "correlation": 409
    }
  },
  {
    "ph": "f", "id": 409, "pid": 870, "tid": 870, "ts": 7429770606535.014,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770606537.607, "dur": 0.141,
    "args": {
      "External id": 1306,      "cid": 370, "correlation": 410
    }
  },
  {
    "ph": "f", "id": 410, "pid": 870, "tid": 870, "ts": 7429770606537.607,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606539.687, "dur": 5.739,
    "args": {
      "External id": 1306,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 411,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 411, "pid": 870, "tid": 870, "ts": 7429770606539.687,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606561.722, "dur": 5.021,
    "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": 412,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 412, "pid": 870, "tid": 870, "ts": 7429770606561.722,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770606608.504, "dur": 0.219,
    "args": {
      "External id": 1325,      "cid": 370, "correlation": 413
    }
  },
  {
    "ph": "f", "id": 413, "pid": 870, "tid": 870, "ts": 7429770606608.504,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770606611.218, "dur": 0.183,
    "args": {
      "External id": 1325,      "cid": 370, "correlation": 414
    }
  },
  {
    "ph": "f", "id": 414, "pid": 870, "tid": 870, "ts": 7429770606611.218,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770606613.518, "dur": 0.149,
    "args": {
      "External id": 1325,      "cid": 370, "correlation": 415
    }
  },
  {
    "ph": "f", "id": 415, "pid": 870, "tid": 870, "ts": 7429770606613.518,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606615.530, "dur": 5.238,
    "args": {
      "External id": 1325,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 416,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 416, "pid": 870, "tid": 870, "ts": 7429770606615.530,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606632.821, "dur": 4.132,
    "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": 417,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 417, "pid": 870, "tid": 870, "ts": 7429770606632.821,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606655.428, "dur": 4.400,
    "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": 418,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 418, "pid": 870, "tid": 870, "ts": 7429770606655.428,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606669.423, "dur": 6.543,
    "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": 419,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 419, "pid": 870, "tid": 870, "ts": 7429770606669.423,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606683.701, "dur": 4.855,
    "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": 420,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 420, "pid": 870, "tid": 870, "ts": 7429770606683.701,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606694.079, "dur": 5.543,
    "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": 421,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 421, "pid": 870, "tid": 870, "ts": 7429770606694.079,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606707.168, "dur": 4.134,
    "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": 422,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 422, "pid": 870, "tid": 870, "ts": 7429770606707.168,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606719.102, "dur": 4.537,
    "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": 423,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 423, "pid": 870, "tid": 870, "ts": 7429770606719.102,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770606758.024, "dur": 0.190,
    "args": {
      "External id": 1346,      "cid": 370, "correlation": 424
    }
  },
  {
    "ph": "f", "id": 424, "pid": 870, "tid": 870, "ts": 7429770606758.024,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770606761.279, "dur": 0.182,
    "args": {
      "External id": 1346,      "cid": 370, "correlation": 425
    }
  },
  {
    "ph": "f", "id": 425, "pid": 870, "tid": 870, "ts": 7429770606761.279,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770606763.547, "dur": 0.150,
    "args": {
      "External id": 1346,      "cid": 370, "correlation": 426
    }
  },
  {
    "ph": "f", "id": 426, "pid": 870, "tid": 870, "ts": 7429770606763.547,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606765.484, "dur": 4.721,
    "args": {
      "External id": 1346,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 427,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 427, "pid": 870, "tid": 870, "ts": 7429770606765.484,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606786.953, "dur": 4.528,
    "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": 428,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 428, "pid": 870, "tid": 870, "ts": 7429770606786.953,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770606818.839, "dur": 0.152,
    "args": {
      "External id": 1357,      "cid": 370, "correlation": 429
    }
  },
  {
    "ph": "f", "id": 429, "pid": 870, "tid": 870, "ts": 7429770606818.839,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770606821.584, "dur": 0.140,
    "args": {
      "External id": 1357,      "cid": 370, "correlation": 430
    }
  },
  {
    "ph": "f", "id": 430, "pid": 870, "tid": 870, "ts": 7429770606821.584,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770606823.624, "dur": 0.140,
    "args": {
      "External id": 1357,      "cid": 370, "correlation": 431
    }
  },
  {
    "ph": "f", "id": 431, "pid": 870, "tid": 870, "ts": 7429770606823.624,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606825.590, "dur": 5.766,
    "args": {
      "External id": 1357,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 432,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 432, "pid": 870, "tid": 870, "ts": 7429770606825.590,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606839.686, "dur": 5.084,
    "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": 433,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 433, "pid": 870, "tid": 870, "ts": 7429770606839.686,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770606868.106, "dur": 0.174,
    "args": {
      "External id": 1368,      "cid": 370, "correlation": 434
    }
  },
  {
    "ph": "f", "id": 434, "pid": 870, "tid": 870, "ts": 7429770606868.106,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770606870.958, "dur": 0.137,
    "args": {
      "External id": 1368,      "cid": 370, "correlation": 435
    }
  },
  {
    "ph": "f", "id": 435, "pid": 870, "tid": 870, "ts": 7429770606870.958,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770606873.324, "dur": 0.140,
    "args": {
      "External id": 1368,      "cid": 370, "correlation": 436
    }
  },
  {
    "ph": "f", "id": 436, "pid": 870, "tid": 870, "ts": 7429770606873.324,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606882.231, "dur": 4.783,
    "args": {
      "External id": 1368,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 437,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 437, "pid": 870, "tid": 870, "ts": 7429770606882.231,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606887.365, "dur": 3.669,
    "args": {
      "External id": 1368,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 438,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 438, "pid": 870, "tid": 870, "ts": 7429770606887.365,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606899.972, "dur": 4.506,
    "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": 439,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 439, "pid": 870, "tid": 870, "ts": 7429770606899.972,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606930.500, "dur": 6.591,
    "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": 440,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 440, "pid": 870, "tid": 870, "ts": 7429770606930.500,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606944.934, "dur": 5.376,
    "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": 441,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 441, "pid": 870, "tid": 870, "ts": 7429770606944.934,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606957.646, "dur": 4.986,
    "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": 442,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 442, "pid": 870, "tid": 870, "ts": 7429770606957.646,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606967.406, "dur": 3.922,
    "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": 443,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 443, "pid": 870, "tid": 870, "ts": 7429770606967.406,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606977.946, "dur": 4.152,
    "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": 444,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 444, "pid": 870, "tid": 870, "ts": 7429770606977.946,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770606989.494, "dur": 5.463,
    "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": 445,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 445, "pid": 870, "tid": 870, "ts": 7429770606989.494,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770607036.700, "dur": 0.207,
    "args": {
      "External id": 1389,      "cid": 370, "correlation": 446
    }
  },
  {
    "ph": "f", "id": 446, "pid": 870, "tid": 870, "ts": 7429770607036.700,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770607041.038, "dur": 0.157,
    "args": {
      "External id": 1389,      "cid": 370, "correlation": 447
    }
  },
  {
    "ph": "f", "id": 447, "pid": 870, "tid": 870, "ts": 7429770607041.038,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770607044.079, "dur": 0.149,
    "args": {
      "External id": 1389,      "cid": 370, "correlation": 448
    }
  },
  {
    "ph": "f", "id": 448, "pid": 870, "tid": 870, "ts": 7429770607044.079,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607046.628, "dur": 4.874,
    "args": {
      "External id": 1389,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 449,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 449, "pid": 870, "tid": 870, "ts": 7429770607046.628,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770607087.399, "dur": 0.175,
    "args": {
      "External id": 1402,      "cid": 370, "correlation": 450
    }
  },
  {
    "ph": "f", "id": 450, "pid": 870, "tid": 870, "ts": 7429770607087.399,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770607090.856, "dur": 0.151,
    "args": {
      "External id": 1402,      "cid": 370, "correlation": 451
    }
  },
  {
    "ph": "f", "id": 451, "pid": 870, "tid": 870, "ts": 7429770607090.856,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770607093.318, "dur": 0.137,
    "args": {
      "External id": 1402,      "cid": 370, "correlation": 452
    }
  },
  {
    "ph": "f", "id": 452, "pid": 870, "tid": 870, "ts": 7429770607093.318,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607094.953, "dur": 4.530,
    "args": {
      "External id": 1402,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 453,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 453, "pid": 870, "tid": 870, "ts": 7429770607094.953,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770607131.563, "dur": 0.190,
    "args": {
      "External id": 1415,      "cid": 370, "correlation": 454
    }
  },
  {
    "ph": "f", "id": 454, "pid": 870, "tid": 870, "ts": 7429770607131.563,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770607134.972, "dur": 0.141,
    "args": {
      "External id": 1415,      "cid": 370, "correlation": 455
    }
  },
  {
    "ph": "f", "id": 455, "pid": 870, "tid": 870, "ts": 7429770607134.972,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770607137.261, "dur": 0.137,
    "args": {
      "External id": 1415,      "cid": 370, "correlation": 456
    }
  },
  {
    "ph": "f", "id": 456, "pid": 870, "tid": 870, "ts": 7429770607137.261,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607138.776, "dur": 5.331,
    "args": {
      "External id": 1415,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 457,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 457, "pid": 870, "tid": 870, "ts": 7429770607138.776,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607166.528, "dur": 3.770,
    "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": 458,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 458, "pid": 870, "tid": 870, "ts": 7429770607166.528,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607189.528, "dur": 4.316,
    "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": 459,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 459, "pid": 870, "tid": 870, "ts": 7429770607189.528,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607205.499, "dur": 5.991,
    "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": 460,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 460, "pid": 870, "tid": 870, "ts": 7429770607205.499,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607219.303, "dur": 4.487,
    "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": 461,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 461, "pid": 870, "tid": 870, "ts": 7429770607219.303,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607229.105, "dur": 4.835,
    "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": 462,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 462, "pid": 870, "tid": 870, "ts": 7429770607229.105,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607239.849, "dur": 5.667,
    "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": 463,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 463, "pid": 870, "tid": 870, "ts": 7429770607239.849,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607257.993, "dur": 4.420,
    "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": 464,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 464, "pid": 870, "tid": 870, "ts": 7429770607257.993,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607269.544, "dur": 5.165,
    "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": 465,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 465, "pid": 870, "tid": 870, "ts": 7429770607269.544,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607280.559, "dur": 3.629,
    "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": 466,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 466, "pid": 870, "tid": 870, "ts": 7429770607280.559,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607290.219, "dur": 4.102,
    "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": 467,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 467, "pid": 870, "tid": 870, "ts": 7429770607290.219,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607350.254, "dur": 4.699,
    "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": 468,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 468, "pid": 870, "tid": 870, "ts": 7429770607350.254,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607385.655, "dur": 6.117,
    "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": 469,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 469, "pid": 870, "tid": 870, "ts": 7429770607385.655,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770607426.160, "dur": 0.265,
    "args": {
      "External id": 1490,      "cid": 370, "correlation": 470
    }
  },
  {
    "ph": "f", "id": 470, "pid": 870, "tid": 870, "ts": 7429770607426.160,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770607430.307, "dur": 0.175,
    "args": {
      "External id": 1490,      "cid": 370, "correlation": 471
    }
  },
  {
    "ph": "f", "id": 471, "pid": 870, "tid": 870, "ts": 7429770607430.307,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770607433.509, "dur": 0.151,
    "args": {
      "External id": 1490,      "cid": 370, "correlation": 472
    }
  },
  {
    "ph": "f", "id": 472, "pid": 870, "tid": 870, "ts": 7429770607433.509,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607436.094, "dur": 5.966,
    "args": {
      "External id": 1490,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 473,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 473, "pid": 870, "tid": 870, "ts": 7429770607436.094,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607453.606, "dur": 4.055,
    "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": 474,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 474, "pid": 870, "tid": 870, "ts": 7429770607453.606,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607472.450, "dur": 3.848,
    "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": 475,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 475, "pid": 870, "tid": 870, "ts": 7429770607472.450,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607488.557, "dur": 4.742,
    "args": {
      "External id": 1503,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 476,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 476, "pid": 870, "tid": 870, "ts": 7429770607488.557,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770607533.028, "dur": 0.204,
    "args": {
      "External id": 1515,      "cid": 370, "correlation": 477
    }
  },
  {
    "ph": "f", "id": 477, "pid": 870, "tid": 870, "ts": 7429770607533.028,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770607536.595, "dur": 0.142,
    "args": {
      "External id": 1515,      "cid": 370, "correlation": 478
    }
  },
  {
    "ph": "f", "id": 478, "pid": 870, "tid": 870, "ts": 7429770607536.595,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770607539.028, "dur": 0.145,
    "args": {
      "External id": 1515,      "cid": 370, "correlation": 479
    }
  },
  {
    "ph": "f", "id": 479, "pid": 870, "tid": 870, "ts": 7429770607539.028,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607541.535, "dur": 4.987,
    "args": {
      "External id": 1515,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 480,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 480, "pid": 870, "tid": 870, "ts": 7429770607541.535,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607562.198, "dur": 4.455,
    "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": 481,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 481, "pid": 870, "tid": 870, "ts": 7429770607562.198,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770607608.598, "dur": 0.283,
    "args": {
      "External id": 1534,      "cid": 370, "correlation": 482
    }
  },
  {
    "ph": "f", "id": 482, "pid": 870, "tid": 870, "ts": 7429770607608.598,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770607612.240, "dur": 0.153,
    "args": {
      "External id": 1534,      "cid": 370, "correlation": 483
    }
  },
  {
    "ph": "f", "id": 483, "pid": 870, "tid": 870, "ts": 7429770607612.240,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770607614.727, "dur": 0.365,
    "args": {
      "External id": 1534,      "cid": 370, "correlation": 484
    }
  },
  {
    "ph": "f", "id": 484, "pid": 870, "tid": 870, "ts": 7429770607614.727,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607616.741, "dur": 7.203,
    "args": {
      "External id": 1534,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 485,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 485, "pid": 870, "tid": 870, "ts": 7429770607616.741,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607666.719, "dur": 5.415,
    "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": 486,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 486, "pid": 870, "tid": 870, "ts": 7429770607666.719,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607689.278, "dur": 4.191,
    "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": 487,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 487, "pid": 870, "tid": 870, "ts": 7429770607689.278,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607703.233, "dur": 4.198,
    "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": 488,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 488, "pid": 870, "tid": 870, "ts": 7429770607703.233,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607715.403, "dur": 3.841,
    "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": 489,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 489, "pid": 870, "tid": 870, "ts": 7429770607715.403,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607724.649, "dur": 5.069,
    "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": 490,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 490, "pid": 870, "tid": 870, "ts": 7429770607724.649,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607736.557, "dur": 4.248,
    "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": 491,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 491, "pid": 870, "tid": 870, "ts": 7429770607736.557,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607748.076, "dur": 4.107,
    "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": 492,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 492, "pid": 870, "tid": 870, "ts": 7429770607748.076,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770607792.265, "dur": 0.383,
    "args": {
      "External id": 1556,      "cid": 290, "correlation": 493
    }
  },
  {
    "ph": "f", "id": 493, "pid": 870, "tid": 870, "ts": 7429770607792.265,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770607793.302, "dur": 75.961,
    "args": {
      "External id": 1556,            "bytes": 79691776, "cid": 108, "correlation": 494, "ptr": "0x7ffd8041f0a8"
    }
  },
  {
    "ph": "s", "id": 494, "pid": 870, "tid": 870, "ts": 7429770607793.302,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770607871.904, "dur": 0.264,
    "args": {
      "External id": 1555,      "cid": 370, "correlation": 495
    }
  },
  {
    "ph": "f", "id": 495, "pid": 870, "tid": 870, "ts": 7429770607871.904,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770607876.820, "dur": 0.171,
    "args": {
      "External id": 1555,      "cid": 370, "correlation": 496
    }
  },
  {
    "ph": "f", "id": 496, "pid": 870, "tid": 870, "ts": 7429770607876.820,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770607879.987, "dur": 0.371,
    "args": {
      "External id": 1555,      "cid": 370, "correlation": 497
    }
  },
  {
    "ph": "f", "id": 497, "pid": 870, "tid": 870, "ts": 7429770607879.987,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607883.865, "dur": 7.283,
    "args": {
      "External id": 1555,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 498,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 498, "pid": 870, "tid": 870, "ts": 7429770607883.865,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607912.283, "dur": 5.136,
    "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": 499,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 499, "pid": 870, "tid": 870, "ts": 7429770607912.283,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770607948.474, "dur": 0.178,
    "args": {
      "External id": 1566,      "cid": 370, "correlation": 500
    }
  },
  {
    "ph": "f", "id": 500, "pid": 870, "tid": 870, "ts": 7429770607948.474,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770607952.019, "dur": 0.144,
    "args": {
      "External id": 1566,      "cid": 370, "correlation": 501
    }
  },
  {
    "ph": "f", "id": 501, "pid": 870, "tid": 870, "ts": 7429770607952.019,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770607954.414, "dur": 0.138,
    "args": {
      "External id": 1566,      "cid": 370, "correlation": 502
    }
  },
  {
    "ph": "f", "id": 502, "pid": 870, "tid": 870, "ts": 7429770607954.414,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607956.262, "dur": 4.885,
    "args": {
      "External id": 1566,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 503,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 503, "pid": 870, "tid": 870, "ts": 7429770607956.262,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770607971.002, "dur": 4.685,
    "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": 504,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 504, "pid": 870, "tid": 870, "ts": 7429770607971.002,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770608000.732, "dur": 0.172,
    "args": {
      "External id": 1577,      "cid": 370, "correlation": 505
    }
  },
  {
    "ph": "f", "id": 505, "pid": 870, "tid": 870, "ts": 7429770608000.732,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770608003.779, "dur": 0.146,
    "args": {
      "External id": 1577,      "cid": 370, "correlation": 506
    }
  },
  {
    "ph": "f", "id": 506, "pid": 870, "tid": 870, "ts": 7429770608003.779,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770608006.176, "dur": 0.140,
    "args": {
      "External id": 1577,      "cid": 370, "correlation": 507
    }
  },
  {
    "ph": "f", "id": 507, "pid": 870, "tid": 870, "ts": 7429770608006.176,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608015.144, "dur": 6.259,
    "args": {
      "External id": 1577,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 508,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 508, "pid": 870, "tid": 870, "ts": 7429770608015.144,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608021.790, "dur": 4.182,
    "args": {
      "External id": 1577,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 509,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 509, "pid": 870, "tid": 870, "ts": 7429770608021.790,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608036.131, "dur": 3.773,
    "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": 510,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 510, "pid": 870, "tid": 870, "ts": 7429770608036.131,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608068.661, "dur": 3.669,
    "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": 511,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 511, "pid": 870, "tid": 870, "ts": 7429770608068.661,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608082.006, "dur": 5.769,
    "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": 512,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 512, "pid": 870, "tid": 870, "ts": 7429770608082.006,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608096.111, "dur": 4.373,
    "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": 513,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 513, "pid": 870, "tid": 870, "ts": 7429770608096.111,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608109.408, "dur": 4.365,
    "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": 514,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 514, "pid": 870, "tid": 870, "ts": 7429770608109.408,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608119.613, "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": 515,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 515, "pid": 870, "tid": 870, "ts": 7429770608119.613,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608131.388, "dur": 3.700,
    "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": 516,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 516, "pid": 870, "tid": 870, "ts": 7429770608131.388,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770608173.518, "dur": 0.230,
    "args": {
      "External id": 1598,      "cid": 370, "correlation": 517
    }
  },
  {
    "ph": "f", "id": 517, "pid": 870, "tid": 870, "ts": 7429770608173.518,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770608176.894, "dur": 0.176,
    "args": {
      "External id": 1598,      "cid": 370, "correlation": 518
    }
  },
  {
    "ph": "f", "id": 518, "pid": 870, "tid": 870, "ts": 7429770608176.894,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770608179.352, "dur": 0.153,
    "args": {
      "External id": 1598,      "cid": 370, "correlation": 519
    }
  },
  {
    "ph": "f", "id": 519, "pid": 870, "tid": 870, "ts": 7429770608179.352,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608181.439, "dur": 4.886,
    "args": {
      "External id": 1598,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 520,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 520, "pid": 870, "tid": 870, "ts": 7429770608181.439,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770608223.584, "dur": 0.294,
    "args": {
      "External id": 1611,      "cid": 370, "correlation": 521
    }
  },
  {
    "ph": "f", "id": 521, "pid": 870, "tid": 870, "ts": 7429770608223.584,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770608226.866, "dur": 0.155,
    "args": {
      "External id": 1611,      "cid": 370, "correlation": 522
    }
  },
  {
    "ph": "f", "id": 522, "pid": 870, "tid": 870, "ts": 7429770608226.866,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770608228.912, "dur": 0.139,
    "args": {
      "External id": 1611,      "cid": 370, "correlation": 523
    }
  },
  {
    "ph": "f", "id": 523, "pid": 870, "tid": 870, "ts": 7429770608228.912,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608230.459, "dur": 5.348,
    "args": {
      "External id": 1611,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 524,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 524, "pid": 870, "tid": 870, "ts": 7429770608230.459,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770608269.753, "dur": 0.195,
    "args": {
      "External id": 1624,      "cid": 370, "correlation": 525
    }
  },
  {
    "ph": "f", "id": 525, "pid": 870, "tid": 870, "ts": 7429770608269.753,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770608272.904, "dur": 0.136,
    "args": {
      "External id": 1624,      "cid": 370, "correlation": 526
    }
  },
  {
    "ph": "f", "id": 526, "pid": 870, "tid": 870, "ts": 7429770608272.904,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770608274.531, "dur": 0.137,
    "args": {
      "External id": 1624,      "cid": 370, "correlation": 527
    }
  },
  {
    "ph": "f", "id": 527, "pid": 870, "tid": 870, "ts": 7429770608274.531,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608275.940, "dur": 5.563,
    "args": {
      "External id": 1624,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 528,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 528, "pid": 870, "tid": 870, "ts": 7429770608275.940,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608303.682, "dur": 4.433,
    "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": 529,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 529, "pid": 870, "tid": 870, "ts": 7429770608303.682,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608329.267, "dur": 4.680,
    "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": 530,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 530, "pid": 870, "tid": 870, "ts": 7429770608329.267,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608348.713, "dur": 9.163,
    "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": 531,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 531, "pid": 870, "tid": 870, "ts": 7429770608348.713,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608365.308, "dur": 4.190,
    "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": 532,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 532, "pid": 870, "tid": 870, "ts": 7429770608365.308,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608374.807, "dur": 4.399,
    "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": 533,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 533, "pid": 870, "tid": 870, "ts": 7429770608374.807,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608385.875, "dur": 4.530,
    "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": 534,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 534, "pid": 870, "tid": 870, "ts": 7429770608385.875,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608402.685, "dur": 4.931,
    "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": 535,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 535, "pid": 870, "tid": 870, "ts": 7429770608402.685,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608413.926, "dur": 4.916,
    "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": 536,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 536, "pid": 870, "tid": 870, "ts": 7429770608413.926,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608425.046, "dur": 4.649,
    "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": 537,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 537, "pid": 870, "tid": 870, "ts": 7429770608425.046,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608434.589, "dur": 3.343,
    "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": 538,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 538, "pid": 870, "tid": 870, "ts": 7429770608434.589,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608497.170, "dur": 4.464,
    "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": 539,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 539, "pid": 870, "tid": 870, "ts": 7429770608497.170,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608531.682, "dur": 4.226,
    "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": 540,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 540, "pid": 870, "tid": 870, "ts": 7429770608531.682,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770608570.832, "dur": 0.220,
    "args": {
      "External id": 1699,      "cid": 370, "correlation": 541
    }
  },
  {
    "ph": "f", "id": 541, "pid": 870, "tid": 870, "ts": 7429770608570.832,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770608574.947, "dur": 0.169,
    "args": {
      "External id": 1699,      "cid": 370, "correlation": 542
    }
  },
  {
    "ph": "f", "id": 542, "pid": 870, "tid": 870, "ts": 7429770608574.947,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770608577.635, "dur": 0.189,
    "args": {
      "External id": 1699,      "cid": 370, "correlation": 543
    }
  },
  {
    "ph": "f", "id": 543, "pid": 870, "tid": 870, "ts": 7429770608577.635,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608580.197, "dur": 5.732,
    "args": {
      "External id": 1699,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 544,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 544, "pid": 870, "tid": 870, "ts": 7429770608580.197,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608597.319, "dur": 3.870,
    "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": 545,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 545, "pid": 870, "tid": 870, "ts": 7429770608597.319,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608615.597, "dur": 3.363,
    "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": 546,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 546, "pid": 870, "tid": 870, "ts": 7429770608615.597,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608631.006, "dur": 4.900,
    "args": {
      "External id": 1712,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 547,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 547, "pid": 870, "tid": 870, "ts": 7429770608631.006,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770608677.179, "dur": 0.226,
    "args": {
      "External id": 1724,      "cid": 370, "correlation": 548
    }
  },
  {
    "ph": "f", "id": 548, "pid": 870, "tid": 870, "ts": 7429770608677.179,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770608681.356, "dur": 0.148,
    "args": {
      "External id": 1724,      "cid": 370, "correlation": 549
    }
  },
  {
    "ph": "f", "id": 549, "pid": 870, "tid": 870, "ts": 7429770608681.356,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770608684.147, "dur": 0.149,
    "args": {
      "External id": 1724,      "cid": 370, "correlation": 550
    }
  },
  {
    "ph": "f", "id": 550, "pid": 870, "tid": 870, "ts": 7429770608684.147,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608686.378, "dur": 5.220,
    "args": {
      "External id": 1724,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 551,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 551, "pid": 870, "tid": 870, "ts": 7429770608686.378,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608708.186, "dur": 4.805,
    "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": 552,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 552, "pid": 870, "tid": 870, "ts": 7429770608708.186,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770608757.928, "dur": 0.188,
    "args": {
      "External id": 1743,      "cid": 370, "correlation": 553
    }
  },
  {
    "ph": "f", "id": 553, "pid": 870, "tid": 870, "ts": 7429770608757.928,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770608761.385, "dur": 0.155,
    "args": {
      "External id": 1743,      "cid": 370, "correlation": 554
    }
  },
  {
    "ph": "f", "id": 554, "pid": 870, "tid": 870, "ts": 7429770608761.385,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770608763.745, "dur": 0.182,
    "args": {
      "External id": 1743,      "cid": 370, "correlation": 555
    }
  },
  {
    "ph": "f", "id": 555, "pid": 870, "tid": 870, "ts": 7429770608763.745,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608765.653, "dur": 5.978,
    "args": {
      "External id": 1743,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 556,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 556, "pid": 870, "tid": 870, "ts": 7429770608765.653,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608784.933, "dur": 3.885,
    "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": 557,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 557, "pid": 870, "tid": 870, "ts": 7429770608784.933,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608807.587, "dur": 4.026,
    "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": 558,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 558, "pid": 870, "tid": 870, "ts": 7429770608807.587,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608821.273, "dur": 5.301,
    "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": 559,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 559, "pid": 870, "tid": 870, "ts": 7429770608821.273,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608834.594, "dur": 4.207,
    "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": 560,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 560, "pid": 870, "tid": 870, "ts": 7429770608834.594,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608844.584, "dur": 6.408,
    "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": 561,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 561, "pid": 870, "tid": 870, "ts": 7429770608844.584,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608857.320, "dur": 4.584,
    "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": 562,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 562, "pid": 870, "tid": 870, "ts": 7429770608857.320,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608870.047, "dur": 3.602,
    "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": 563,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 563, "pid": 870, "tid": 870, "ts": 7429770608870.047,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770608921.687, "dur": 0.209,
    "args": {
      "External id": 1764,      "cid": 370, "correlation": 564
    }
  },
  {
    "ph": "f", "id": 564, "pid": 870, "tid": 870, "ts": 7429770608921.687,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770608924.565, "dur": 0.146,
    "args": {
      "External id": 1764,      "cid": 370, "correlation": 565
    }
  },
  {
    "ph": "f", "id": 565, "pid": 870, "tid": 870, "ts": 7429770608924.565,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770608926.880, "dur": 0.151,
    "args": {
      "External id": 1764,      "cid": 370, "correlation": 566
    }
  },
  {
    "ph": "f", "id": 566, "pid": 870, "tid": 870, "ts": 7429770608926.880,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608929.049, "dur": 4.899,
    "args": {
      "External id": 1764,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 567,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 567, "pid": 870, "tid": 870, "ts": 7429770608929.049,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608950.377, "dur": 4.012,
    "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": 568,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 568, "pid": 870, "tid": 870, "ts": 7429770608950.377,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770608980.946, "dur": 0.188,
    "args": {
      "External id": 1775,      "cid": 370, "correlation": 569
    }
  },
  {
    "ph": "f", "id": 569, "pid": 870, "tid": 870, "ts": 7429770608980.946,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770608983.373, "dur": 0.138,
    "args": {
      "External id": 1775,      "cid": 370, "correlation": 570
    }
  },
  {
    "ph": "f", "id": 570, "pid": 870, "tid": 870, "ts": 7429770608983.373,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770608985.607, "dur": 0.140,
    "args": {
      "External id": 1775,      "cid": 370, "correlation": 571
    }
  },
  {
    "ph": "f", "id": 571, "pid": 870, "tid": 870, "ts": 7429770608985.607,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770608987.286, "dur": 6.040,
    "args": {
      "External id": 1775,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 572,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 572, "pid": 870, "tid": 870, "ts": 7429770608987.286,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609001.480, "dur": 4.933,
    "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": 573,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 573, "pid": 870, "tid": 870, "ts": 7429770609001.480,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770609034.925, "dur": 0.178,
    "args": {
      "External id": 1786,      "cid": 370, "correlation": 574
    }
  },
  {
    "ph": "f", "id": 574, "pid": 870, "tid": 870, "ts": 7429770609034.925,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770609037.295, "dur": 0.139,
    "args": {
      "External id": 1786,      "cid": 370, "correlation": 575
    }
  },
  {
    "ph": "f", "id": 575, "pid": 870, "tid": 870, "ts": 7429770609037.295,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770609039.973, "dur": 0.142,
    "args": {
      "External id": 1786,      "cid": 370, "correlation": 576
    }
  },
  {
    "ph": "f", "id": 576, "pid": 870, "tid": 870, "ts": 7429770609039.973,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609048.140, "dur": 6.123,
    "args": {
      "External id": 1786,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 577,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 577, "pid": 870, "tid": 870, "ts": 7429770609048.140,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609054.606, "dur": 3.832,
    "args": {
      "External id": 1786,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 578,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 578, "pid": 870, "tid": 870, "ts": 7429770609054.606,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609067.836, "dur": 5.793,
    "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": 579,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 579, "pid": 870, "tid": 870, "ts": 7429770609067.836,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609099.672, "dur": 3.990,
    "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": 580,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 580, "pid": 870, "tid": 870, "ts": 7429770609099.672,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609111.233, "dur": 4.433,
    "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": 581,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 581, "pid": 870, "tid": 870, "ts": 7429770609111.233,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609122.783, "dur": 3.955,
    "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": 582,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 582, "pid": 870, "tid": 870, "ts": 7429770609122.783,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609132.780, "dur": 5.396,
    "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": 583,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 583, "pid": 870, "tid": 870, "ts": 7429770609132.780,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609143.978, "dur": 3.888,
    "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": 584,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 584, "pid": 870, "tid": 870, "ts": 7429770609143.978,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609156.770, "dur": 3.458,
    "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": 585,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 585, "pid": 870, "tid": 870, "ts": 7429770609156.770,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770609198.672, "dur": 0.288,
    "args": {
      "External id": 1807,      "cid": 370, "correlation": 586
    }
  },
  {
    "ph": "f", "id": 586, "pid": 870, "tid": 870, "ts": 7429770609198.672,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770609202.548, "dur": 0.152,
    "args": {
      "External id": 1807,      "cid": 370, "correlation": 587
    }
  },
  {
    "ph": "f", "id": 587, "pid": 870, "tid": 870, "ts": 7429770609202.548,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770609204.781, "dur": 0.145,
    "args": {
      "External id": 1807,      "cid": 370, "correlation": 588
    }
  },
  {
    "ph": "f", "id": 588, "pid": 870, "tid": 870, "ts": 7429770609204.781,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609206.804, "dur": 5.130,
    "args": {
      "External id": 1807,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 589,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 589, "pid": 870, "tid": 870, "ts": 7429770609206.804,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770609247.476, "dur": 0.220,
    "args": {
      "External id": 1820,      "cid": 370, "correlation": 590
    }
  },
  {
    "ph": "f", "id": 590, "pid": 870, "tid": 870, "ts": 7429770609247.476,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770609250.628, "dur": 0.154,
    "args": {
      "External id": 1820,      "cid": 370, "correlation": 591
    }
  },
  {
    "ph": "f", "id": 591, "pid": 870, "tid": 870, "ts": 7429770609250.628,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770609252.981, "dur": 0.140,
    "args": {
      "External id": 1820,      "cid": 370, "correlation": 592
    }
  },
  {
    "ph": "f", "id": 592, "pid": 870, "tid": 870, "ts": 7429770609252.981,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609254.529, "dur": 4.938,
    "args": {
      "External id": 1820,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 593,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 593, "pid": 870, "tid": 870, "ts": 7429770609254.529,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770609292.422, "dur": 0.202,
    "args": {
      "External id": 1833,      "cid": 370, "correlation": 594
    }
  },
  {
    "ph": "f", "id": 594, "pid": 870, "tid": 870, "ts": 7429770609292.422,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770609295.542, "dur": 0.139,
    "args": {
      "External id": 1833,      "cid": 370, "correlation": 595
    }
  },
  {
    "ph": "f", "id": 595, "pid": 870, "tid": 870, "ts": 7429770609295.542,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770609297.209, "dur": 0.138,
    "args": {
      "External id": 1833,      "cid": 370, "correlation": 596
    }
  },
  {
    "ph": "f", "id": 596, "pid": 870, "tid": 870, "ts": 7429770609297.209,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609298.591, "dur": 5.360,
    "args": {
      "External id": 1833,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 597,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 597, "pid": 870, "tid": 870, "ts": 7429770609298.591,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609326.226, "dur": 4.173,
    "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": 598,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 598, "pid": 870, "tid": 870, "ts": 7429770609326.226,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609350.832, "dur": 4.004,
    "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": 599,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 599, "pid": 870, "tid": 870, "ts": 7429770609350.832,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609367.795, "dur": 6.835,
    "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": 600,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 600, "pid": 870, "tid": 870, "ts": 7429770609367.795,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609382.153, "dur": 4.168,
    "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": 601,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 601, "pid": 870, "tid": 870, "ts": 7429770609382.153,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609392.538, "dur": 4.333,
    "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": 602,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 602, "pid": 870, "tid": 870, "ts": 7429770609392.538,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609402.684, "dur": 4.017,
    "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": 603,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 603, "pid": 870, "tid": 870, "ts": 7429770609402.684,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609420.393, "dur": 4.509,
    "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": 604,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 604, "pid": 870, "tid": 870, "ts": 7429770609420.393,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609431.282, "dur": 6.142,
    "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": 605,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 605, "pid": 870, "tid": 870, "ts": 7429770609431.282,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609444.301, "dur": 3.821,
    "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": 606,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 606, "pid": 870, "tid": 870, "ts": 7429770609444.301,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609452.537, "dur": 3.525,
    "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": 607,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 607, "pid": 870, "tid": 870, "ts": 7429770609452.537,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609510.691, "dur": 6.981,
    "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": 608,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 608, "pid": 870, "tid": 870, "ts": 7429770609510.691,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609547.330, "dur": 4.101,
    "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": 609,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 609, "pid": 870, "tid": 870, "ts": 7429770609547.330,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770609585.238, "dur": 0.253,
    "args": {
      "External id": 1908,      "cid": 370, "correlation": 610
    }
  },
  {
    "ph": "f", "id": 610, "pid": 870, "tid": 870, "ts": 7429770609585.238,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770609589.235, "dur": 0.163,
    "args": {
      "External id": 1908,      "cid": 370, "correlation": 611
    }
  },
  {
    "ph": "f", "id": 611, "pid": 870, "tid": 870, "ts": 7429770609589.235,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770609592.135, "dur": 0.151,
    "args": {
      "External id": 1908,      "cid": 370, "correlation": 612
    }
  },
  {
    "ph": "f", "id": 612, "pid": 870, "tid": 870, "ts": 7429770609592.135,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609594.488, "dur": 5.147,
    "args": {
      "External id": 1908,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 613,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 613, "pid": 870, "tid": 870, "ts": 7429770609594.488,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609610.749, "dur": 4.179,
    "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": 614,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 614, "pid": 870, "tid": 870, "ts": 7429770609610.749,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609628.846, "dur": 4.044,
    "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": 615,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 615, "pid": 870, "tid": 870, "ts": 7429770609628.846,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609643.470, "dur": 4.927,
    "args": {
      "External id": 1921,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 616,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 616, "pid": 870, "tid": 870, "ts": 7429770609643.470,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770609687.552, "dur": 0.209,
    "args": {
      "External id": 1933,      "cid": 370, "correlation": 617
    }
  },
  {
    "ph": "f", "id": 617, "pid": 870, "tid": 870, "ts": 7429770609687.552,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770609690.750, "dur": 0.141,
    "args": {
      "External id": 1933,      "cid": 370, "correlation": 618
    }
  },
  {
    "ph": "f", "id": 618, "pid": 870, "tid": 870, "ts": 7429770609690.750,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770609693.373, "dur": 0.153,
    "args": {
      "External id": 1933,      "cid": 370, "correlation": 619
    }
  },
  {
    "ph": "f", "id": 619, "pid": 870, "tid": 870, "ts": 7429770609693.373,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609695.758, "dur": 4.748,
    "args": {
      "External id": 1933,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 620,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 620, "pid": 870, "tid": 870, "ts": 7429770609695.758,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609716.529, "dur": 4.748,
    "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": 621,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 621, "pid": 870, "tid": 870, "ts": 7429770609716.529,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770609759.696, "dur": 0.298,
    "args": {
      "External id": 1952,      "cid": 370, "correlation": 622
    }
  },
  {
    "ph": "f", "id": 622, "pid": 870, "tid": 870, "ts": 7429770609759.696,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770609762.918, "dur": 0.154,
    "args": {
      "External id": 1952,      "cid": 370, "correlation": 623
    }
  },
  {
    "ph": "f", "id": 623, "pid": 870, "tid": 870, "ts": 7429770609762.918,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770609765.272, "dur": 0.340,
    "args": {
      "External id": 1952,      "cid": 370, "correlation": 624
    }
  },
  {
    "ph": "f", "id": 624, "pid": 870, "tid": 870, "ts": 7429770609765.272,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609767.197, "dur": 6.516,
    "args": {
      "External id": 1952,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 625,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 625, "pid": 870, "tid": 870, "ts": 7429770609767.197,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609785.113, "dur": 4.043,
    "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": 626,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 626, "pid": 870, "tid": 870, "ts": 7429770609785.113,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609805.467, "dur": 4.016,
    "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": 627,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 627, "pid": 870, "tid": 870, "ts": 7429770609805.467,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609818.560, "dur": 5.235,
    "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": 628,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 628, "pid": 870, "tid": 870, "ts": 7429770609818.560,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609831.176, "dur": 4.487,
    "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": 629,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 629, "pid": 870, "tid": 870, "ts": 7429770609831.176,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609841.766, "dur": 4.448,
    "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": 630,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 630, "pid": 870, "tid": 870, "ts": 7429770609841.766,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609852.199, "dur": 5.172,
    "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": 631,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 631, "pid": 870, "tid": 870, "ts": 7429770609852.199,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609866.272, "dur": 3.960,
    "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": 632,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 632, "pid": 870, "tid": 870, "ts": 7429770609866.272,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770609904.092, "dur": 0.206,
    "args": {
      "External id": 1973,      "cid": 370, "correlation": 633
    }
  },
  {
    "ph": "f", "id": 633, "pid": 870, "tid": 870, "ts": 7429770609904.092,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770609907.526, "dur": 0.141,
    "args": {
      "External id": 1973,      "cid": 370, "correlation": 634
    }
  },
  {
    "ph": "f", "id": 634, "pid": 870, "tid": 870, "ts": 7429770609907.526,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770609909.987, "dur": 0.152,
    "args": {
      "External id": 1973,      "cid": 370, "correlation": 635
    }
  },
  {
    "ph": "f", "id": 635, "pid": 870, "tid": 870, "ts": 7429770609909.987,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609912.004, "dur": 4.741,
    "args": {
      "External id": 1973,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 636,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 636, "pid": 870, "tid": 870, "ts": 7429770609912.004,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609931.941, "dur": 4.789,
    "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": 637,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 637, "pid": 870, "tid": 870, "ts": 7429770609931.941,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770609962.060, "dur": 0.166,
    "args": {
      "External id": 1984,      "cid": 370, "correlation": 638
    }
  },
  {
    "ph": "f", "id": 638, "pid": 870, "tid": 870, "ts": 7429770609962.060,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770609965.220, "dur": 0.141,
    "args": {
      "External id": 1984,      "cid": 370, "correlation": 639
    }
  },
  {
    "ph": "f", "id": 639, "pid": 870, "tid": 870, "ts": 7429770609965.220,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770609967.657, "dur": 0.151,
    "args": {
      "External id": 1984,      "cid": 370, "correlation": 640
    }
  },
  {
    "ph": "f", "id": 640, "pid": 870, "tid": 870, "ts": 7429770609967.657,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609969.320, "dur": 5.377,
    "args": {
      "External id": 1984,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 641,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 641, "pid": 870, "tid": 870, "ts": 7429770609969.320,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770609983.126, "dur": 4.222,
    "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": 642,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 642, "pid": 870, "tid": 870, "ts": 7429770609983.126,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770610016.524, "dur": 0.348,
    "args": {
      "External id": 1996,      "cid": 290, "correlation": 643
    }
  },
  {
    "ph": "f", "id": 643, "pid": 870, "tid": 870, "ts": 7429770610016.524,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770610017.558, "dur": 90.554,
    "args": {
      "External id": 1996,            "bytes": 79691776, "cid": 108, "correlation": 644, "ptr": "0x7ffd8041f0c0"
    }
  },
  {
    "ph": "s", "id": 644, "pid": 870, "tid": 870, "ts": 7429770610017.558,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770610111.608, "dur": 0.367,
    "args": {
      "External id": 1995,      "cid": 370, "correlation": 645
    }
  },
  {
    "ph": "f", "id": 645, "pid": 870, "tid": 870, "ts": 7429770610111.608,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770610117.510, "dur": 0.147,
    "args": {
      "External id": 1995,      "cid": 370, "correlation": 646
    }
  },
  {
    "ph": "f", "id": 646, "pid": 870, "tid": 870, "ts": 7429770610117.510,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770610120.900, "dur": 0.342,
    "args": {
      "External id": 1995,      "cid": 370, "correlation": 647
    }
  },
  {
    "ph": "f", "id": 647, "pid": 870, "tid": 870, "ts": 7429770610120.900,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770610133.007, "dur": 7.674,
    "args": {
      "External id": 1995,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 648,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 648, "pid": 870, "tid": 870, "ts": 7429770610133.007,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770610141.065, "dur": 4.160,
    "args": {
      "External id": 1995,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 649,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 649, "pid": 870, "tid": 870, "ts": 7429770610141.065,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770610166.368, "dur": 4.379,
    "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": 650,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 650, "pid": 870, "tid": 870, "ts": 7429770610166.368,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770610205.152, "dur": 6.069,
    "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": 651,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 651, "pid": 870, "tid": 870, "ts": 7429770610205.152,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770610222.811, "dur": 6.212,
    "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": 652,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 652, "pid": 870, "tid": 870, "ts": 7429770610222.811,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770610238.823, "dur": 4.419,
    "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": 653,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 653, "pid": 870, "tid": 870, "ts": 7429770610238.823,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770610248.622, "dur": 4.622,
    "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": 654,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 654, "pid": 870, "tid": 870, "ts": 7429770610248.622,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770610260.890, "dur": 4.161,
    "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": 655,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 655, "pid": 870, "tid": 870, "ts": 7429770610260.890,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770610272.455, "dur": 5.205,
    "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": 656,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 656, "pid": 870, "tid": 870, "ts": 7429770610272.455,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770610324.279, "dur": 0.250,
    "args": {
      "External id": 2016,      "cid": 370, "correlation": 657
    }
  },
  {
    "ph": "f", "id": 657, "pid": 870, "tid": 870, "ts": 7429770610324.279,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770610328.202, "dur": 0.155,
    "args": {
      "External id": 2016,      "cid": 370, "correlation": 658
    }
  },
  {
    "ph": "f", "id": 658, "pid": 870, "tid": 870, "ts": 7429770610328.202,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770610330.846, "dur": 0.355,
    "args": {
      "External id": 2016,      "cid": 370, "correlation": 659
    }
  },
  {
    "ph": "f", "id": 659, "pid": 870, "tid": 870, "ts": 7429770610330.846,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770610333.473, "dur": 4.822,
    "args": {
      "External id": 2016,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 660,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 660, "pid": 870, "tid": 870, "ts": 7429770610333.473,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770610577.099, "dur": 0.283,
    "args": {
      "External id": 2029,      "cid": 370, "correlation": 661
    }
  },
  {
    "ph": "f", "id": 661, "pid": 870, "tid": 870, "ts": 7429770610577.099,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770610581.783, "dur": 0.140,
    "args": {
      "External id": 2029,      "cid": 370, "correlation": 662
    }
  },
  {
    "ph": "f", "id": 662, "pid": 870, "tid": 870, "ts": 7429770610581.783,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770610584.311, "dur": 0.160,
    "args": {
      "External id": 2029,      "cid": 370, "correlation": 663
    }
  },
  {
    "ph": "f", "id": 663, "pid": 870, "tid": 870, "ts": 7429770610584.311,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770610587.356, "dur": 7.731,
    "args": {
      "External id": 2029,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 664,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 664, "pid": 870, "tid": 870, "ts": 7429770610587.356,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770610641.394, "dur": 0.221,
    "args": {
      "External id": 2042,      "cid": 370, "correlation": 665
    }
  },
  {
    "ph": "f", "id": 665, "pid": 870, "tid": 870, "ts": 7429770610641.394,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770610644.583, "dur": 0.163,
    "args": {
      "External id": 2042,      "cid": 370, "correlation": 666
    }
  },
  {
    "ph": "f", "id": 666, "pid": 870, "tid": 870, "ts": 7429770610644.583,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770610646.448, "dur": 0.137,
    "args": {
      "External id": 2042,      "cid": 370, "correlation": 667
    }
  },
  {
    "ph": "f", "id": 667, "pid": 870, "tid": 870, "ts": 7429770610646.448,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770610648.032, "dur": 5.158,
    "args": {
      "External id": 2042,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 668,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 668, "pid": 870, "tid": 870, "ts": 7429770610648.032,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770610713.589, "dur": 5.892,
    "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": 669,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 669, "pid": 870, "tid": 870, "ts": 7429770610713.589,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770610745.441, "dur": 5.738,
    "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": 670,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 670, "pid": 870, "tid": 870, "ts": 7429770610745.441,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770610766.523, "dur": 7.354,
    "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": 671,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 671, "pid": 870, "tid": 870, "ts": 7429770610766.523,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770610782.263, "dur": 4.333,
    "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": 672,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 672, "pid": 870, "tid": 870, "ts": 7429770610782.263,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770610793.617, "dur": 4.576,
    "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": 673,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 673, "pid": 870, "tid": 870, "ts": 7429770610793.617,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770610803.634, "dur": 4.017,
    "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": 674,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 674, "pid": 870, "tid": 870, "ts": 7429770610803.634,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770610821.507, "dur": 4.907,
    "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": 675,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 675, "pid": 870, "tid": 870, "ts": 7429770610821.507,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770610833.778, "dur": 6.385,
    "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": 676,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 676, "pid": 870, "tid": 870, "ts": 7429770610833.778,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770610846.747, "dur": 3.626,
    "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": 677,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 677, "pid": 870, "tid": 870, "ts": 7429770610846.747,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770610854.930, "dur": 3.956,
    "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": 678,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 678, "pid": 870, "tid": 870, "ts": 7429770610854.930,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770610924.386, "dur": 5.313,
    "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": 679,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 679, "pid": 870, "tid": 870, "ts": 7429770610924.386,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770610959.270, "dur": 4.091,
    "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": 680,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 680, "pid": 870, "tid": 870, "ts": 7429770610959.270,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770611000.690, "dur": 0.296,
    "args": {
      "External id": 2117,      "cid": 370, "correlation": 681
    }
  },
  {
    "ph": "f", "id": 681, "pid": 870, "tid": 870, "ts": 7429770611000.690,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770611004.956, "dur": 0.151,
    "args": {
      "External id": 2117,      "cid": 370, "correlation": 682
    }
  },
  {
    "ph": "f", "id": 682, "pid": 870, "tid": 870, "ts": 7429770611004.956,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770611008.162, "dur": 0.183,
    "args": {
      "External id": 2117,      "cid": 370, "correlation": 683
    }
  },
  {
    "ph": "f", "id": 683, "pid": 870, "tid": 870, "ts": 7429770611008.162,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611011.307, "dur": 6.185,
    "args": {
      "External id": 2117,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 684,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 684, "pid": 870, "tid": 870, "ts": 7429770611011.307,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611030.751, "dur": 4.394,
    "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": 685,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 685, "pid": 870, "tid": 870, "ts": 7429770611030.751,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611050.099, "dur": 4.084,
    "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": 686,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 686, "pid": 870, "tid": 870, "ts": 7429770611050.099,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611068.375, "dur": 5.956,
    "args": {
      "External id": 2130,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 687,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 687, "pid": 870, "tid": 870, "ts": 7429770611068.375,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770611119.653, "dur": 0.192,
    "args": {
      "External id": 2142,      "cid": 370, "correlation": 688
    }
  },
  {
    "ph": "f", "id": 688, "pid": 870, "tid": 870, "ts": 7429770611119.653,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770611123.212, "dur": 0.141,
    "args": {
      "External id": 2142,      "cid": 370, "correlation": 689
    }
  },
  {
    "ph": "f", "id": 689, "pid": 870, "tid": 870, "ts": 7429770611123.212,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770611126.056, "dur": 0.170,
    "args": {
      "External id": 2142,      "cid": 370, "correlation": 690
    }
  },
  {
    "ph": "f", "id": 690, "pid": 870, "tid": 870, "ts": 7429770611126.056,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611128.220, "dur": 5.210,
    "args": {
      "External id": 2142,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 691,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 691, "pid": 870, "tid": 870, "ts": 7429770611128.220,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611150.530, "dur": 4.315,
    "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": 692,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 692, "pid": 870, "tid": 870, "ts": 7429770611150.530,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770611197.812, "dur": 0.201,
    "args": {
      "External id": 2161,      "cid": 370, "correlation": 693
    }
  },
  {
    "ph": "f", "id": 693, "pid": 870, "tid": 870, "ts": 7429770611197.812,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770611201.246, "dur": 0.152,
    "args": {
      "External id": 2161,      "cid": 370, "correlation": 694
    }
  },
  {
    "ph": "f", "id": 694, "pid": 870, "tid": 870, "ts": 7429770611201.246,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770611203.388, "dur": 0.596,
    "args": {
      "External id": 2161,      "cid": 370, "correlation": 695
    }
  },
  {
    "ph": "f", "id": 695, "pid": 870, "tid": 870, "ts": 7429770611203.388,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611205.993, "dur": 6.070,
    "args": {
      "External id": 2161,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 696,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 696, "pid": 870, "tid": 870, "ts": 7429770611205.993,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611226.071, "dur": 4.452,
    "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": 697,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 697, "pid": 870, "tid": 870, "ts": 7429770611226.071,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611258.325, "dur": 6.900,
    "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": 698,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 698, "pid": 870, "tid": 870, "ts": 7429770611258.325,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611277.339, "dur": 5.096,
    "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": 699,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 699, "pid": 870, "tid": 870, "ts": 7429770611277.339,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611291.401, "dur": 5.028,
    "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": 700,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 700, "pid": 870, "tid": 870, "ts": 7429770611291.401,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611302.948, "dur": 4.560,
    "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": 701,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 701, "pid": 870, "tid": 870, "ts": 7429770611302.948,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611314.294, "dur": 4.855,
    "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": 702,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 702, "pid": 870, "tid": 870, "ts": 7429770611314.294,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611326.767, "dur": 3.791,
    "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": 703,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 703, "pid": 870, "tid": 870, "ts": 7429770611326.767,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770611370.569, "dur": 0.206,
    "args": {
      "External id": 2182,      "cid": 370, "correlation": 704
    }
  },
  {
    "ph": "f", "id": 704, "pid": 870, "tid": 870, "ts": 7429770611370.569,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770611373.996, "dur": 0.142,
    "args": {
      "External id": 2182,      "cid": 370, "correlation": 705
    }
  },
  {
    "ph": "f", "id": 705, "pid": 870, "tid": 870, "ts": 7429770611373.996,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770611376.623, "dur": 0.155,
    "args": {
      "External id": 2182,      "cid": 370, "correlation": 706
    }
  },
  {
    "ph": "f", "id": 706, "pid": 870, "tid": 870, "ts": 7429770611376.623,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611378.646, "dur": 4.862,
    "args": {
      "External id": 2182,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 707,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 707, "pid": 870, "tid": 870, "ts": 7429770611378.646,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611401.760, "dur": 4.673,
    "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": 708,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 708, "pid": 870, "tid": 870, "ts": 7429770611401.760,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770611433.827, "dur": 0.145,
    "args": {
      "External id": 2193,      "cid": 370, "correlation": 709
    }
  },
  {
    "ph": "f", "id": 709, "pid": 870, "tid": 870, "ts": 7429770611433.827,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770611436.761, "dur": 0.138,
    "args": {
      "External id": 2193,      "cid": 370, "correlation": 710
    }
  },
  {
    "ph": "f", "id": 710, "pid": 870, "tid": 870, "ts": 7429770611436.761,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770611438.679, "dur": 0.140,
    "args": {
      "External id": 2193,      "cid": 370, "correlation": 711
    }
  },
  {
    "ph": "f", "id": 711, "pid": 870, "tid": 870, "ts": 7429770611438.679,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611440.698, "dur": 5.746,
    "args": {
      "External id": 2193,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 712,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 712, "pid": 870, "tid": 870, "ts": 7429770611440.698,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611455.273, "dur": 4.474,
    "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": 713,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 713, "pid": 870, "tid": 870, "ts": 7429770611455.273,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770611485.128, "dur": 0.159,
    "args": {
      "External id": 2204,      "cid": 370, "correlation": 714
    }
  },
  {
    "ph": "f", "id": 714, "pid": 870, "tid": 870, "ts": 7429770611485.128,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770611488.243, "dur": 0.139,
    "args": {
      "External id": 2204,      "cid": 370, "correlation": 715
    }
  },
  {
    "ph": "f", "id": 715, "pid": 870, "tid": 870, "ts": 7429770611488.243,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770611491.025, "dur": 0.144,
    "args": {
      "External id": 2204,      "cid": 370, "correlation": 716
    }
  },
  {
    "ph": "f", "id": 716, "pid": 870, "tid": 870, "ts": 7429770611491.025,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611499.704, "dur": 6.911,
    "args": {
      "External id": 2204,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 717,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 717, "pid": 870, "tid": 870, "ts": 7429770611499.704,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611506.976, "dur": 4.319,
    "args": {
      "External id": 2204,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 718,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 718, "pid": 870, "tid": 870, "ts": 7429770611506.976,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611520.957, "dur": 3.785,
    "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": 719,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 719, "pid": 870, "tid": 870, "ts": 7429770611520.957,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611552.467, "dur": 3.731,
    "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": 720,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 720, "pid": 870, "tid": 870, "ts": 7429770611552.467,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611565.960, "dur": 4.671,
    "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": 721,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 721, "pid": 870, "tid": 870, "ts": 7429770611565.960,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611577.660, "dur": 4.799,
    "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": 722,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 722, "pid": 870, "tid": 870, "ts": 7429770611577.660,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611588.763, "dur": 4.905,
    "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": 723,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 723, "pid": 870, "tid": 870, "ts": 7429770611588.763,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611599.738, "dur": 3.739,
    "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": 724,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 724, "pid": 870, "tid": 870, "ts": 7429770611599.738,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611611.384, "dur": 3.631,
    "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": 725,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 725, "pid": 870, "tid": 870, "ts": 7429770611611.384,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770611654.706, "dur": 0.276,
    "args": {
      "External id": 2225,      "cid": 370, "correlation": 726
    }
  },
  {
    "ph": "f", "id": 726, "pid": 870, "tid": 870, "ts": 7429770611654.706,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770611658.214, "dur": 0.164,
    "args": {
      "External id": 2225,      "cid": 370, "correlation": 727
    }
  },
  {
    "ph": "f", "id": 727, "pid": 870, "tid": 870, "ts": 7429770611658.214,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770611660.960, "dur": 0.151,
    "args": {
      "External id": 2225,      "cid": 370, "correlation": 728
    }
  },
  {
    "ph": "f", "id": 728, "pid": 870, "tid": 870, "ts": 7429770611660.960,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611663.107, "dur": 4.637,
    "args": {
      "External id": 2225,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 729,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 729, "pid": 870, "tid": 870, "ts": 7429770611663.107,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770611704.121, "dur": 0.212,
    "args": {
      "External id": 2238,      "cid": 370, "correlation": 730
    }
  },
  {
    "ph": "f", "id": 730, "pid": 870, "tid": 870, "ts": 7429770611704.121,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770611707.380, "dur": 0.152,
    "args": {
      "External id": 2238,      "cid": 370, "correlation": 731
    }
  },
  {
    "ph": "f", "id": 731, "pid": 870, "tid": 870, "ts": 7429770611707.380,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770611709.591, "dur": 0.148,
    "args": {
      "External id": 2238,      "cid": 370, "correlation": 732
    }
  },
  {
    "ph": "f", "id": 732, "pid": 870, "tid": 870, "ts": 7429770611709.591,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611711.166, "dur": 5.419,
    "args": {
      "External id": 2238,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 733,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 733, "pid": 870, "tid": 870, "ts": 7429770611711.166,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770611749.553, "dur": 0.189,
    "args": {
      "External id": 2251,      "cid": 370, "correlation": 734
    }
  },
  {
    "ph": "f", "id": 734, "pid": 870, "tid": 870, "ts": 7429770611749.553,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770611752.728, "dur": 0.140,
    "args": {
      "External id": 2251,      "cid": 370, "correlation": 735
    }
  },
  {
    "ph": "f", "id": 735, "pid": 870, "tid": 870, "ts": 7429770611752.728,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770611755.008, "dur": 0.151,
    "args": {
      "External id": 2251,      "cid": 370, "correlation": 736
    }
  },
  {
    "ph": "f", "id": 736, "pid": 870, "tid": 870, "ts": 7429770611755.008,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611756.455, "dur": 5.917,
    "args": {
      "External id": 2251,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 737,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 737, "pid": 870, "tid": 870, "ts": 7429770611756.455,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611784.701, "dur": 4.955,
    "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": 738,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 738, "pid": 870, "tid": 870, "ts": 7429770611784.701,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611810.858, "dur": 4.806,
    "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": 739,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 739, "pid": 870, "tid": 870, "ts": 7429770611810.858,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611828.167, "dur": 7.627,
    "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": 740,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 740, "pid": 870, "tid": 870, "ts": 7429770611828.167,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611842.990, "dur": 3.986,
    "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": 741,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 741, "pid": 870, "tid": 870, "ts": 7429770611842.990,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611851.952, "dur": 5.741,
    "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": 742,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 742, "pid": 870, "tid": 870, "ts": 7429770611851.952,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611864.706, "dur": 4.023,
    "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": 743,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 743, "pid": 870, "tid": 870, "ts": 7429770611864.706,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611882.577, "dur": 5.886,
    "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": 744,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 744, "pid": 870, "tid": 870, "ts": 7429770611882.577,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611894.378, "dur": 5.039,
    "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": 745,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 745, "pid": 870, "tid": 870, "ts": 7429770611894.378,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611906.086, "dur": 5.076,
    "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": 746,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 746, "pid": 870, "tid": 870, "ts": 7429770611906.086,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611915.911, "dur": 3.702,
    "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": 747,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 747, "pid": 870, "tid": 870, "ts": 7429770611915.911,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770611974.531, "dur": 5.086,
    "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": 748,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 748, "pid": 870, "tid": 870, "ts": 7429770611974.531,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612009.447, "dur": 3.771,
    "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": 749,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 749, "pid": 870, "tid": 870, "ts": 7429770612009.447,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770612047.051, "dur": 0.237,
    "args": {
      "External id": 2326,      "cid": 370, "correlation": 750
    }
  },
  {
    "ph": "f", "id": 750, "pid": 870, "tid": 870, "ts": 7429770612047.051,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770612050.572, "dur": 0.156,
    "args": {
      "External id": 2326,      "cid": 370, "correlation": 751
    }
  },
  {
    "ph": "f", "id": 751, "pid": 870, "tid": 870, "ts": 7429770612050.572,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770612053.668, "dur": 0.152,
    "args": {
      "External id": 2326,      "cid": 370, "correlation": 752
    }
  },
  {
    "ph": "f", "id": 752, "pid": 870, "tid": 870, "ts": 7429770612053.668,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612055.985, "dur": 6.292,
    "args": {
      "External id": 2326,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 753,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 753, "pid": 870, "tid": 870, "ts": 7429770612055.985,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612073.605, "dur": 3.938,
    "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": 754,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 754, "pid": 870, "tid": 870, "ts": 7429770612073.605,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612091.302, "dur": 3.880,
    "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": 755,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 755, "pid": 870, "tid": 870, "ts": 7429770612091.302,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612106.024, "dur": 6.041,
    "args": {
      "External id": 2339,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 756,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 756, "pid": 870, "tid": 870, "ts": 7429770612106.024,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770612152.138, "dur": 0.182,
    "args": {
      "External id": 2351,      "cid": 370, "correlation": 757
    }
  },
  {
    "ph": "f", "id": 757, "pid": 870, "tid": 870, "ts": 7429770612152.138,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770612155.290, "dur": 0.138,
    "args": {
      "External id": 2351,      "cid": 370, "correlation": 758
    }
  },
  {
    "ph": "f", "id": 758, "pid": 870, "tid": 870, "ts": 7429770612155.290,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770612158.038, "dur": 0.142,
    "args": {
      "External id": 2351,      "cid": 370, "correlation": 759
    }
  },
  {
    "ph": "f", "id": 759, "pid": 870, "tid": 870, "ts": 7429770612158.038,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612160.432, "dur": 4.956,
    "args": {
      "External id": 2351,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 760,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 760, "pid": 870, "tid": 870, "ts": 7429770612160.432,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612182.142, "dur": 3.960,
    "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": 761,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 761, "pid": 870, "tid": 870, "ts": 7429770612182.142,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770612225.984, "dur": 0.261,
    "args": {
      "External id": 2370,      "cid": 370, "correlation": 762
    }
  },
  {
    "ph": "f", "id": 762, "pid": 870, "tid": 870, "ts": 7429770612225.984,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770612229.038, "dur": 0.152,
    "args": {
      "External id": 2370,      "cid": 370, "correlation": 763
    }
  },
  {
    "ph": "f", "id": 763, "pid": 870, "tid": 870, "ts": 7429770612229.038,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770612231.478, "dur": 0.368,
    "args": {
      "External id": 2370,      "cid": 370, "correlation": 764
    }
  },
  {
    "ph": "f", "id": 764, "pid": 870, "tid": 870, "ts": 7429770612231.478,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612233.495, "dur": 5.576,
    "args": {
      "External id": 2370,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 765,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 765, "pid": 870, "tid": 870, "ts": 7429770612233.495,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612251.331, "dur": 3.981,
    "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": 766,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 766, "pid": 870, "tid": 870, "ts": 7429770612251.331,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612271.227, "dur": 3.958,
    "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": 767,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 767, "pid": 870, "tid": 870, "ts": 7429770612271.227,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612284.320, "dur": 5.196,
    "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": 768,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 768, "pid": 870, "tid": 870, "ts": 7429770612284.320,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612298.698, "dur": 11.903,
    "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": 769,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 769, "pid": 870, "tid": 870, "ts": 7429770612298.698,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612320.511, "dur": 4.304,
    "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": 770,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 770, "pid": 870, "tid": 870, "ts": 7429770612320.511,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612333.103, "dur": 5.582,
    "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": 771,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 771, "pid": 870, "tid": 870, "ts": 7429770612333.103,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612346.868, "dur": 4.946,
    "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": 772,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 772, "pid": 870, "tid": 870, "ts": 7429770612346.868,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770612393.085, "dur": 0.251,
    "args": {
      "External id": 2391,      "cid": 370, "correlation": 773
    }
  },
  {
    "ph": "f", "id": 773, "pid": 870, "tid": 870, "ts": 7429770612393.085,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770612397.112, "dur": 0.141,
    "args": {
      "External id": 2391,      "cid": 370, "correlation": 774
    }
  },
  {
    "ph": "f", "id": 774, "pid": 870, "tid": 870, "ts": 7429770612397.112,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770612399.560, "dur": 0.138,
    "args": {
      "External id": 2391,      "cid": 370, "correlation": 775
    }
  },
  {
    "ph": "f", "id": 775, "pid": 870, "tid": 870, "ts": 7429770612399.560,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612402.667, "dur": 5.361,
    "args": {
      "External id": 2391,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 776,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 776, "pid": 870, "tid": 870, "ts": 7429770612402.667,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612424.645, "dur": 4.166,
    "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": 777,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 777, "pid": 870, "tid": 870, "ts": 7429770612424.645,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770612454.578, "dur": 0.165,
    "args": {
      "External id": 2402,      "cid": 370, "correlation": 778
    }
  },
  {
    "ph": "f", "id": 778, "pid": 870, "tid": 870, "ts": 7429770612454.578,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770612457.683, "dur": 0.144,
    "args": {
      "External id": 2402,      "cid": 370, "correlation": 779
    }
  },
  {
    "ph": "f", "id": 779, "pid": 870, "tid": 870, "ts": 7429770612457.683,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770612459.853, "dur": 0.139,
    "args": {
      "External id": 2402,      "cid": 370, "correlation": 780
    }
  },
  {
    "ph": "f", "id": 780, "pid": 870, "tid": 870, "ts": 7429770612459.853,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612461.552, "dur": 5.317,
    "args": {
      "External id": 2402,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 781,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 781, "pid": 870, "tid": 870, "ts": 7429770612461.552,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612475.335, "dur": 4.292,
    "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": 782,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 782, "pid": 870, "tid": 870, "ts": 7429770612475.335,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770612503.223, "dur": 0.178,
    "args": {
      "External id": 2413,      "cid": 370, "correlation": 783
    }
  },
  {
    "ph": "f", "id": 783, "pid": 870, "tid": 870, "ts": 7429770612503.223,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770612506.887, "dur": 0.137,
    "args": {
      "External id": 2413,      "cid": 370, "correlation": 784
    }
  },
  {
    "ph": "f", "id": 784, "pid": 870, "tid": 870, "ts": 7429770612506.887,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770612509.465, "dur": 0.143,
    "args": {
      "External id": 2413,      "cid": 370, "correlation": 785
    }
  },
  {
    "ph": "f", "id": 785, "pid": 870, "tid": 870, "ts": 7429770612509.465,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612517.976, "dur": 6.519,
    "args": {
      "External id": 2413,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 786,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 786, "pid": 870, "tid": 870, "ts": 7429770612517.976,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612524.861, "dur": 4.931,
    "args": {
      "External id": 2413,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 787,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 787, "pid": 870, "tid": 870, "ts": 7429770612524.861,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612539.028, "dur": 4.037,
    "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": 788,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 788, "pid": 870, "tid": 870, "ts": 7429770612539.028,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612569.792, "dur": 3.826,
    "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": 789,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 789, "pid": 870, "tid": 870, "ts": 7429770612569.792,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612582.619, "dur": 5.302,
    "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": 790,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 790, "pid": 870, "tid": 870, "ts": 7429770612582.619,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612596.411, "dur": 4.862,
    "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": 791,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 791, "pid": 870, "tid": 870, "ts": 7429770612596.411,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612606.418, "dur": 6.533,
    "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": 792,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 792, "pid": 870, "tid": 870, "ts": 7429770612606.418,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612619.316, "dur": 4.604,
    "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": 793,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 793, "pid": 870, "tid": 870, "ts": 7429770612619.316,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612630.766, "dur": 5.372,
    "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": 794,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 794, "pid": 870, "tid": 870, "ts": 7429770612630.766,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770612673.223, "dur": 0.239,
    "args": {
      "External id": 2434,      "cid": 370, "correlation": 795
    }
  },
  {
    "ph": "f", "id": 795, "pid": 870, "tid": 870, "ts": 7429770612673.223,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770612677.122, "dur": 0.156,
    "args": {
      "External id": 2434,      "cid": 370, "correlation": 796
    }
  },
  {
    "ph": "f", "id": 796, "pid": 870, "tid": 870, "ts": 7429770612677.122,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770612679.297, "dur": 0.150,
    "args": {
      "External id": 2434,      "cid": 370, "correlation": 797
    }
  },
  {
    "ph": "f", "id": 797, "pid": 870, "tid": 870, "ts": 7429770612679.297,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612681.310, "dur": 4.972,
    "args": {
      "External id": 2434,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 798,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 798, "pid": 870, "tid": 870, "ts": 7429770612681.310,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770612720.689, "dur": 0.192,
    "args": {
      "External id": 2447,      "cid": 370, "correlation": 799
    }
  },
  {
    "ph": "f", "id": 799, "pid": 870, "tid": 870, "ts": 7429770612720.689,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770612724.010, "dur": 0.150,
    "args": {
      "External id": 2447,      "cid": 370, "correlation": 800
    }
  },
  {
    "ph": "f", "id": 800, "pid": 870, "tid": 870, "ts": 7429770612724.010,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770612726.057, "dur": 0.151,
    "args": {
      "External id": 2447,      "cid": 370, "correlation": 801
    }
  },
  {
    "ph": "f", "id": 801, "pid": 870, "tid": 870, "ts": 7429770612726.057,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612727.627, "dur": 4.588,
    "args": {
      "External id": 2447,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 802,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 802, "pid": 870, "tid": 870, "ts": 7429770612727.627,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770612763.985, "dur": 0.166,
    "args": {
      "External id": 2460,      "cid": 370, "correlation": 803
    }
  },
  {
    "ph": "f", "id": 803, "pid": 870, "tid": 870, "ts": 7429770612763.985,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770612767.248, "dur": 0.159,
    "args": {
      "External id": 2460,      "cid": 370, "correlation": 804
    }
  },
  {
    "ph": "f", "id": 804, "pid": 870, "tid": 870, "ts": 7429770612767.248,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770612769.096, "dur": 0.151,
    "args": {
      "External id": 2460,      "cid": 370, "correlation": 805
    }
  },
  {
    "ph": "f", "id": 805, "pid": 870, "tid": 870, "ts": 7429770612769.096,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612770.483, "dur": 5.601,
    "args": {
      "External id": 2460,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 806,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 806, "pid": 870, "tid": 870, "ts": 7429770612770.483,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612800.991, "dur": 5.152,
    "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": 807,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 807, "pid": 870, "tid": 870, "ts": 7429770612800.991,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612826.493, "dur": 4.432,
    "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": 808,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 808, "pid": 870, "tid": 870, "ts": 7429770612826.493,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612843.416, "dur": 6.660,
    "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": 809,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 809, "pid": 870, "tid": 870, "ts": 7429770612843.416,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612857.458, "dur": 3.748,
    "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": 810,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 810, "pid": 870, "tid": 870, "ts": 7429770612857.458,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612867.122, "dur": 3.526,
    "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": 811,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 811, "pid": 870, "tid": 870, "ts": 7429770612867.122,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612876.011, "dur": 3.748,
    "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": 812,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 812, "pid": 870, "tid": 870, "ts": 7429770612876.011,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612893.509, "dur": 4.639,
    "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": 813,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 813, "pid": 870, "tid": 870, "ts": 7429770612893.509,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612905.019, "dur": 6.568,
    "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": 814,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 814, "pid": 870, "tid": 870, "ts": 7429770612905.019,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612918.054, "dur": 3.763,
    "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": 815,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 815, "pid": 870, "tid": 870, "ts": 7429770612918.054,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612926.403, "dur": 4.012,
    "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": 816,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 816, "pid": 870, "tid": 870, "ts": 7429770612926.403,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770612987.935, "dur": 4.487,
    "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": 817,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 817, "pid": 870, "tid": 870, "ts": 7429770612987.935,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613022.963, "dur": 3.754,
    "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": 818,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 818, "pid": 870, "tid": 870, "ts": 7429770613022.963,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770613063.850, "dur": 0.205,
    "args": {
      "External id": 2535,      "cid": 370, "correlation": 819
    }
  },
  {
    "ph": "f", "id": 819, "pid": 870, "tid": 870, "ts": 7429770613063.850,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770613067.831, "dur": 0.159,
    "args": {
      "External id": 2535,      "cid": 370, "correlation": 820
    }
  },
  {
    "ph": "f", "id": 820, "pid": 870, "tid": 870, "ts": 7429770613067.831,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770613070.302, "dur": 0.143,
    "args": {
      "External id": 2535,      "cid": 370, "correlation": 821
    }
  },
  {
    "ph": "f", "id": 821, "pid": 870, "tid": 870, "ts": 7429770613070.302,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613072.707, "dur": 6.193,
    "args": {
      "External id": 2535,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 822,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 822, "pid": 870, "tid": 870, "ts": 7429770613072.707,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613090.520, "dur": 4.287,
    "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": 823,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 823, "pid": 870, "tid": 870, "ts": 7429770613090.520,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613109.917, "dur": 4.117,
    "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": 824,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 824, "pid": 870, "tid": 870, "ts": 7429770613109.917,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613126.666, "dur": 4.827,
    "args": {
      "External id": 2548,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 825,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 825, "pid": 870, "tid": 870, "ts": 7429770613126.666,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770613209.556, "dur": 0.451,
    "args": {
      "External id": 2561,      "cid": 290, "correlation": 826
    }
  },
  {
    "ph": "f", "id": 826, "pid": 870, "tid": 870, "ts": 7429770613209.556,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770613210.753, "dur": 72.398,
    "args": {
      "External id": 2561,            "bytes": 79691776, "cid": 108, "correlation": 827, "ptr": "0x7ffd8041fd60"
    }
  },
  {
    "ph": "s", "id": 827, "pid": 870, "tid": 870, "ts": 7429770613210.753,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770613286.001, "dur": 0.230,
    "args": {
      "External id": 2560,      "cid": 370, "correlation": 828
    }
  },
  {
    "ph": "f", "id": 828, "pid": 870, "tid": 870, "ts": 7429770613286.001,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770613289.791, "dur": 0.143,
    "args": {
      "External id": 2560,      "cid": 370, "correlation": 829
    }
  },
  {
    "ph": "f", "id": 829, "pid": 870, "tid": 870, "ts": 7429770613289.791,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770613292.470, "dur": 0.156,
    "args": {
      "External id": 2560,      "cid": 370, "correlation": 830
    }
  },
  {
    "ph": "f", "id": 830, "pid": 870, "tid": 870, "ts": 7429770613292.470,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613295.896, "dur": 6.665,
    "args": {
      "External id": 2560,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 831,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 831, "pid": 870, "tid": 870, "ts": 7429770613295.896,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613324.094, "dur": 5.002,
    "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": 832,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 832, "pid": 870, "tid": 870, "ts": 7429770613324.094,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770613380.800, "dur": 0.266,
    "args": {
      "External id": 2579,      "cid": 370, "correlation": 833
    }
  },
  {
    "ph": "f", "id": 833, "pid": 870, "tid": 870, "ts": 7429770613380.800,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770613384.345, "dur": 0.206,
    "args": {
      "External id": 2579,      "cid": 370, "correlation": 834
    }
  },
  {
    "ph": "f", "id": 834, "pid": 870, "tid": 870, "ts": 7429770613384.345,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770613386.738, "dur": 0.555,
    "args": {
      "External id": 2579,      "cid": 370, "correlation": 835
    }
  },
  {
    "ph": "f", "id": 835, "pid": 870, "tid": 870, "ts": 7429770613386.738,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613389.429, "dur": 6.590,
    "args": {
      "External id": 2579,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 836,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 836, "pid": 870, "tid": 870, "ts": 7429770613389.429,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613411.321, "dur": 4.437,
    "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": 837,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 837, "pid": 870, "tid": 870, "ts": 7429770613411.321,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613435.330, "dur": 6.909,
    "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": 838,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 838, "pid": 870, "tid": 870, "ts": 7429770613435.330,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613452.868, "dur": 6.260,
    "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": 839,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 839, "pid": 870, "tid": 870, "ts": 7429770613452.868,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613468.096, "dur": 4.314,
    "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": 840,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 840, "pid": 870, "tid": 870, "ts": 7429770613468.096,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613480.160, "dur": 5.696,
    "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": 841,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 841, "pid": 870, "tid": 870, "ts": 7429770613480.160,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613494.443, "dur": 5.519,
    "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": 842,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 842, "pid": 870, "tid": 870, "ts": 7429770613494.443,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613508.026, "dur": 4.938,
    "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": 843,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 843, "pid": 870, "tid": 870, "ts": 7429770613508.026,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770613551.886, "dur": 0.183,
    "args": {
      "External id": 2600,      "cid": 370, "correlation": 844
    }
  },
  {
    "ph": "f", "id": 844, "pid": 870, "tid": 870, "ts": 7429770613551.886,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770613555.251, "dur": 0.138,
    "args": {
      "External id": 2600,      "cid": 370, "correlation": 845
    }
  },
  {
    "ph": "f", "id": 845, "pid": 870, "tid": 870, "ts": 7429770613555.251,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770613557.669, "dur": 0.141,
    "args": {
      "External id": 2600,      "cid": 370, "correlation": 846
    }
  },
  {
    "ph": "f", "id": 846, "pid": 870, "tid": 870, "ts": 7429770613557.669,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613559.780, "dur": 5.359,
    "args": {
      "External id": 2600,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 847,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 847, "pid": 870, "tid": 870, "ts": 7429770613559.780,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613583.777, "dur": 5.282,
    "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": 848,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 848, "pid": 870, "tid": 870, "ts": 7429770613583.777,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770613616.360, "dur": 0.177,
    "args": {
      "External id": 2611,      "cid": 370, "correlation": 849
    }
  },
  {
    "ph": "f", "id": 849, "pid": 870, "tid": 870, "ts": 7429770613616.360,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770613619.524, "dur": 0.140,
    "args": {
      "External id": 2611,      "cid": 370, "correlation": 850
    }
  },
  {
    "ph": "f", "id": 850, "pid": 870, "tid": 870, "ts": 7429770613619.524,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770613621.738, "dur": 0.141,
    "args": {
      "External id": 2611,      "cid": 370, "correlation": 851
    }
  },
  {
    "ph": "f", "id": 851, "pid": 870, "tid": 870, "ts": 7429770613621.738,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613623.532, "dur": 5.075,
    "args": {
      "External id": 2611,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 852,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 852, "pid": 870, "tid": 870, "ts": 7429770613623.532,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613637.040, "dur": 4.295,
    "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": 853,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 853, "pid": 870, "tid": 870, "ts": 7429770613637.040,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770613667.583, "dur": 0.163,
    "args": {
      "External id": 2622,      "cid": 370, "correlation": 854
    }
  },
  {
    "ph": "f", "id": 854, "pid": 870, "tid": 870, "ts": 7429770613667.583,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770613670.731, "dur": 0.142,
    "args": {
      "External id": 2622,      "cid": 370, "correlation": 855
    }
  },
  {
    "ph": "f", "id": 855, "pid": 870, "tid": 870, "ts": 7429770613670.731,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770613673.619, "dur": 0.142,
    "args": {
      "External id": 2622,      "cid": 370, "correlation": 856
    }
  },
  {
    "ph": "f", "id": 856, "pid": 870, "tid": 870, "ts": 7429770613673.619,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613682.270, "dur": 5.787,
    "args": {
      "External id": 2622,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 857,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 857, "pid": 870, "tid": 870, "ts": 7429770613682.270,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613688.437, "dur": 4.145,
    "args": {
      "External id": 2622,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 858,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 858, "pid": 870, "tid": 870, "ts": 7429770613688.437,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613702.214, "dur": 4.367,
    "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": 859,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 859, "pid": 870, "tid": 870, "ts": 7429770613702.214,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613732.353, "dur": 3.840,
    "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": 860,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 860, "pid": 870, "tid": 870, "ts": 7429770613732.353,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613744.575, "dur": 4.815,
    "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": 861,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 861, "pid": 870, "tid": 870, "ts": 7429770613744.575,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613756.040, "dur": 4.511,
    "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": 862,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 862, "pid": 870, "tid": 870, "ts": 7429770613756.040,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613765.963, "dur": 5.754,
    "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": 863,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 863, "pid": 870, "tid": 870, "ts": 7429770613765.963,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613777.485, "dur": 3.808,
    "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": 864,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 864, "pid": 870, "tid": 870, "ts": 7429770613777.485,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613789.641, "dur": 3.611,
    "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": 865,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 865, "pid": 870, "tid": 870, "ts": 7429770613789.641,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770613831.834, "dur": 0.282,
    "args": {
      "External id": 2643,      "cid": 370, "correlation": 866
    }
  },
  {
    "ph": "f", "id": 866, "pid": 870, "tid": 870, "ts": 7429770613831.834,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770613835.566, "dur": 0.153,
    "args": {
      "External id": 2643,      "cid": 370, "correlation": 867
    }
  },
  {
    "ph": "f", "id": 867, "pid": 870, "tid": 870, "ts": 7429770613835.566,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770613838.270, "dur": 0.148,
    "args": {
      "External id": 2643,      "cid": 370, "correlation": 868
    }
  },
  {
    "ph": "f", "id": 868, "pid": 870, "tid": 870, "ts": 7429770613838.270,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613840.309, "dur": 5.653,
    "args": {
      "External id": 2643,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 869,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 869, "pid": 870, "tid": 870, "ts": 7429770613840.309,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770613881.997, "dur": 0.176,
    "args": {
      "External id": 2656,      "cid": 370, "correlation": 870
    }
  },
  {
    "ph": "f", "id": 870, "pid": 870, "tid": 870, "ts": 7429770613881.997,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770613885.186, "dur": 0.154,
    "args": {
      "External id": 2656,      "cid": 370, "correlation": 871
    }
  },
  {
    "ph": "f", "id": 871, "pid": 870, "tid": 870, "ts": 7429770613885.186,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770613887.895, "dur": 0.147,
    "args": {
      "External id": 2656,      "cid": 370, "correlation": 872
    }
  },
  {
    "ph": "f", "id": 872, "pid": 870, "tid": 870, "ts": 7429770613887.895,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613889.479, "dur": 5.581,
    "args": {
      "External id": 2656,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 873,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 873, "pid": 870, "tid": 870, "ts": 7429770613889.479,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770613927.525, "dur": 0.168,
    "args": {
      "External id": 2669,      "cid": 370, "correlation": 874
    }
  },
  {
    "ph": "f", "id": 874, "pid": 870, "tid": 870, "ts": 7429770613927.525,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770613930.393, "dur": 0.139,
    "args": {
      "External id": 2669,      "cid": 370, "correlation": 875
    }
  },
  {
    "ph": "f", "id": 875, "pid": 870, "tid": 870, "ts": 7429770613930.393,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770613932.304, "dur": 0.150,
    "args": {
      "External id": 2669,      "cid": 370, "correlation": 876
    }
  },
  {
    "ph": "f", "id": 876, "pid": 870, "tid": 870, "ts": 7429770613932.304,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613933.757, "dur": 5.897,
    "args": {
      "External id": 2669,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 877,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 877, "pid": 870, "tid": 870, "ts": 7429770613933.757,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613962.743, "dur": 5.140,
    "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": 878,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 878, "pid": 870, "tid": 870, "ts": 7429770613962.743,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770613989.385, "dur": 5.242,
    "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": 879,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 879, "pid": 870, "tid": 870, "ts": 7429770613989.385,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614008.435, "dur": 7.576,
    "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": 880,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 880, "pid": 870, "tid": 870, "ts": 7429770614008.435,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614023.931, "dur": 3.826,
    "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": 881,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 881, "pid": 870, "tid": 870, "ts": 7429770614023.931,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614032.970, "dur": 3.817,
    "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": 882,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 882, "pid": 870, "tid": 870, "ts": 7429770614032.970,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614042.987, "dur": 4.017,
    "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": 883,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 883, "pid": 870, "tid": 870, "ts": 7429770614042.987,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614058.875, "dur": 6.489,
    "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": 884,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 884, "pid": 870, "tid": 870, "ts": 7429770614058.875,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614091.474, "dur": 6.841,
    "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": 885,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 885, "pid": 870, "tid": 870, "ts": 7429770614091.474,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614108.214, "dur": 5.137,
    "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": 886,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 886, "pid": 870, "tid": 870, "ts": 7429770614108.214,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614119.553, "dur": 3.732,
    "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": 887,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 887, "pid": 870, "tid": 870, "ts": 7429770614119.553,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614184.747, "dur": 4.542,
    "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": 888,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 888, "pid": 870, "tid": 870, "ts": 7429770614184.747,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614220.001, "dur": 4.125,
    "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": 889,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 889, "pid": 870, "tid": 870, "ts": 7429770614220.001,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770614260.295, "dur": 0.254,
    "args": {
      "External id": 2744,      "cid": 370, "correlation": 890
    }
  },
  {
    "ph": "f", "id": 890, "pid": 870, "tid": 870, "ts": 7429770614260.295,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770614264.588, "dur": 0.164,
    "args": {
      "External id": 2744,      "cid": 370, "correlation": 891
    }
  },
  {
    "ph": "f", "id": 891, "pid": 870, "tid": 870, "ts": 7429770614264.588,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770614268.154, "dur": 0.153,
    "args": {
      "External id": 2744,      "cid": 370, "correlation": 892
    }
  },
  {
    "ph": "f", "id": 892, "pid": 870, "tid": 870, "ts": 7429770614268.154,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614271.396, "dur": 6.226,
    "args": {
      "External id": 2744,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 893,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 893, "pid": 870, "tid": 870, "ts": 7429770614271.396,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614289.749, "dur": 4.242,
    "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": 894,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 894, "pid": 870, "tid": 870, "ts": 7429770614289.749,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614308.654, "dur": 3.578,
    "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": 895,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 895, "pid": 870, "tid": 870, "ts": 7429770614308.654,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614324.975, "dur": 5.551,
    "args": {
      "External id": 2757,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 896,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 896, "pid": 870, "tid": 870, "ts": 7429770614324.975,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770614373.823, "dur": 0.204,
    "args": {
      "External id": 2769,      "cid": 370, "correlation": 897
    }
  },
  {
    "ph": "f", "id": 897, "pid": 870, "tid": 870, "ts": 7429770614373.823,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770614377.150, "dur": 0.144,
    "args": {
      "External id": 2769,      "cid": 370, "correlation": 898
    }
  },
  {
    "ph": "f", "id": 898, "pid": 870, "tid": 870, "ts": 7429770614377.150,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770614379.758, "dur": 0.167,
    "args": {
      "External id": 2769,      "cid": 370, "correlation": 899
    }
  },
  {
    "ph": "f", "id": 899, "pid": 870, "tid": 870, "ts": 7429770614379.758,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614382.095, "dur": 4.945,
    "args": {
      "External id": 2769,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 900,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 900, "pid": 870, "tid": 870, "ts": 7429770614382.095,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614402.967, "dur": 4.741,
    "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": 901,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 901, "pid": 870, "tid": 870, "ts": 7429770614402.967,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770614451.592, "dur": 0.188,
    "args": {
      "External id": 2788,      "cid": 370, "correlation": 902
    }
  },
  {
    "ph": "f", "id": 902, "pid": 870, "tid": 870, "ts": 7429770614451.592,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770614454.794, "dur": 0.150,
    "args": {
      "External id": 2788,      "cid": 370, "correlation": 903
    }
  },
  {
    "ph": "f", "id": 903, "pid": 870, "tid": 870, "ts": 7429770614454.794,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770614457.500, "dur": 0.361,
    "args": {
      "External id": 2788,      "cid": 370, "correlation": 904
    }
  },
  {
    "ph": "f", "id": 904, "pid": 870, "tid": 870, "ts": 7429770614457.500,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614459.448, "dur": 6.039,
    "args": {
      "External id": 2788,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 905,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 905, "pid": 870, "tid": 870, "ts": 7429770614459.448,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614477.310, "dur": 5.730,
    "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": 906,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 906, "pid": 870, "tid": 870, "ts": 7429770614477.310,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614500.150, "dur": 4.259,
    "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": 907,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 907, "pid": 870, "tid": 870, "ts": 7429770614500.150,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614513.922, "dur": 5.738,
    "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": 908,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 908, "pid": 870, "tid": 870, "ts": 7429770614513.922,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614528.419, "dur": 4.441,
    "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": 909,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 909, "pid": 870, "tid": 870, "ts": 7429770614528.419,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614538.964, "dur": 4.590,
    "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": 910,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 910, "pid": 870, "tid": 870, "ts": 7429770614538.964,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614550.928, "dur": 5.112,
    "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": 911,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 911, "pid": 870, "tid": 870, "ts": 7429770614550.928,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614564.405, "dur": 4.493,
    "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": 912,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 912, "pid": 870, "tid": 870, "ts": 7429770614564.405,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770614604.296, "dur": 0.175,
    "args": {
      "External id": 2809,      "cid": 370, "correlation": 913
    }
  },
  {
    "ph": "f", "id": 913, "pid": 870, "tid": 870, "ts": 7429770614604.296,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770614607.256, "dur": 0.137,
    "args": {
      "External id": 2809,      "cid": 370, "correlation": 914
    }
  },
  {
    "ph": "f", "id": 914, "pid": 870, "tid": 870, "ts": 7429770614607.256,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770614609.865, "dur": 0.142,
    "args": {
      "External id": 2809,      "cid": 370, "correlation": 915
    }
  },
  {
    "ph": "f", "id": 915, "pid": 870, "tid": 870, "ts": 7429770614609.865,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614612.007, "dur": 4.730,
    "args": {
      "External id": 2809,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 916,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 916, "pid": 870, "tid": 870, "ts": 7429770614612.007,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614633.082, "dur": 4.975,
    "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": 917,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 917, "pid": 870, "tid": 870, "ts": 7429770614633.082,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770614665.500, "dur": 0.173,
    "args": {
      "External id": 2820,      "cid": 370, "correlation": 918
    }
  },
  {
    "ph": "f", "id": 918, "pid": 870, "tid": 870, "ts": 7429770614665.500,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770614668.850, "dur": 0.158,
    "args": {
      "External id": 2820,      "cid": 370, "correlation": 919
    }
  },
  {
    "ph": "f", "id": 919, "pid": 870, "tid": 870, "ts": 7429770614668.850,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770614671.293, "dur": 0.140,
    "args": {
      "External id": 2820,      "cid": 370, "correlation": 920
    }
  },
  {
    "ph": "f", "id": 920, "pid": 870, "tid": 870, "ts": 7429770614671.293,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614672.928, "dur": 4.713,
    "args": {
      "External id": 2820,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 921,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 921, "pid": 870, "tid": 870, "ts": 7429770614672.928,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614686.024, "dur": 4.239,
    "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": 922,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 922, "pid": 870, "tid": 870, "ts": 7429770614686.024,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770614717.424, "dur": 0.182,
    "args": {
      "External id": 2831,      "cid": 370, "correlation": 923
    }
  },
  {
    "ph": "f", "id": 923, "pid": 870, "tid": 870, "ts": 7429770614717.424,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770614720.667, "dur": 0.138,
    "args": {
      "External id": 2831,      "cid": 370, "correlation": 924
    }
  },
  {
    "ph": "f", "id": 924, "pid": 870, "tid": 870, "ts": 7429770614720.667,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770614723.152, "dur": 0.140,
    "args": {
      "External id": 2831,      "cid": 370, "correlation": 925
    }
  },
  {
    "ph": "f", "id": 925, "pid": 870, "tid": 870, "ts": 7429770614723.152,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614731.457, "dur": 5.590,
    "args": {
      "External id": 2831,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 926,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 926, "pid": 870, "tid": 870, "ts": 7429770614731.457,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614737.443, "dur": 3.899,
    "args": {
      "External id": 2831,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 927,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 927, "pid": 870, "tid": 870, "ts": 7429770614737.443,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614751.023, "dur": 4.199,
    "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": 928,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 928, "pid": 870, "tid": 870, "ts": 7429770614751.023,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614780.997, "dur": 3.661,
    "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": 929,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 929, "pid": 870, "tid": 870, "ts": 7429770614780.997,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614793.215, "dur": 4.635,
    "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": 930,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 930, "pid": 870, "tid": 870, "ts": 7429770614793.215,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614804.759, "dur": 5.711,
    "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": 931,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 931, "pid": 870, "tid": 870, "ts": 7429770614804.759,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614816.623, "dur": 5.327,
    "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": 932,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 932, "pid": 870, "tid": 870, "ts": 7429770614816.623,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614827.545, "dur": 4.043,
    "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": 933,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 933, "pid": 870, "tid": 870, "ts": 7429770614827.545,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614839.843, "dur": 3.547,
    "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": 934,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 934, "pid": 870, "tid": 870, "ts": 7429770614839.843,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770614882.293, "dur": 0.292,
    "args": {
      "External id": 2852,      "cid": 370, "correlation": 935
    }
  },
  {
    "ph": "f", "id": 935, "pid": 870, "tid": 870, "ts": 7429770614882.293,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770614886.100, "dur": 0.142,
    "args": {
      "External id": 2852,      "cid": 370, "correlation": 936
    }
  },
  {
    "ph": "f", "id": 936, "pid": 870, "tid": 870, "ts": 7429770614886.100,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770614888.462, "dur": 0.154,
    "args": {
      "External id": 2852,      "cid": 370, "correlation": 937
    }
  },
  {
    "ph": "f", "id": 937, "pid": 870, "tid": 870, "ts": 7429770614888.462,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614890.618, "dur": 4.913,
    "args": {
      "External id": 2852,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 938,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 938, "pid": 870, "tid": 870, "ts": 7429770614890.618,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770614931.300, "dur": 0.173,
    "args": {
      "External id": 2865,      "cid": 370, "correlation": 939
    }
  },
  {
    "ph": "f", "id": 939, "pid": 870, "tid": 870, "ts": 7429770614931.300,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770614934.580, "dur": 0.155,
    "args": {
      "External id": 2865,      "cid": 370, "correlation": 940
    }
  },
  {
    "ph": "f", "id": 940, "pid": 870, "tid": 870, "ts": 7429770614934.580,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770614936.817, "dur": 0.153,
    "args": {
      "External id": 2865,      "cid": 370, "correlation": 941
    }
  },
  {
    "ph": "f", "id": 941, "pid": 870, "tid": 870, "ts": 7429770614936.817,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614938.372, "dur": 5.276,
    "args": {
      "External id": 2865,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 942,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 942, "pid": 870, "tid": 870, "ts": 7429770614938.372,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770614975.400, "dur": 0.165,
    "args": {
      "External id": 2878,      "cid": 370, "correlation": 943
    }
  },
  {
    "ph": "f", "id": 943, "pid": 870, "tid": 870, "ts": 7429770614975.400,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770614978.523, "dur": 0.158,
    "args": {
      "External id": 2878,      "cid": 370, "correlation": 944
    }
  },
  {
    "ph": "f", "id": 944, "pid": 870, "tid": 870, "ts": 7429770614978.523,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770614980.513, "dur": 0.152,
    "args": {
      "External id": 2878,      "cid": 370, "correlation": 945
    }
  },
  {
    "ph": "f", "id": 945, "pid": 870, "tid": 870, "ts": 7429770614980.513,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770614981.945, "dur": 4.709,
    "args": {
      "External id": 2878,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 946,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 946, "pid": 870, "tid": 870, "ts": 7429770614981.945,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615008.465, "dur": 3.904,
    "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": 947,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 947, "pid": 870, "tid": 870, "ts": 7429770615008.465,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615033.160, "dur": 4.478,
    "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": 948,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 948, "pid": 870, "tid": 870, "ts": 7429770615033.160,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615050.874, "dur": 5.922,
    "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": 949,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 949, "pid": 870, "tid": 870, "ts": 7429770615050.874,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615065.337, "dur": 4.182,
    "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": 950,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 950, "pid": 870, "tid": 870, "ts": 7429770615065.337,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615074.737, "dur": 5.071,
    "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": 951,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 951, "pid": 870, "tid": 870, "ts": 7429770615074.737,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615085.752, "dur": 5.295,
    "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": 952,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 952, "pid": 870, "tid": 870, "ts": 7429770615085.752,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615102.822, "dur": 5.651,
    "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": 953,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 953, "pid": 870, "tid": 870, "ts": 7429770615102.822,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615116.572, "dur": 5.282,
    "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": 954,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 954, "pid": 870, "tid": 870, "ts": 7429770615116.572,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615127.385, "dur": 3.980,
    "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": 955,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 955, "pid": 870, "tid": 870, "ts": 7429770615127.385,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615136.107, "dur": 3.056,
    "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": 956,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 956, "pid": 870, "tid": 870, "ts": 7429770615136.107,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615194.285, "dur": 4.007,
    "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": 957,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 957, "pid": 870, "tid": 870, "ts": 7429770615194.285,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615226.629, "dur": 3.466,
    "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": 958,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 958, "pid": 870, "tid": 870, "ts": 7429770615226.629,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770615263.268, "dur": 0.219,
    "args": {
      "External id": 2953,      "cid": 370, "correlation": 959
    }
  },
  {
    "ph": "f", "id": 959, "pid": 870, "tid": 870, "ts": 7429770615263.268,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770615267.440, "dur": 0.157,
    "args": {
      "External id": 2953,      "cid": 370, "correlation": 960
    }
  },
  {
    "ph": "f", "id": 960, "pid": 870, "tid": 870, "ts": 7429770615267.440,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770615269.931, "dur": 0.142,
    "args": {
      "External id": 2953,      "cid": 370, "correlation": 961
    }
  },
  {
    "ph": "f", "id": 961, "pid": 870, "tid": 870, "ts": 7429770615269.931,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615272.257, "dur": 5.398,
    "args": {
      "External id": 2953,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 962,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 962, "pid": 870, "tid": 870, "ts": 7429770615272.257,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615289.388, "dur": 4.125,
    "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": 963,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 963, "pid": 870, "tid": 870, "ts": 7429770615289.388,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615307.167, "dur": 3.671,
    "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": 964,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 964, "pid": 870, "tid": 870, "ts": 7429770615307.167,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615321.742, "dur": 4.921,
    "args": {
      "External id": 2966,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 965,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 965, "pid": 870, "tid": 870, "ts": 7429770615321.742,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770615367.647, "dur": 0.190,
    "args": {
      "External id": 2978,      "cid": 370, "correlation": 966
    }
  },
  {
    "ph": "f", "id": 966, "pid": 870, "tid": 870, "ts": 7429770615367.647,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770615370.923, "dur": 0.139,
    "args": {
      "External id": 2978,      "cid": 370, "correlation": 967
    }
  },
  {
    "ph": "f", "id": 967, "pid": 870, "tid": 870, "ts": 7429770615370.923,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770615373.538, "dur": 0.146,
    "args": {
      "External id": 2978,      "cid": 370, "correlation": 968
    }
  },
  {
    "ph": "f", "id": 968, "pid": 870, "tid": 870, "ts": 7429770615373.538,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615375.574, "dur": 4.382,
    "args": {
      "External id": 2978,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 969,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 969, "pid": 870, "tid": 870, "ts": 7429770615375.574,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615395.346, "dur": 4.769,
    "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": 970,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 970, "pid": 870, "tid": 870, "ts": 7429770615395.346,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770615440.834, "dur": 0.176,
    "args": {
      "External id": 2997,      "cid": 370, "correlation": 971
    }
  },
  {
    "ph": "f", "id": 971, "pid": 870, "tid": 870, "ts": 7429770615440.834,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770615444.040, "dur": 0.151,
    "args": {
      "External id": 2997,      "cid": 370, "correlation": 972
    }
  },
  {
    "ph": "f", "id": 972, "pid": 870, "tid": 870, "ts": 7429770615444.040,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770615446.135, "dur": 0.418,
    "args": {
      "External id": 2997,      "cid": 370, "correlation": 973
    }
  },
  {
    "ph": "f", "id": 973, "pid": 870, "tid": 870, "ts": 7429770615446.135,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615448.056, "dur": 5.358,
    "args": {
      "External id": 2997,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 974,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 974, "pid": 870, "tid": 870, "ts": 7429770615448.056,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615465.266, "dur": 4.132,
    "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": 975,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 975, "pid": 870, "tid": 870, "ts": 7429770615465.266,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615487.452, "dur": 5.385,
    "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": 976,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 976, "pid": 870, "tid": 870, "ts": 7429770615487.452,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615500.958, "dur": 5.919,
    "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": 977,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 977, "pid": 870, "tid": 870, "ts": 7429770615500.958,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615513.984, "dur": 4.437,
    "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": 978,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 978, "pid": 870, "tid": 870, "ts": 7429770615513.984,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615523.441, "dur": 3.920,
    "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": 979,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 979, "pid": 870, "tid": 870, "ts": 7429770615523.441,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615534.818, "dur": 5.065,
    "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": 980,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 980, "pid": 870, "tid": 870, "ts": 7429770615534.818,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615547.796, "dur": 6.350,
    "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": 981,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 981, "pid": 870, "tid": 870, "ts": 7429770615547.796,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770615591.287, "dur": 0.168,
    "args": {
      "External id": 3018,      "cid": 370, "correlation": 982
    }
  },
  {
    "ph": "f", "id": 982, "pid": 870, "tid": 870, "ts": 7429770615591.287,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770615594.451, "dur": 0.138,
    "args": {
      "External id": 3018,      "cid": 370, "correlation": 983
    }
  },
  {
    "ph": "f", "id": 983, "pid": 870, "tid": 870, "ts": 7429770615594.451,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770615597.178, "dur": 0.138,
    "args": {
      "External id": 3018,      "cid": 370, "correlation": 984
    }
  },
  {
    "ph": "f", "id": 984, "pid": 870, "tid": 870, "ts": 7429770615597.178,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615599.140, "dur": 4.687,
    "args": {
      "External id": 3018,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 985,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 985, "pid": 870, "tid": 870, "ts": 7429770615599.140,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615619.575, "dur": 4.858,
    "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": 986,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 986, "pid": 870, "tid": 870, "ts": 7429770615619.575,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770615653.292, "dur": 0.492,
    "args": {
      "External id": 3030,      "cid": 290, "correlation": 987
    }
  },
  {
    "ph": "f", "id": 987, "pid": 870, "tid": 870, "ts": 7429770615653.292,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770615654.442, "dur": 70.945,
    "args": {
      "External id": 3030,            "bytes": 79691776, "cid": 108, "correlation": 988, "ptr": "0x7ffd8041f0c0"
    }
  },
  {
    "ph": "s", "id": 988, "pid": 870, "tid": 870, "ts": 7429770615654.442,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770615727.481, "dur": 0.178,
    "args": {
      "External id": 3029,      "cid": 370, "correlation": 989
    }
  },
  {
    "ph": "f", "id": 989, "pid": 870, "tid": 870, "ts": 7429770615727.481,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770615731.569, "dur": 0.141,
    "args": {
      "External id": 3029,      "cid": 370, "correlation": 990
    }
  },
  {
    "ph": "f", "id": 990, "pid": 870, "tid": 870, "ts": 7429770615731.569,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770615734.069, "dur": 0.143,
    "args": {
      "External id": 3029,      "cid": 370, "correlation": 991
    }
  },
  {
    "ph": "f", "id": 991, "pid": 870, "tid": 870, "ts": 7429770615734.069,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615737.577, "dur": 6.003,
    "args": {
      "External id": 3029,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 992,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 992, "pid": 870, "tid": 870, "ts": 7429770615737.577,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615755.096, "dur": 4.464,
    "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": 993,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 993, "pid": 870, "tid": 870, "ts": 7429770615755.096,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770615789.796, "dur": 0.169,
    "args": {
      "External id": 3040,      "cid": 370, "correlation": 994
    }
  },
  {
    "ph": "f", "id": 994, "pid": 870, "tid": 870, "ts": 7429770615789.796,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770615792.629, "dur": 0.137,
    "args": {
      "External id": 3040,      "cid": 370, "correlation": 995
    }
  },
  {
    "ph": "f", "id": 995, "pid": 870, "tid": 870, "ts": 7429770615792.629,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770615795.326, "dur": 0.140,
    "args": {
      "External id": 3040,      "cid": 370, "correlation": 996
    }
  },
  {
    "ph": "f", "id": 996, "pid": 870, "tid": 870, "ts": 7429770615795.326,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615804.100, "dur": 5.685,
    "args": {
      "External id": 3040,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 997,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 997, "pid": 870, "tid": 870, "ts": 7429770615804.100,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615810.180, "dur": 4.450,
    "args": {
      "External id": 3040,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 998,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 998, "pid": 870, "tid": 870, "ts": 7429770615810.180,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615824.677, "dur": 5.198,
    "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": 999,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 999, "pid": 870, "tid": 870, "ts": 7429770615824.677,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615858.577, "dur": 4.030,
    "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": 1000,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1000, "pid": 870, "tid": 870, "ts": 7429770615858.577,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615871.575, "dur": 4.652,
    "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": 1001,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1001, "pid": 870, "tid": 870, "ts": 7429770615871.575,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615883.960, "dur": 4.826,
    "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": 1002,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1002, "pid": 870, "tid": 870, "ts": 7429770615883.960,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615894.289, "dur": 6.211,
    "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": 1003,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1003, "pid": 870, "tid": 870, "ts": 7429770615894.289,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615906.863, "dur": 4.345,
    "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": 1004,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1004, "pid": 870, "tid": 870, "ts": 7429770615906.863,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615919.888, "dur": 3.978,
    "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": 1005,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1005, "pid": 870, "tid": 870, "ts": 7429770615919.888,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770615964.619, "dur": 0.177,
    "args": {
      "External id": 3061,      "cid": 370, "correlation": 1006
    }
  },
  {
    "ph": "f", "id": 1006, "pid": 870, "tid": 870, "ts": 7429770615964.619,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770615968.187, "dur": 0.155,
    "args": {
      "External id": 3061,      "cid": 370, "correlation": 1007
    }
  },
  {
    "ph": "f", "id": 1007, "pid": 870, "tid": 870, "ts": 7429770615968.187,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770615970.773, "dur": 0.149,
    "args": {
      "External id": 3061,      "cid": 370, "correlation": 1008
    }
  },
  {
    "ph": "f", "id": 1008, "pid": 870, "tid": 870, "ts": 7429770615970.773,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770615972.734, "dur": 5.002,
    "args": {
      "External id": 3061,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1009,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1009, "pid": 870, "tid": 870, "ts": 7429770615972.734,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770616042.527, "dur": 0.180,
    "args": {
      "External id": 3074,      "cid": 370, "correlation": 1010
    }
  },
  {
    "ph": "f", "id": 1010, "pid": 870, "tid": 870, "ts": 7429770616042.527,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770616046.021, "dur": 0.178,
    "args": {
      "External id": 3074,      "cid": 370, "correlation": 1011
    }
  },
  {
    "ph": "f", "id": 1011, "pid": 870, "tid": 870, "ts": 7429770616046.021,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770616048.269, "dur": 0.150,
    "args": {
      "External id": 3074,      "cid": 370, "correlation": 1012
    }
  },
  {
    "ph": "f", "id": 1012, "pid": 870, "tid": 870, "ts": 7429770616048.269,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616049.863, "dur": 5.528,
    "args": {
      "External id": 3074,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1013,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1013, "pid": 870, "tid": 870, "ts": 7429770616049.863,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770616086.776, "dur": 0.180,
    "args": {
      "External id": 3087,      "cid": 370, "correlation": 1014
    }
  },
  {
    "ph": "f", "id": 1014, "pid": 870, "tid": 870, "ts": 7429770616086.776,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770616089.373, "dur": 0.163,
    "args": {
      "External id": 3087,      "cid": 370, "correlation": 1015
    }
  },
  {
    "ph": "f", "id": 1015, "pid": 870, "tid": 870, "ts": 7429770616089.373,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770616091.947, "dur": 0.157,
    "args": {
      "External id": 3087,      "cid": 370, "correlation": 1016
    }
  },
  {
    "ph": "f", "id": 1016, "pid": 870, "tid": 870, "ts": 7429770616091.947,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616093.432, "dur": 5.604,
    "args": {
      "External id": 3087,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1017,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1017, "pid": 870, "tid": 870, "ts": 7429770616093.432,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616120.439, "dur": 4.267,
    "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": 1018,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1018, "pid": 870, "tid": 870, "ts": 7429770616120.439,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616147.135, "dur": 4.638,
    "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": 1019,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1019, "pid": 870, "tid": 870, "ts": 7429770616147.135,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616165.331, "dur": 6.702,
    "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": 1020,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1020, "pid": 870, "tid": 870, "ts": 7429770616165.331,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616178.680, "dur": 3.909,
    "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": 1021,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1021, "pid": 870, "tid": 870, "ts": 7429770616178.680,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616187.614, "dur": 5.489,
    "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": 1022,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1022, "pid": 870, "tid": 870, "ts": 7429770616187.614,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616199.378, "dur": 3.646,
    "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": 1023,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1023, "pid": 870, "tid": 870, "ts": 7429770616199.378,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616216.126, "dur": 5.276,
    "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": 1024,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1024, "pid": 870, "tid": 870, "ts": 7429770616216.126,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616227.590, "dur": 5.685,
    "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": 1025,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1025, "pid": 870, "tid": 870, "ts": 7429770616227.590,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616245.558, "dur": 3.225,
    "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": 1026,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1026, "pid": 870, "tid": 870, "ts": 7429770616245.558,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616253.097, "dur": 3.075,
    "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": 1027,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1027, "pid": 870, "tid": 870, "ts": 7429770616253.097,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616311.146, "dur": 4.773,
    "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": 1028,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1028, "pid": 870, "tid": 870, "ts": 7429770616311.146,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616342.204, "dur": 4.395,
    "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": 1029,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1029, "pid": 870, "tid": 870, "ts": 7429770616342.204,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770616369.301, "dur": 0.414,
    "args": {
      "External id": 3162,      "cid": 290, "correlation": 1030
    }
  },
  {
    "ph": "f", "id": 1030, "pid": 870, "tid": 870, "ts": 7429770616369.301,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770616370.323, "dur": 53.423,
    "args": {
      "External id": 3162,            "bytes": 2097152, "cid": 108, "correlation": 1031, "ptr": "0x7ffd80420410"
    }
  },
  {
    "ph": "s", "id": 1031, "pid": 870, "tid": 870, "ts": 7429770616370.323,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770616439.384, "dur": 0.306,
    "args": {
      "External id": 3162,      "cid": 370, "correlation": 1032
    }
  },
  {
    "ph": "f", "id": 1032, "pid": 870, "tid": 870, "ts": 7429770616439.384,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770616443.701, "dur": 0.172,
    "args": {
      "External id": 3162,      "cid": 370, "correlation": 1033
    }
  },
  {
    "ph": "f", "id": 1033, "pid": 870, "tid": 870, "ts": 7429770616443.701,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770616447.021, "dur": 0.181,
    "args": {
      "External id": 3162,      "cid": 370, "correlation": 1034
    }
  },
  {
    "ph": "f", "id": 1034, "pid": 870, "tid": 870, "ts": 7429770616447.021,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616450.856, "dur": 6.579,
    "args": {
      "External id": 3162,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1035,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1035, "pid": 870, "tid": 870, "ts": 7429770616450.856,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616471.853, "dur": 4.840,
    "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": 1036,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1036, "pid": 870, "tid": 870, "ts": 7429770616471.853,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616492.383, "dur": 3.893,
    "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": 1037,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1037, "pid": 870, "tid": 870, "ts": 7429770616492.383,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616509.352, "dur": 5.869,
    "args": {
      "External id": 3175,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 1038,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1038, "pid": 870, "tid": 870, "ts": 7429770616509.352,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770616558.610, "dur": 0.188,
    "args": {
      "External id": 3187,      "cid": 370, "correlation": 1039
    }
  },
  {
    "ph": "f", "id": 1039, "pid": 870, "tid": 870, "ts": 7429770616558.610,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770616562.261, "dur": 0.140,
    "args": {
      "External id": 3187,      "cid": 370, "correlation": 1040
    }
  },
  {
    "ph": "f", "id": 1040, "pid": 870, "tid": 870, "ts": 7429770616562.261,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770616564.982, "dur": 0.146,
    "args": {
      "External id": 3187,      "cid": 370, "correlation": 1041
    }
  },
  {
    "ph": "f", "id": 1041, "pid": 870, "tid": 870, "ts": 7429770616564.982,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616567.222, "dur": 5.162,
    "args": {
      "External id": 3187,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1042,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1042, "pid": 870, "tid": 870, "ts": 7429770616567.222,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616589.529, "dur": 4.662,
    "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": 1043,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1043, "pid": 870, "tid": 870, "ts": 7429770616589.529,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770616639.560, "dur": 0.225,
    "args": {
      "External id": 3206,      "cid": 370, "correlation": 1044
    }
  },
  {
    "ph": "f", "id": 1044, "pid": 870, "tid": 870, "ts": 7429770616639.560,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770616642.984, "dur": 0.185,
    "args": {
      "External id": 3206,      "cid": 370, "correlation": 1045
    }
  },
  {
    "ph": "f", "id": 1045, "pid": 870, "tid": 870, "ts": 7429770616642.984,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770616645.547, "dur": 0.359,
    "args": {
      "External id": 3206,      "cid": 370, "correlation": 1046
    }
  },
  {
    "ph": "f", "id": 1046, "pid": 870, "tid": 870, "ts": 7429770616645.547,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616647.512, "dur": 6.542,
    "args": {
      "External id": 3206,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1047,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1047, "pid": 870, "tid": 870, "ts": 7429770616647.512,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616667.215, "dur": 5.393,
    "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": 1048,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1048, "pid": 870, "tid": 870, "ts": 7429770616667.215,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616692.178, "dur": 4.833,
    "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": 1049,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1049, "pid": 870, "tid": 870, "ts": 7429770616692.178,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616707.217, "dur": 6.449,
    "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": 1050,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1050, "pid": 870, "tid": 870, "ts": 7429770616707.217,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616722.316, "dur": 4.114,
    "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": 1051,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1051, "pid": 870, "tid": 870, "ts": 7429770616722.316,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616732.349, "dur": 5.297,
    "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": 1052,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1052, "pid": 870, "tid": 870, "ts": 7429770616732.349,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616744.857, "dur": 5.228,
    "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": 1053,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1053, "pid": 870, "tid": 870, "ts": 7429770616744.857,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616757.881, "dur": 5.842,
    "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": 1054,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1054, "pid": 870, "tid": 870, "ts": 7429770616757.881,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770616802.398, "dur": 0.176,
    "args": {
      "External id": 3227,      "cid": 370, "correlation": 1055
    }
  },
  {
    "ph": "f", "id": 1055, "pid": 870, "tid": 870, "ts": 7429770616802.398,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770616805.330, "dur": 0.155,
    "args": {
      "External id": 3227,      "cid": 370, "correlation": 1056
    }
  },
  {
    "ph": "f", "id": 1056, "pid": 870, "tid": 870, "ts": 7429770616805.330,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770616807.882, "dur": 0.135,
    "args": {
      "External id": 3227,      "cid": 370, "correlation": 1057
    }
  },
  {
    "ph": "f", "id": 1057, "pid": 870, "tid": 870, "ts": 7429770616807.882,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616810.013, "dur": 5.522,
    "args": {
      "External id": 3227,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1058,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1058, "pid": 870, "tid": 870, "ts": 7429770616810.013,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616832.553, "dur": 6.087,
    "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": 1059,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1059, "pid": 870, "tid": 870, "ts": 7429770616832.553,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770616865.452, "dur": 0.190,
    "args": {
      "External id": 3238,      "cid": 370, "correlation": 1060
    }
  },
  {
    "ph": "f", "id": 1060, "pid": 870, "tid": 870, "ts": 7429770616865.452,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770616868.327, "dur": 0.159,
    "args": {
      "External id": 3238,      "cid": 370, "correlation": 1061
    }
  },
  {
    "ph": "f", "id": 1061, "pid": 870, "tid": 870, "ts": 7429770616868.327,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770616870.806, "dur": 0.137,
    "args": {
      "External id": 3238,      "cid": 370, "correlation": 1062
    }
  },
  {
    "ph": "f", "id": 1062, "pid": 870, "tid": 870, "ts": 7429770616870.806,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616872.428, "dur": 5.245,
    "args": {
      "External id": 3238,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1063,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1063, "pid": 870, "tid": 870, "ts": 7429770616872.428,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616886.224, "dur": 4.213,
    "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": 1064,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1064, "pid": 870, "tid": 870, "ts": 7429770616886.224,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770616914.259, "dur": 0.158,
    "args": {
      "External id": 3249,      "cid": 370, "correlation": 1065
    }
  },
  {
    "ph": "f", "id": 1065, "pid": 870, "tid": 870, "ts": 7429770616914.259,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770616918.080, "dur": 0.141,
    "args": {
      "External id": 3249,      "cid": 370, "correlation": 1066
    }
  },
  {
    "ph": "f", "id": 1066, "pid": 870, "tid": 870, "ts": 7429770616918.080,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770616920.342, "dur": 0.139,
    "args": {
      "External id": 3249,      "cid": 370, "correlation": 1067
    }
  },
  {
    "ph": "f", "id": 1067, "pid": 870, "tid": 870, "ts": 7429770616920.342,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616928.694, "dur": 6.035,
    "args": {
      "External id": 3249,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1068,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1068, "pid": 870, "tid": 870, "ts": 7429770616928.694,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616935.112, "dur": 6.016,
    "args": {
      "External id": 3249,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1069,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1069, "pid": 870, "tid": 870, "ts": 7429770616935.112,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616951.057, "dur": 3.824,
    "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": 1070,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1070, "pid": 870, "tid": 870, "ts": 7429770616951.057,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616979.761, "dur": 4.355,
    "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": 1071,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1071, "pid": 870, "tid": 870, "ts": 7429770616979.761,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770616992.771, "dur": 5.791,
    "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": 1072,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1072, "pid": 870, "tid": 870, "ts": 7429770616992.771,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617005.347, "dur": 4.862,
    "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": 1073,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1073, "pid": 870, "tid": 870, "ts": 7429770617005.347,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617015.971, "dur": 4.705,
    "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": 1074,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1074, "pid": 870, "tid": 870, "ts": 7429770617015.971,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617026.665, "dur": 4.257,
    "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": 1075,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1075, "pid": 870, "tid": 870, "ts": 7429770617026.665,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617039.466, "dur": 3.708,
    "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": 1076,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1076, "pid": 870, "tid": 870, "ts": 7429770617039.466,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770617082.837, "dur": 0.173,
    "args": {
      "External id": 3270,      "cid": 370, "correlation": 1077
    }
  },
  {
    "ph": "f", "id": 1077, "pid": 870, "tid": 870, "ts": 7429770617082.837,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770617086.954, "dur": 0.154,
    "args": {
      "External id": 3270,      "cid": 370, "correlation": 1078
    }
  },
  {
    "ph": "f", "id": 1078, "pid": 870, "tid": 870, "ts": 7429770617086.954,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770617089.290, "dur": 0.150,
    "args": {
      "External id": 3270,      "cid": 370, "correlation": 1079
    }
  },
  {
    "ph": "f", "id": 1079, "pid": 870, "tid": 870, "ts": 7429770617089.290,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617091.375, "dur": 5.067,
    "args": {
      "External id": 3270,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1080,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1080, "pid": 870, "tid": 870, "ts": 7429770617091.375,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770617133.323, "dur": 0.184,
    "args": {
      "External id": 3283,      "cid": 370, "correlation": 1081
    }
  },
  {
    "ph": "f", "id": 1081, "pid": 870, "tid": 870, "ts": 7429770617133.323,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770617136.627, "dur": 0.174,
    "args": {
      "External id": 3283,      "cid": 370, "correlation": 1082
    }
  },
  {
    "ph": "f", "id": 1082, "pid": 870, "tid": 870, "ts": 7429770617136.627,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770617138.870, "dur": 0.148,
    "args": {
      "External id": 3283,      "cid": 370, "correlation": 1083
    }
  },
  {
    "ph": "f", "id": 1083, "pid": 870, "tid": 870, "ts": 7429770617138.870,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617140.606, "dur": 4.826,
    "args": {
      "External id": 3283,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1084,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1084, "pid": 870, "tid": 870, "ts": 7429770617140.606,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770617179.032, "dur": 0.166,
    "args": {
      "External id": 3296,      "cid": 370, "correlation": 1085
    }
  },
  {
    "ph": "f", "id": 1085, "pid": 870, "tid": 870, "ts": 7429770617179.032,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770617182.004, "dur": 0.162,
    "args": {
      "External id": 3296,      "cid": 370, "correlation": 1086
    }
  },
  {
    "ph": "f", "id": 1086, "pid": 870, "tid": 870, "ts": 7429770617182.004,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770617183.667, "dur": 0.149,
    "args": {
      "External id": 3296,      "cid": 370, "correlation": 1087
    }
  },
  {
    "ph": "f", "id": 1087, "pid": 870, "tid": 870, "ts": 7429770617183.667,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617185.117, "dur": 6.074,
    "args": {
      "External id": 3296,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1088,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1088, "pid": 870, "tid": 870, "ts": 7429770617185.117,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617213.315, "dur": 5.167,
    "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": 1089,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1089, "pid": 870, "tid": 870, "ts": 7429770617213.315,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617240.366, "dur": 5.578,
    "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": 1090,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1090, "pid": 870, "tid": 870, "ts": 7429770617240.366,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617259.205, "dur": 7.360,
    "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": 1091,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1091, "pid": 870, "tid": 870, "ts": 7429770617259.205,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617274.801, "dur": 4.314,
    "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": 1092,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1092, "pid": 870, "tid": 870, "ts": 7429770617274.801,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617284.656, "dur": 3.749,
    "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": 1093,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1093, "pid": 870, "tid": 870, "ts": 7429770617284.656,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617295.454, "dur": 4.661,
    "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": 1094,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1094, "pid": 870, "tid": 870, "ts": 7429770617295.454,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617311.900, "dur": 5.489,
    "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": 1095,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1095, "pid": 870, "tid": 870, "ts": 7429770617311.900,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617323.563, "dur": 7.115,
    "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": 1096,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1096, "pid": 870, "tid": 870, "ts": 7429770617323.563,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617336.403, "dur": 4.231,
    "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": 1097,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1097, "pid": 870, "tid": 870, "ts": 7429770617336.403,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617345.753, "dur": 3.564,
    "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": 1098,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1098, "pid": 870, "tid": 870, "ts": 7429770617345.753,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617407.281, "dur": 4.718,
    "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": 1099,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1099, "pid": 870, "tid": 870, "ts": 7429770617407.281,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617441.364, "dur": 3.835,
    "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": 1100,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1100, "pid": 870, "tid": 870, "ts": 7429770617441.364,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770617469.276, "dur": 0.433,
    "args": {
      "External id": 3371,      "cid": 290, "correlation": 1101
    }
  },
  {
    "ph": "f", "id": 1101, "pid": 870, "tid": 870, "ts": 7429770617469.276,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770617470.394, "dur": 56.176,
    "args": {
      "External id": 3371,            "bytes": 2097152, "cid": 108, "correlation": 1102, "ptr": "0x7ffd80420410"
    }
  },
  {
    "ph": "s", "id": 1102, "pid": 870, "tid": 870, "ts": 7429770617470.394,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770617543.316, "dur": 0.247,
    "args": {
      "External id": 3371,      "cid": 370, "correlation": 1103
    }
  },
  {
    "ph": "f", "id": 1103, "pid": 870, "tid": 870, "ts": 7429770617543.316,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770617547.779, "dur": 0.173,
    "args": {
      "External id": 3371,      "cid": 370, "correlation": 1104
    }
  },
  {
    "ph": "f", "id": 1104, "pid": 870, "tid": 870, "ts": 7429770617547.779,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770617551.025, "dur": 0.206,
    "args": {
      "External id": 3371,      "cid": 370, "correlation": 1105
    }
  },
  {
    "ph": "f", "id": 1105, "pid": 870, "tid": 870, "ts": 7429770617551.025,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617554.630, "dur": 5.844,
    "args": {
      "External id": 3371,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1106,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1106, "pid": 870, "tid": 870, "ts": 7429770617554.630,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617575.377, "dur": 5.353,
    "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": 1107,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1107, "pid": 870, "tid": 870, "ts": 7429770617575.377,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617596.574, "dur": 4.194,
    "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": 1108,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1108, "pid": 870, "tid": 870, "ts": 7429770617596.574,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617613.389, "dur": 5.883,
    "args": {
      "External id": 3384,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 1109,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1109, "pid": 870, "tid": 870, "ts": 7429770617613.389,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770617664.140, "dur": 0.201,
    "args": {
      "External id": 3396,      "cid": 370, "correlation": 1110
    }
  },
  {
    "ph": "f", "id": 1110, "pid": 870, "tid": 870, "ts": 7429770617664.140,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770617667.338, "dur": 0.140,
    "args": {
      "External id": 3396,      "cid": 370, "correlation": 1111
    }
  },
  {
    "ph": "f", "id": 1111, "pid": 870, "tid": 870, "ts": 7429770617667.338,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770617669.897, "dur": 0.143,
    "args": {
      "External id": 3396,      "cid": 370, "correlation": 1112
    }
  },
  {
    "ph": "f", "id": 1112, "pid": 870, "tid": 870, "ts": 7429770617669.897,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617672.159, "dur": 5.439,
    "args": {
      "External id": 3396,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1113,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1113, "pid": 870, "tid": 870, "ts": 7429770617672.159,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617696.473, "dur": 4.379,
    "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": 1114,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1114, "pid": 870, "tid": 870, "ts": 7429770617696.473,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770617747.506, "dur": 0.207,
    "args": {
      "External id": 3415,      "cid": 370, "correlation": 1115
    }
  },
  {
    "ph": "f", "id": 1115, "pid": 870, "tid": 870, "ts": 7429770617747.506,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770617751.267, "dur": 0.181,
    "args": {
      "External id": 3415,      "cid": 370, "correlation": 1116
    }
  },
  {
    "ph": "f", "id": 1116, "pid": 870, "tid": 870, "ts": 7429770617751.267,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770617753.944, "dur": 0.359,
    "args": {
      "External id": 3415,      "cid": 370, "correlation": 1117
    }
  },
  {
    "ph": "f", "id": 1117, "pid": 870, "tid": 870, "ts": 7429770617753.944,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617756.177, "dur": 5.574,
    "args": {
      "External id": 3415,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1118,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1118, "pid": 870, "tid": 870, "ts": 7429770617756.177,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617775.604, "dur": 4.637,
    "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": 1119,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1119, "pid": 870, "tid": 870, "ts": 7429770617775.604,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617798.372, "dur": 4.468,
    "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": 1120,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1120, "pid": 870, "tid": 870, "ts": 7429770617798.372,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617812.049, "dur": 5.896,
    "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": 1121,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1121, "pid": 870, "tid": 870, "ts": 7429770617812.049,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617827.245, "dur": 4.949,
    "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": 1122,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1122, "pid": 870, "tid": 870, "ts": 7429770617827.245,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617839.176, "dur": 5.116,
    "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": 1123,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1123, "pid": 870, "tid": 870, "ts": 7429770617839.176,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617851.177, "dur": 5.089,
    "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": 1124,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1124, "pid": 870, "tid": 870, "ts": 7429770617851.177,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617864.591, "dur": 4.691,
    "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": 1125,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1125, "pid": 870, "tid": 870, "ts": 7429770617864.591,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770617910.377, "dur": 0.186,
    "args": {
      "External id": 3436,      "cid": 370, "correlation": 1126
    }
  },
  {
    "ph": "f", "id": 1126, "pid": 870, "tid": 870, "ts": 7429770617910.377,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770617913.696, "dur": 0.173,
    "args": {
      "External id": 3436,      "cid": 370, "correlation": 1127
    }
  },
  {
    "ph": "f", "id": 1127, "pid": 870, "tid": 870, "ts": 7429770617913.696,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770617916.160, "dur": 0.139,
    "args": {
      "External id": 3436,      "cid": 370, "correlation": 1128
    }
  },
  {
    "ph": "f", "id": 1128, "pid": 870, "tid": 870, "ts": 7429770617916.160,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617918.103, "dur": 5.412,
    "args": {
      "External id": 3436,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1129,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1129, "pid": 870, "tid": 870, "ts": 7429770617918.103,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617940.650, "dur": 5.408,
    "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": 1130,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1130, "pid": 870, "tid": 870, "ts": 7429770617940.650,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770617971.996, "dur": 0.159,
    "args": {
      "External id": 3447,      "cid": 370, "correlation": 1131
    }
  },
  {
    "ph": "f", "id": 1131, "pid": 870, "tid": 870, "ts": 7429770617971.996,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770617975.040, "dur": 0.139,
    "args": {
      "External id": 3447,      "cid": 370, "correlation": 1132
    }
  },
  {
    "ph": "f", "id": 1132, "pid": 870, "tid": 870, "ts": 7429770617975.040,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770617977.014, "dur": 0.137,
    "args": {
      "External id": 3447,      "cid": 370, "correlation": 1133
    }
  },
  {
    "ph": "f", "id": 1133, "pid": 870, "tid": 870, "ts": 7429770617977.014,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617978.840, "dur": 7.618,
    "args": {
      "External id": 3447,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1134,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1134, "pid": 870, "tid": 870, "ts": 7429770617978.840,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770617995.080, "dur": 4.099,
    "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": 1135,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1135, "pid": 870, "tid": 870, "ts": 7429770617995.080,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770618022.658, "dur": 0.168,
    "args": {
      "External id": 3458,      "cid": 370, "correlation": 1136
    }
  },
  {
    "ph": "f", "id": 1136, "pid": 870, "tid": 870, "ts": 7429770618022.658,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770618025.576, "dur": 0.139,
    "args": {
      "External id": 3458,      "cid": 370, "correlation": 1137
    }
  },
  {
    "ph": "f", "id": 1137, "pid": 870, "tid": 870, "ts": 7429770618025.576,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770618028.025, "dur": 0.141,
    "args": {
      "External id": 3458,      "cid": 370, "correlation": 1138
    }
  },
  {
    "ph": "f", "id": 1138, "pid": 870, "tid": 870, "ts": 7429770618028.025,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770618036.247, "dur": 5.696,
    "args": {
      "External id": 3458,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1139,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1139, "pid": 870, "tid": 870, "ts": 7429770618036.247,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770618042.460, "dur": 4.232,
    "args": {
      "External id": 3458,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1140,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1140, "pid": 870, "tid": 870, "ts": 7429770618042.460,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770618055.820, "dur": 4.010,
    "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": 1141,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1141, "pid": 870, "tid": 870, "ts": 7429770618055.820,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770618109.661, "dur": 4.250,
    "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": 1142,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1142, "pid": 870, "tid": 870, "ts": 7429770618109.661,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770618123.550, "dur": 6.150,
    "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": 1143,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1143, "pid": 870, "tid": 870, "ts": 7429770618123.550,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770618139.925, "dur": 3.680,
    "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": 1144,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1144, "pid": 870, "tid": 870, "ts": 7429770618139.925,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770618148.562, "dur": 3.798,
    "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": 1145,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1145, "pid": 870, "tid": 870, "ts": 7429770618148.562,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770618158.583, "dur": 5.774,
    "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": 1146,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1146, "pid": 870, "tid": 870, "ts": 7429770618158.583,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770618172.149, "dur": 5.540,
    "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": 1147,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1147, "pid": 870, "tid": 870, "ts": 7429770618172.149,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770618223.434, "dur": 0.319,
    "args": {
      "External id": 3480,      "cid": 290, "correlation": 1148
    }
  },
  {
    "ph": "f", "id": 1148, "pid": 870, "tid": 870, "ts": 7429770618223.434,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770618224.411, "dur": 60.967,
    "args": {
      "External id": 3480,            "bytes": 79691776, "cid": 108, "correlation": 1149, "ptr": "0x100000001"
    }
  },
  {
    "ph": "s", "id": 1149, "pid": 870, "tid": 870, "ts": 7429770618224.411,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770618287.568, "dur": 0.256,
    "args": {
      "External id": 3479,      "cid": 370, "correlation": 1150
    }
  },
  {
    "ph": "f", "id": 1150, "pid": 870, "tid": 870, "ts": 7429770618287.568,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770618292.433, "dur": 0.160,
    "args": {
      "External id": 3479,      "cid": 370, "correlation": 1151
    }
  },
  {
    "ph": "f", "id": 1151, "pid": 870, "tid": 870, "ts": 7429770618292.433,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770618295.226, "dur": 0.350,
    "args": {
      "External id": 3479,      "cid": 370, "correlation": 1152
    }
  },
  {
    "ph": "f", "id": 1152, "pid": 870, "tid": 870, "ts": 7429770618295.226,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770618299.446, "dur": 5.756,
    "args": {
      "External id": 3479,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1153,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1153, "pid": 870, "tid": 870, "ts": 7429770618299.446,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770618351.134, "dur": 0.159,
    "args": {
      "External id": 3492,      "cid": 370, "correlation": 1154
    }
  },
  {
    "ph": "f", "id": 1154, "pid": 870, "tid": 870, "ts": 7429770618351.134,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770618354.176, "dur": 0.152,
    "args": {
      "External id": 3492,      "cid": 370, "correlation": 1155
    }
  },
  {
    "ph": "f", "id": 1155, "pid": 870, "tid": 870, "ts": 7429770618354.176,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770618356.930, "dur": 0.148,
    "args": {
      "External id": 3492,      "cid": 370, "correlation": 1156
    }
  },
  {
    "ph": "f", "id": 1156, "pid": 870, "tid": 870, "ts": 7429770618356.930,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770618358.724, "dur": 5.545,
    "args": {
      "External id": 3492,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1157,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1157, "pid": 870, "tid": 870, "ts": 7429770618358.724,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770618399.798, "dur": 0.169,
    "args": {
      "External id": 3505,      "cid": 370, "correlation": 1158
    }
  },
  {
    "ph": "f", "id": 1158, "pid": 870, "tid": 870, "ts": 7429770618399.798,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770618402.924, "dur": 0.139,
    "args": {
      "External id": 3505,      "cid": 370, "correlation": 1159
    }
  },
  {
    "ph": "f", "id": 1159, "pid": 870, "tid": 870, "ts": 7429770618402.924,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770618404.918, "dur": 0.151,
    "args": {
      "External id": 3505,      "cid": 370, "correlation": 1160
    }
  },
  {
    "ph": "f", "id": 1160, "pid": 870, "tid": 870, "ts": 7429770618404.918,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770618406.430, "dur": 5.826,
    "args": {
      "External id": 3505,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1161,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1161, "pid": 870, "tid": 870, "ts": 7429770618406.430,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770618437.030, "dur": 4.280,
    "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": 1162,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1162, "pid": 870, "tid": 870, "ts": 7429770618437.030,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770618466.339, "dur": 5.354,
    "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": 1163,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1163, "pid": 870, "tid": 870, "ts": 7429770618466.339,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770618487.212, "dur": 7.313,
    "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": 1164,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1164, "pid": 870, "tid": 870, "ts": 7429770618487.212,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770618502.587, "dur": 3.953,
    "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": 1165,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1165, "pid": 870, "tid": 870, "ts": 7429770618502.587,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770618512.320, "dur": 4.143,
    "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": 1166,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1166, "pid": 870, "tid": 870, "ts": 7429770618512.320,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770618523.394, "dur": 4.022,
    "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": 1167,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1167, "pid": 870, "tid": 870, "ts": 7429770618523.394,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770618540.089, "dur": 6.196,
    "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": 1168,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1168, "pid": 870, "tid": 870, "ts": 7429770618540.089,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770618552.852, "dur": 6.494,
    "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": 1169,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1169, "pid": 870, "tid": 870, "ts": 7429770618552.852,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770618565.613, "dur": 4.109,
    "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": 1170,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1170, "pid": 870, "tid": 870, "ts": 7429770618565.613,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770618574.182, "dur": 3.163,
    "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": 1171,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1171, "pid": 870, "tid": 870, "ts": 7429770618574.182,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770618638.661, "dur": 4.486,
    "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": 1172,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1172, "pid": 870, "tid": 870, "ts": 7429770618638.661,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770618674.324, "dur": 3.943,
    "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": 1173,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1173, "pid": 870, "tid": 870, "ts": 7429770618674.324,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770618703.067, "dur": 0.272,
    "args": {
      "External id": 3580,      "cid": 290, "correlation": 1174
    }
  },
  {
    "ph": "f", "id": 1174, "pid": 870, "tid": 870, "ts": 7429770618703.067,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770618703.866, "dur": 43.705,
    "args": {
      "External id": 3580,            "bytes": 2097152, "cid": 108, "correlation": 1175, "ptr": "0x7ffd80420410"
    }
  },
  {
    "ph": "s", "id": 1175, "pid": 870, "tid": 870, "ts": 7429770618703.866,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770618763.259, "dur": 0.337,
    "args": {
      "External id": 3580,      "cid": 370, "correlation": 1176
    }
  },
  {
    "ph": "f", "id": 1176, "pid": 870, "tid": 870, "ts": 7429770618763.259,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770618767.812, "dur": 0.182,
    "args": {
      "External id": 3580,      "cid": 370, "correlation": 1177
    }
  },
  {
    "ph": "f", "id": 1177, "pid": 870, "tid": 870, "ts": 7429770618767.812,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770618771.102, "dur": 0.204,
    "args": {
      "External id": 3580,      "cid": 370, "correlation": 1178
    }
  },
  {
    "ph": "f", "id": 1178, "pid": 870, "tid": 870, "ts": 7429770618771.102,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770618774.306, "dur": 5.377,
    "args": {
      "External id": 3580,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1179,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1179, "pid": 870, "tid": 870, "ts": 7429770618774.306,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770618793.943, "dur": 4.966,
    "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": 1180,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1180, "pid": 870, "tid": 870, "ts": 7429770618793.943,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770618848.103, "dur": 3.898,
    "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": 1181,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1181, "pid": 870, "tid": 870, "ts": 7429770618848.103,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770618864.731, "dur": 5.825,
    "args": {
      "External id": 3593,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 1182,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1182, "pid": 870, "tid": 870, "ts": 7429770618864.731,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770618914.898, "dur": 0.186,
    "args": {
      "External id": 3605,      "cid": 370, "correlation": 1183
    }
  },
  {
    "ph": "f", "id": 1183, "pid": 870, "tid": 870, "ts": 7429770618914.898,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770618918.101, "dur": 0.144,
    "args": {
      "External id": 3605,      "cid": 370, "correlation": 1184
    }
  },
  {
    "ph": "f", "id": 1184, "pid": 870, "tid": 870, "ts": 7429770618918.101,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770618920.835, "dur": 0.209,
    "args": {
      "External id": 3605,      "cid": 370, "correlation": 1185
    }
  },
  {
    "ph": "f", "id": 1185, "pid": 870, "tid": 870, "ts": 7429770618920.835,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770618923.251, "dur": 5.369,
    "args": {
      "External id": 3605,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1186,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1186, "pid": 870, "tid": 870, "ts": 7429770618923.251,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770618945.826, "dur": 4.001,
    "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": 1187,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1187, "pid": 870, "tid": 870, "ts": 7429770618945.826,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770618994.381, "dur": 0.227,
    "args": {
      "External id": 3624,      "cid": 370, "correlation": 1188
    }
  },
  {
    "ph": "f", "id": 1188, "pid": 870, "tid": 870, "ts": 7429770618994.381,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770618997.997, "dur": 0.152,
    "args": {
      "External id": 3624,      "cid": 370, "correlation": 1189
    }
  },
  {
    "ph": "f", "id": 1189, "pid": 870, "tid": 870, "ts": 7429770618997.997,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770619000.643, "dur": 0.360,
    "args": {
      "External id": 3624,      "cid": 370, "correlation": 1190
    }
  },
  {
    "ph": "f", "id": 1190, "pid": 870, "tid": 870, "ts": 7429770619000.643,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619002.719, "dur": 5.224,
    "args": {
      "External id": 3624,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1191,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1191, "pid": 870, "tid": 870, "ts": 7429770619002.719,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619020.749, "dur": 4.660,
    "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": 1192,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1192, "pid": 870, "tid": 870, "ts": 7429770619020.749,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619045.194, "dur": 4.622,
    "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": 1193,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1193, "pid": 870, "tid": 870, "ts": 7429770619045.194,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619060.566, "dur": 5.054,
    "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": 1194,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1194, "pid": 870, "tid": 870, "ts": 7429770619060.566,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619073.444, "dur": 3.908,
    "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": 1195,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1195, "pid": 870, "tid": 870, "ts": 7429770619073.444,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619085.287, "dur": 4.997,
    "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": 1196,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1196, "pid": 870, "tid": 870, "ts": 7429770619085.287,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619096.760, "dur": 5.906,
    "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": 1197,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1197, "pid": 870, "tid": 870, "ts": 7429770619096.760,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619110.686, "dur": 4.234,
    "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": 1198,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1198, "pid": 870, "tid": 870, "ts": 7429770619110.686,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770619161.102, "dur": 0.200,
    "args": {
      "External id": 3645,      "cid": 370, "correlation": 1199
    }
  },
  {
    "ph": "f", "id": 1199, "pid": 870, "tid": 870, "ts": 7429770619161.102,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770619164.033, "dur": 0.160,
    "args": {
      "External id": 3645,      "cid": 370, "correlation": 1200
    }
  },
  {
    "ph": "f", "id": 1200, "pid": 870, "tid": 870, "ts": 7429770619164.033,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770619166.502, "dur": 0.138,
    "args": {
      "External id": 3645,      "cid": 370, "correlation": 1201
    }
  },
  {
    "ph": "f", "id": 1201, "pid": 870, "tid": 870, "ts": 7429770619166.502,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619168.943, "dur": 5.646,
    "args": {
      "External id": 3645,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1202,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1202, "pid": 870, "tid": 870, "ts": 7429770619168.943,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619191.924, "dur": 5.406,
    "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": 1203,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1203, "pid": 870, "tid": 870, "ts": 7429770619191.924,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770619219.590, "dur": 0.159,
    "args": {
      "External id": 3656,      "cid": 370, "correlation": 1204
    }
  },
  {
    "ph": "f", "id": 1204, "pid": 870, "tid": 870, "ts": 7429770619219.590,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770619221.885, "dur": 0.153,
    "args": {
      "External id": 3656,      "cid": 370, "correlation": 1205
    }
  },
  {
    "ph": "f", "id": 1205, "pid": 870, "tid": 870, "ts": 7429770619221.885,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770619224.067, "dur": 0.142,
    "args": {
      "External id": 3656,      "cid": 370, "correlation": 1206
    }
  },
  {
    "ph": "f", "id": 1206, "pid": 870, "tid": 870, "ts": 7429770619224.067,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619225.912, "dur": 4.458,
    "args": {
      "External id": 3656,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1207,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1207, "pid": 870, "tid": 870, "ts": 7429770619225.912,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619238.426, "dur": 3.954,
    "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": 1208,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1208, "pid": 870, "tid": 870, "ts": 7429770619238.426,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770619264.704, "dur": 0.162,
    "args": {
      "External id": 3667,      "cid": 370, "correlation": 1209
    }
  },
  {
    "ph": "f", "id": 1209, "pid": 870, "tid": 870, "ts": 7429770619264.704,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770619268.228, "dur": 0.138,
    "args": {
      "External id": 3667,      "cid": 370, "correlation": 1210
    }
  },
  {
    "ph": "f", "id": 1210, "pid": 870, "tid": 870, "ts": 7429770619268.228,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770619270.445, "dur": 0.137,
    "args": {
      "External id": 3667,      "cid": 370, "correlation": 1211
    }
  },
  {
    "ph": "f", "id": 1211, "pid": 870, "tid": 870, "ts": 7429770619270.445,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619279.637, "dur": 5.828,
    "args": {
      "External id": 3667,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1212,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1212, "pid": 870, "tid": 870, "ts": 7429770619279.637,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619285.799, "dur": 4.719,
    "args": {
      "External id": 3667,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1213,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1213, "pid": 870, "tid": 870, "ts": 7429770619285.799,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619300.635, "dur": 4.136,
    "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": 1214,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1214, "pid": 870, "tid": 870, "ts": 7429770619300.635,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619330.348, "dur": 3.718,
    "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": 1215,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1215, "pid": 870, "tid": 870, "ts": 7429770619330.348,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619342.951, "dur": 4.915,
    "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": 1216,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1216, "pid": 870, "tid": 870, "ts": 7429770619342.951,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619355.130, "dur": 4.287,
    "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": 1217,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1217, "pid": 870, "tid": 870, "ts": 7429770619355.130,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619365.510, "dur": 3.935,
    "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": 1218,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1218, "pid": 870, "tid": 870, "ts": 7429770619365.510,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619375.112, "dur": 4.354,
    "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": 1219,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1219, "pid": 870, "tid": 870, "ts": 7429770619375.112,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619388.402, "dur": 3.744,
    "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": 1220,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1220, "pid": 870, "tid": 870, "ts": 7429770619388.402,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770619432.010, "dur": 0.290,
    "args": {
      "External id": 3688,      "cid": 370, "correlation": 1221
    }
  },
  {
    "ph": "f", "id": 1221, "pid": 870, "tid": 870, "ts": 7429770619432.010,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770619435.609, "dur": 0.141,
    "args": {
      "External id": 3688,      "cid": 370, "correlation": 1222
    }
  },
  {
    "ph": "f", "id": 1222, "pid": 870, "tid": 870, "ts": 7429770619435.609,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770619437.746, "dur": 0.150,
    "args": {
      "External id": 3688,      "cid": 370, "correlation": 1223
    }
  },
  {
    "ph": "f", "id": 1223, "pid": 870, "tid": 870, "ts": 7429770619437.746,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619439.830, "dur": 5.184,
    "args": {
      "External id": 3688,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1224,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1224, "pid": 870, "tid": 870, "ts": 7429770619439.830,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770619480.767, "dur": 0.184,
    "args": {
      "External id": 3701,      "cid": 370, "correlation": 1225
    }
  },
  {
    "ph": "f", "id": 1225, "pid": 870, "tid": 870, "ts": 7429770619480.767,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770619484.028, "dur": 0.162,
    "args": {
      "External id": 3701,      "cid": 370, "correlation": 1226
    }
  },
  {
    "ph": "f", "id": 1226, "pid": 870, "tid": 870, "ts": 7429770619484.028,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770619486.456, "dur": 0.153,
    "args": {
      "External id": 3701,      "cid": 370, "correlation": 1227
    }
  },
  {
    "ph": "f", "id": 1227, "pid": 870, "tid": 870, "ts": 7429770619486.456,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619488.253, "dur": 4.977,
    "args": {
      "External id": 3701,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1228,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1228, "pid": 870, "tid": 870, "ts": 7429770619488.253,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770619527.279, "dur": 0.189,
    "args": {
      "External id": 3714,      "cid": 370, "correlation": 1229
    }
  },
  {
    "ph": "f", "id": 1229, "pid": 870, "tid": 870, "ts": 7429770619527.279,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770619530.349, "dur": 0.159,
    "args": {
      "External id": 3714,      "cid": 370, "correlation": 1230
    }
  },
  {
    "ph": "f", "id": 1230, "pid": 870, "tid": 870, "ts": 7429770619530.349,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770619532.347, "dur": 0.152,
    "args": {
      "External id": 3714,      "cid": 370, "correlation": 1231
    }
  },
  {
    "ph": "f", "id": 1231, "pid": 870, "tid": 870, "ts": 7429770619532.347,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619533.741, "dur": 5.383,
    "args": {
      "External id": 3714,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1232,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1232, "pid": 870, "tid": 870, "ts": 7429770619533.741,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619561.768, "dur": 6.478,
    "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": 1233,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1233, "pid": 870, "tid": 870, "ts": 7429770619561.768,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619589.610, "dur": 5.458,
    "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": 1234,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1234, "pid": 870, "tid": 870, "ts": 7429770619589.610,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619609.849, "dur": 6.294,
    "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": 1235,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1235, "pid": 870, "tid": 870, "ts": 7429770619609.849,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619623.877, "dur": 3.887,
    "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": 1236,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1236, "pid": 870, "tid": 870, "ts": 7429770619623.877,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619634.031, "dur": 4.439,
    "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": 1237,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1237, "pid": 870, "tid": 870, "ts": 7429770619634.031,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619644.258, "dur": 4.997,
    "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": 1238,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1238, "pid": 870, "tid": 870, "ts": 7429770619644.258,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619663.922, "dur": 4.423,
    "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": 1239,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1239, "pid": 870, "tid": 870, "ts": 7429770619663.922,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619675.080, "dur": 6.799,
    "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": 1240,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1240, "pid": 870, "tid": 870, "ts": 7429770619675.080,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619689.065, "dur": 4.363,
    "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": 1241,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1241, "pid": 870, "tid": 870, "ts": 7429770619689.065,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619697.990, "dur": 3.242,
    "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": 1242,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1242, "pid": 870, "tid": 870, "ts": 7429770619697.990,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619759.036, "dur": 4.333,
    "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": 1243,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1243, "pid": 870, "tid": 870, "ts": 7429770619759.036,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619791.345, "dur": 3.578,
    "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": 1244,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1244, "pid": 870, "tid": 870, "ts": 7429770619791.345,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770619818.940, "dur": 0.320,
    "args": {
      "External id": 3789,      "cid": 290, "correlation": 1245
    }
  },
  {
    "ph": "f", "id": 1245, "pid": 870, "tid": 870, "ts": 7429770619818.940,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770619819.952, "dur": 61.173,
    "args": {
      "External id": 3789,            "bytes": 2097152, "cid": 108, "correlation": 1246, "ptr": "0x7ffd80420410"
    }
  },
  {
    "ph": "s", "id": 1246, "pid": 870, "tid": 870, "ts": 7429770619819.952,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770619895.931, "dur": 0.288,
    "args": {
      "External id": 3789,      "cid": 370, "correlation": 1247
    }
  },
  {
    "ph": "f", "id": 1247, "pid": 870, "tid": 870, "ts": 7429770619895.931,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770619900.937, "dur": 0.144,
    "args": {
      "External id": 3789,      "cid": 370, "correlation": 1248
    }
  },
  {
    "ph": "f", "id": 1248, "pid": 870, "tid": 870, "ts": 7429770619900.937,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770619904.143, "dur": 0.194,
    "args": {
      "External id": 3789,      "cid": 370, "correlation": 1249
    }
  },
  {
    "ph": "f", "id": 1249, "pid": 870, "tid": 870, "ts": 7429770619904.143,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619907.678, "dur": 6.552,
    "args": {
      "External id": 3789,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1250,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1250, "pid": 870, "tid": 870, "ts": 7429770619907.678,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619929.424, "dur": 4.443,
    "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": 1251,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1251, "pid": 870, "tid": 870, "ts": 7429770619929.424,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619950.212, "dur": 5.943,
    "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": 1252,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1252, "pid": 870, "tid": 870, "ts": 7429770619950.212,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770619968.579, "dur": 6.207,
    "args": {
      "External id": 3802,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 1253,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1253, "pid": 870, "tid": 870, "ts": 7429770619968.579,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770620019.312, "dur": 0.225,
    "args": {
      "External id": 3814,      "cid": 370, "correlation": 1254
    }
  },
  {
    "ph": "f", "id": 1254, "pid": 870, "tid": 870, "ts": 7429770620019.312,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770620023.051, "dur": 0.141,
    "args": {
      "External id": 3814,      "cid": 370, "correlation": 1255
    }
  },
  {
    "ph": "f", "id": 1255, "pid": 870, "tid": 870, "ts": 7429770620023.051,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770620025.955, "dur": 0.160,
    "args": {
      "External id": 3814,      "cid": 370, "correlation": 1256
    }
  },
  {
    "ph": "f", "id": 1256, "pid": 870, "tid": 870, "ts": 7429770620025.955,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620028.326, "dur": 6.225,
    "args": {
      "External id": 3814,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1257,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1257, "pid": 870, "tid": 870, "ts": 7429770620028.326,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620052.306, "dur": 4.134,
    "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": 1258,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1258, "pid": 870, "tid": 870, "ts": 7429770620052.306,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770620103.026, "dur": 0.222,
    "args": {
      "External id": 3833,      "cid": 370, "correlation": 1259
    }
  },
  {
    "ph": "f", "id": 1259, "pid": 870, "tid": 870, "ts": 7429770620103.026,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770620106.485, "dur": 0.206,
    "args": {
      "External id": 3833,      "cid": 370, "correlation": 1260
    }
  },
  {
    "ph": "f", "id": 1260, "pid": 870, "tid": 870, "ts": 7429770620106.485,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770620108.613, "dur": 0.361,
    "args": {
      "External id": 3833,      "cid": 370, "correlation": 1261
    }
  },
  {
    "ph": "f", "id": 1261, "pid": 870, "tid": 870, "ts": 7429770620108.613,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620110.822, "dur": 5.053,
    "args": {
      "External id": 3833,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1262,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1262, "pid": 870, "tid": 870, "ts": 7429770620110.822,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620128.607, "dur": 4.464,
    "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": 1263,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1263, "pid": 870, "tid": 870, "ts": 7429770620128.607,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620152.563, "dur": 4.015,
    "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": 1264,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1264, "pid": 870, "tid": 870, "ts": 7429770620152.563,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620166.334, "dur": 7.180,
    "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": 1265,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1265, "pid": 870, "tid": 870, "ts": 7429770620166.334,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620181.630, "dur": 4.358,
    "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": 1266,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1266, "pid": 870, "tid": 870, "ts": 7429770620181.630,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620191.990, "dur": 3.847,
    "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": 1267,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1267, "pid": 870, "tid": 870, "ts": 7429770620191.990,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620203.469, "dur": 4.154,
    "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": 1268,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1268, "pid": 870, "tid": 870, "ts": 7429770620203.469,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620215.232, "dur": 6.365,
    "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": 1269,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1269, "pid": 870, "tid": 870, "ts": 7429770620215.232,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770620257.426, "dur": 0.224,
    "args": {
      "External id": 3854,      "cid": 370, "correlation": 1270
    }
  },
  {
    "ph": "f", "id": 1270, "pid": 870, "tid": 870, "ts": 7429770620257.426,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770620260.835, "dur": 0.179,
    "args": {
      "External id": 3854,      "cid": 370, "correlation": 1271
    }
  },
  {
    "ph": "f", "id": 1271, "pid": 870, "tid": 870, "ts": 7429770620260.835,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770620263.576, "dur": 0.141,
    "args": {
      "External id": 3854,      "cid": 370, "correlation": 1272
    }
  },
  {
    "ph": "f", "id": 1272, "pid": 870, "tid": 870, "ts": 7429770620263.576,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620265.629, "dur": 6.497,
    "args": {
      "External id": 3854,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1273,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1273, "pid": 870, "tid": 870, "ts": 7429770620265.629,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620288.541, "dur": 5.351,
    "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": 1274,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1274, "pid": 870, "tid": 870, "ts": 7429770620288.541,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770620320.184, "dur": 0.183,
    "args": {
      "External id": 3865,      "cid": 370, "correlation": 1275
    }
  },
  {
    "ph": "f", "id": 1275, "pid": 870, "tid": 870, "ts": 7429770620320.184,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770620323.210, "dur": 0.144,
    "args": {
      "External id": 3865,      "cid": 370, "correlation": 1276
    }
  },
  {
    "ph": "f", "id": 1276, "pid": 870, "tid": 870, "ts": 7429770620323.210,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770620325.069, "dur": 0.139,
    "args": {
      "External id": 3865,      "cid": 370, "correlation": 1277
    }
  },
  {
    "ph": "f", "id": 1277, "pid": 870, "tid": 870, "ts": 7429770620325.069,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620326.784, "dur": 7.450,
    "args": {
      "External id": 3865,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1278,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1278, "pid": 870, "tid": 870, "ts": 7429770620326.784,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620343.115, "dur": 4.200,
    "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": 1279,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1279, "pid": 870, "tid": 870, "ts": 7429770620343.115,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770620370.669, "dur": 0.176,
    "args": {
      "External id": 3876,      "cid": 370, "correlation": 1280
    }
  },
  {
    "ph": "f", "id": 1280, "pid": 870, "tid": 870, "ts": 7429770620370.669,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770620374.396, "dur": 0.139,
    "args": {
      "External id": 3876,      "cid": 370, "correlation": 1281
    }
  },
  {
    "ph": "f", "id": 1281, "pid": 870, "tid": 870, "ts": 7429770620374.396,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770620377.069, "dur": 0.138,
    "args": {
      "External id": 3876,      "cid": 370, "correlation": 1282
    }
  },
  {
    "ph": "f", "id": 1282, "pid": 870, "tid": 870, "ts": 7429770620377.069,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620385.276, "dur": 6.691,
    "args": {
      "External id": 3876,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1283,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1283, "pid": 870, "tid": 870, "ts": 7429770620385.276,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620392.507, "dur": 5.022,
    "args": {
      "External id": 3876,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1284,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1284, "pid": 870, "tid": 870, "ts": 7429770620392.507,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620406.284, "dur": 4.422,
    "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": 1285,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1285, "pid": 870, "tid": 870, "ts": 7429770620406.284,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620435.159, "dur": 3.286,
    "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": 1286,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1286, "pid": 870, "tid": 870, "ts": 7429770620435.159,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620446.125, "dur": 5.759,
    "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": 1287,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1287, "pid": 870, "tid": 870, "ts": 7429770620446.125,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620459.338, "dur": 3.919,
    "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": 1288,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1288, "pid": 870, "tid": 870, "ts": 7429770620459.338,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620467.599, "dur": 3.744,
    "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": 1289,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1289, "pid": 870, "tid": 870, "ts": 7429770620467.599,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620477.757, "dur": 4.581,
    "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": 1290,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1290, "pid": 870, "tid": 870, "ts": 7429770620477.757,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620488.690, "dur": 5.228,
    "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": 1291,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1291, "pid": 870, "tid": 870, "ts": 7429770620488.690,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770620529.428, "dur": 0.204,
    "args": {
      "External id": 3897,      "cid": 370, "correlation": 1292
    }
  },
  {
    "ph": "f", "id": 1292, "pid": 870, "tid": 870, "ts": 7429770620529.428,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770620532.835, "dur": 0.153,
    "args": {
      "External id": 3897,      "cid": 370, "correlation": 1293
    }
  },
  {
    "ph": "f", "id": 1293, "pid": 870, "tid": 870, "ts": 7429770620532.835,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770620535.014, "dur": 0.147,
    "args": {
      "External id": 3897,      "cid": 370, "correlation": 1294
    }
  },
  {
    "ph": "f", "id": 1294, "pid": 870, "tid": 870, "ts": 7429770620535.014,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620537.001, "dur": 4.793,
    "args": {
      "External id": 3897,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1295,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1295, "pid": 870, "tid": 870, "ts": 7429770620537.001,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770620576.698, "dur": 0.173,
    "args": {
      "External id": 3910,      "cid": 370, "correlation": 1296
    }
  },
  {
    "ph": "f", "id": 1296, "pid": 870, "tid": 870, "ts": 7429770620576.698,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770620579.788, "dur": 0.163,
    "args": {
      "External id": 3910,      "cid": 370, "correlation": 1297
    }
  },
  {
    "ph": "f", "id": 1297, "pid": 870, "tid": 870, "ts": 7429770620579.788,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770620582.091, "dur": 0.151,
    "args": {
      "External id": 3910,      "cid": 370, "correlation": 1298
    }
  },
  {
    "ph": "f", "id": 1298, "pid": 870, "tid": 870, "ts": 7429770620582.091,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620583.652, "dur": 6.059,
    "args": {
      "External id": 3910,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1299,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1299, "pid": 870, "tid": 870, "ts": 7429770620583.652,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770620623.535, "dur": 0.184,
    "args": {
      "External id": 3923,      "cid": 370, "correlation": 1300
    }
  },
  {
    "ph": "f", "id": 1300, "pid": 870, "tid": 870, "ts": 7429770620623.535,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770620626.599, "dur": 0.162,
    "args": {
      "External id": 3923,      "cid": 370, "correlation": 1301
    }
  },
  {
    "ph": "f", "id": 1301, "pid": 870, "tid": 870, "ts": 7429770620626.599,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770620628.473, "dur": 0.153,
    "args": {
      "External id": 3923,      "cid": 370, "correlation": 1302
    }
  },
  {
    "ph": "f", "id": 1302, "pid": 870, "tid": 870, "ts": 7429770620628.473,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620629.891, "dur": 5.017,
    "args": {
      "External id": 3923,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1303,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1303, "pid": 870, "tid": 870, "ts": 7429770620629.891,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620657.267, "dur": 4.161,
    "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": 1304,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1304, "pid": 870, "tid": 870, "ts": 7429770620657.267,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620679.724, "dur": 4.848,
    "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": 1305,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1305, "pid": 870, "tid": 870, "ts": 7429770620679.724,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620698.383, "dur": 7.268,
    "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": 1306,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1306, "pid": 870, "tid": 870, "ts": 7429770620698.383,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620712.360, "dur": 4.642,
    "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": 1307,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1307, "pid": 870, "tid": 870, "ts": 7429770620712.360,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620723.043, "dur": 3.530,
    "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": 1308,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1308, "pid": 870, "tid": 870, "ts": 7429770620723.043,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620731.924, "dur": 3.240,
    "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": 1309,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1309, "pid": 870, "tid": 870, "ts": 7429770620731.924,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620748.599, "dur": 3.563,
    "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": 1310,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1310, "pid": 870, "tid": 870, "ts": 7429770620748.599,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620759.388, "dur": 6.108,
    "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": 1311,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1311, "pid": 870, "tid": 870, "ts": 7429770620759.388,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620772.007, "dur": 4.096,
    "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": 1312,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1312, "pid": 870, "tid": 870, "ts": 7429770620772.007,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620780.788, "dur": 3.171,
    "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": 1313,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1313, "pid": 870, "tid": 870, "ts": 7429770620780.788,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620839.126, "dur": 4.392,
    "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": 1314,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1314, "pid": 870, "tid": 870, "ts": 7429770620839.126,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620872.171, "dur": 4.536,
    "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": 1315,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1315, "pid": 870, "tid": 870, "ts": 7429770620872.171,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770620901.230, "dur": 0.315,
    "args": {
      "External id": 3998,      "cid": 290, "correlation": 1316
    }
  },
  {
    "ph": "f", "id": 1316, "pid": 870, "tid": 870, "ts": 7429770620901.230,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770620902.114, "dur": 54.824,
    "args": {
      "External id": 3998,            "bytes": 2097152, "cid": 108, "correlation": 1317, "ptr": "0x7ffd80420410"
    }
  },
  {
    "ph": "s", "id": 1317, "pid": 870, "tid": 870, "ts": 7429770620902.114,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770620972.654, "dur": 0.277,
    "args": {
      "External id": 3998,      "cid": 370, "correlation": 1318
    }
  },
  {
    "ph": "f", "id": 1318, "pid": 870, "tid": 870, "ts": 7429770620972.654,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770620977.343, "dur": 0.150,
    "args": {
      "External id": 3998,      "cid": 370, "correlation": 1319
    }
  },
  {
    "ph": "f", "id": 1319, "pid": 870, "tid": 870, "ts": 7429770620977.343,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770620980.371, "dur": 0.152,
    "args": {
      "External id": 3998,      "cid": 370, "correlation": 1320
    }
  },
  {
    "ph": "f", "id": 1320, "pid": 870, "tid": 870, "ts": 7429770620980.371,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770620983.746, "dur": 4.845,
    "args": {
      "External id": 3998,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1321,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1321, "pid": 870, "tid": 870, "ts": 7429770620983.746,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621002.632, "dur": 4.698,
    "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": 1322,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1322, "pid": 870, "tid": 870, "ts": 7429770621002.632,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621023.300, "dur": 7.290,
    "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": 1323,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1323, "pid": 870, "tid": 870, "ts": 7429770621023.300,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621042.501, "dur": 5.459,
    "args": {
      "External id": 4011,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 1324,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1324, "pid": 870, "tid": 870, "ts": 7429770621042.501,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770621093.121, "dur": 0.200,
    "args": {
      "External id": 4023,      "cid": 370, "correlation": 1325
    }
  },
  {
    "ph": "f", "id": 1325, "pid": 870, "tid": 870, "ts": 7429770621093.121,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770621096.725, "dur": 0.142,
    "args": {
      "External id": 4023,      "cid": 370, "correlation": 1326
    }
  },
  {
    "ph": "f", "id": 1326, "pid": 870, "tid": 870, "ts": 7429770621096.725,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770621099.421, "dur": 0.157,
    "args": {
      "External id": 4023,      "cid": 370, "correlation": 1327
    }
  },
  {
    "ph": "f", "id": 1327, "pid": 870, "tid": 870, "ts": 7429770621099.421,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621101.737, "dur": 5.975,
    "args": {
      "External id": 4023,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1328,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1328, "pid": 870, "tid": 870, "ts": 7429770621101.737,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621125.062, "dur": 4.162,
    "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": 1329,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1329, "pid": 870, "tid": 870, "ts": 7429770621125.062,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770621173.982, "dur": 0.190,
    "args": {
      "External id": 4042,      "cid": 370, "correlation": 1330
    }
  },
  {
    "ph": "f", "id": 1330, "pid": 870, "tid": 870, "ts": 7429770621173.982,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770621177.681, "dur": 0.152,
    "args": {
      "External id": 4042,      "cid": 370, "correlation": 1331
    }
  },
  {
    "ph": "f", "id": 1331, "pid": 870, "tid": 870, "ts": 7429770621177.681,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770621180.293, "dur": 0.339,
    "args": {
      "External id": 4042,      "cid": 370, "correlation": 1332
    }
  },
  {
    "ph": "f", "id": 1332, "pid": 870, "tid": 870, "ts": 7429770621180.293,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621182.403, "dur": 5.430,
    "args": {
      "External id": 4042,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1333,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1333, "pid": 870, "tid": 870, "ts": 7429770621182.403,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621200.652, "dur": 4.449,
    "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": 1334,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1334, "pid": 870, "tid": 870, "ts": 7429770621200.652,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621222.054, "dur": 5.879,
    "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": 1335,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1335, "pid": 870, "tid": 870, "ts": 7429770621222.054,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621238.191, "dur": 5.209,
    "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": 1336,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1336, "pid": 870, "tid": 870, "ts": 7429770621238.191,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621251.100, "dur": 3.771,
    "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": 1337,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1337, "pid": 870, "tid": 870, "ts": 7429770621251.100,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621262.981, "dur": 4.632,
    "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": 1338,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1338, "pid": 870, "tid": 870, "ts": 7429770621262.981,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621274.224, "dur": 4.329,
    "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": 1339,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1339, "pid": 870, "tid": 870, "ts": 7429770621274.224,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621286.806, "dur": 3.813,
    "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": 1340,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1340, "pid": 870, "tid": 870, "ts": 7429770621286.806,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770621328.327, "dur": 0.309,
    "args": {
      "External id": 4064,      "cid": 290, "correlation": 1341
    }
  },
  {
    "ph": "f", "id": 1341, "pid": 870, "tid": 870, "ts": 7429770621328.327,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770621329.227, "dur": 60.515,
    "args": {
      "External id": 4064,            "bytes": 79691776, "cid": 108, "correlation": 1342, "ptr": "0x7ffd8041f0a8"
    }
  },
  {
    "ph": "s", "id": 1342, "pid": 870, "tid": 870, "ts": 7429770621329.227,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770621392.467, "dur": 0.224,
    "args": {
      "External id": 4063,      "cid": 370, "correlation": 1343
    }
  },
  {
    "ph": "f", "id": 1343, "pid": 870, "tid": 870, "ts": 7429770621392.467,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770621397.033, "dur": 0.139,
    "args": {
      "External id": 4063,      "cid": 370, "correlation": 1344
    }
  },
  {
    "ph": "f", "id": 1344, "pid": 870, "tid": 870, "ts": 7429770621397.033,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770621400.200, "dur": 0.158,
    "args": {
      "External id": 4063,      "cid": 370, "correlation": 1345
    }
  },
  {
    "ph": "f", "id": 1345, "pid": 870, "tid": 870, "ts": 7429770621400.200,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621403.795, "dur": 7.495,
    "args": {
      "External id": 4063,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1346,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1346, "pid": 870, "tid": 870, "ts": 7429770621403.795,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621431.459, "dur": 4.989,
    "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": 1347,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1347, "pid": 870, "tid": 870, "ts": 7429770621431.459,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770621466.634, "dur": 0.163,
    "args": {
      "External id": 4074,      "cid": 370, "correlation": 1348
    }
  },
  {
    "ph": "f", "id": 1348, "pid": 870, "tid": 870, "ts": 7429770621466.634,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770621469.845, "dur": 0.160,
    "args": {
      "External id": 4074,      "cid": 370, "correlation": 1349
    }
  },
  {
    "ph": "f", "id": 1349, "pid": 870, "tid": 870, "ts": 7429770621469.845,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770621472.241, "dur": 0.137,
    "args": {
      "External id": 4074,      "cid": 370, "correlation": 1350
    }
  },
  {
    "ph": "f", "id": 1350, "pid": 870, "tid": 870, "ts": 7429770621472.241,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621473.943, "dur": 5.104,
    "args": {
      "External id": 4074,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1351,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1351, "pid": 870, "tid": 870, "ts": 7429770621473.943,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621487.995, "dur": 4.150,
    "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": 1352,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1352, "pid": 870, "tid": 870, "ts": 7429770621487.995,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770621514.123, "dur": 0.189,
    "args": {
      "External id": 4085,      "cid": 370, "correlation": 1353
    }
  },
  {
    "ph": "f", "id": 1353, "pid": 870, "tid": 870, "ts": 7429770621514.123,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770621518.043, "dur": 0.140,
    "args": {
      "External id": 4085,      "cid": 370, "correlation": 1354
    }
  },
  {
    "ph": "f", "id": 1354, "pid": 870, "tid": 870, "ts": 7429770621518.043,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770621520.708, "dur": 0.140,
    "args": {
      "External id": 4085,      "cid": 370, "correlation": 1355
    }
  },
  {
    "ph": "f", "id": 1355, "pid": 870, "tid": 870, "ts": 7429770621520.708,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621529.252, "dur": 5.531,
    "args": {
      "External id": 4085,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1356,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1356, "pid": 870, "tid": 870, "ts": 7429770621529.252,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621535.196, "dur": 4.797,
    "args": {
      "External id": 4085,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1357,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1357, "pid": 870, "tid": 870, "ts": 7429770621535.196,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621550.146, "dur": 5.118,
    "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": 1358,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1358, "pid": 870, "tid": 870, "ts": 7429770621550.146,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621581.925, "dur": 4.454,
    "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": 1359,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1359, "pid": 870, "tid": 870, "ts": 7429770621581.925,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621595.341, "dur": 5.516,
    "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": 1360,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1360, "pid": 870, "tid": 870, "ts": 7429770621595.341,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621609.418, "dur": 3.843,
    "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": 1361,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1361, "pid": 870, "tid": 870, "ts": 7429770621609.418,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621618.394, "dur": 4.287,
    "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": 1362,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1362, "pid": 870, "tid": 870, "ts": 7429770621618.394,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621629.433, "dur": 4.817,
    "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": 1363,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1363, "pid": 870, "tid": 870, "ts": 7429770621629.433,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621679.789, "dur": 5.180,
    "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": 1364,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1364, "pid": 870, "tid": 870, "ts": 7429770621679.789,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770621727.445, "dur": 0.299,
    "args": {
      "External id": 4106,      "cid": 370, "correlation": 1365
    }
  },
  {
    "ph": "f", "id": 1365, "pid": 870, "tid": 870, "ts": 7429770621727.445,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770621730.982, "dur": 0.140,
    "args": {
      "External id": 4106,      "cid": 370, "correlation": 1366
    }
  },
  {
    "ph": "f", "id": 1366, "pid": 870, "tid": 870, "ts": 7429770621730.982,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770621733.299, "dur": 0.151,
    "args": {
      "External id": 4106,      "cid": 370, "correlation": 1367
    }
  },
  {
    "ph": "f", "id": 1367, "pid": 870, "tid": 870, "ts": 7429770621733.299,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621735.411, "dur": 4.923,
    "args": {
      "External id": 4106,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1368,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1368, "pid": 870, "tid": 870, "ts": 7429770621735.411,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770621779.493, "dur": 0.197,
    "args": {
      "External id": 4119,      "cid": 370, "correlation": 1369
    }
  },
  {
    "ph": "f", "id": 1369, "pid": 870, "tid": 870, "ts": 7429770621779.493,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770621782.523, "dur": 0.139,
    "args": {
      "External id": 4119,      "cid": 370, "correlation": 1370
    }
  },
  {
    "ph": "f", "id": 1370, "pid": 870, "tid": 870, "ts": 7429770621782.523,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770621784.906, "dur": 0.151,
    "args": {
      "External id": 4119,      "cid": 370, "correlation": 1371
    }
  },
  {
    "ph": "f", "id": 1371, "pid": 870, "tid": 870, "ts": 7429770621784.906,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621786.645, "dur": 6.401,
    "args": {
      "External id": 4119,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1372,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1372, "pid": 870, "tid": 870, "ts": 7429770621786.645,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770621838.036, "dur": 0.201,
    "args": {
      "External id": 4132,      "cid": 370, "correlation": 1373
    }
  },
  {
    "ph": "f", "id": 1373, "pid": 870, "tid": 870, "ts": 7429770621838.036,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770621840.648, "dur": 0.142,
    "args": {
      "External id": 4132,      "cid": 370, "correlation": 1374
    }
  },
  {
    "ph": "f", "id": 1374, "pid": 870, "tid": 870, "ts": 7429770621840.648,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770621842.378, "dur": 0.151,
    "args": {
      "External id": 4132,      "cid": 370, "correlation": 1375
    }
  },
  {
    "ph": "f", "id": 1375, "pid": 870, "tid": 870, "ts": 7429770621842.378,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621843.910, "dur": 5.224,
    "args": {
      "External id": 4132,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1376,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1376, "pid": 870, "tid": 870, "ts": 7429770621843.910,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621873.055, "dur": 6.285,
    "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": 1377,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1377, "pid": 870, "tid": 870, "ts": 7429770621873.055,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621901.123, "dur": 4.635,
    "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": 1378,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1378, "pid": 870, "tid": 870, "ts": 7429770621901.123,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621920.028, "dur": 7.713,
    "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": 1379,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1379, "pid": 870, "tid": 870, "ts": 7429770621920.028,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621935.589, "dur": 3.818,
    "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": 1380,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1380, "pid": 870, "tid": 870, "ts": 7429770621935.589,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621944.767, "dur": 4.032,
    "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": 1381,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1381, "pid": 870, "tid": 870, "ts": 7429770621944.767,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621954.702, "dur": 4.239,
    "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": 1382,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1382, "pid": 870, "tid": 870, "ts": 7429770621954.702,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621972.836, "dur": 3.615,
    "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": 1383,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1383, "pid": 870, "tid": 870, "ts": 7429770621972.836,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621983.718, "dur": 6.150,
    "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": 1384,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1384, "pid": 870, "tid": 870, "ts": 7429770621983.718,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770621995.770, "dur": 5.126,
    "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": 1385,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1385, "pid": 870, "tid": 870, "ts": 7429770621995.770,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622005.450, "dur": 5.107,
    "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": 1386,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1386, "pid": 870, "tid": 870, "ts": 7429770622005.450,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622089.614, "dur": 5.641,
    "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": 1387,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1387, "pid": 870, "tid": 870, "ts": 7429770622089.614,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622131.612, "dur": 4.774,
    "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": 1388,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1388, "pid": 870, "tid": 870, "ts": 7429770622131.612,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770622163.651, "dur": 0.341,
    "args": {
      "External id": 4207,      "cid": 290, "correlation": 1389
    }
  },
  {
    "ph": "f", "id": 1389, "pid": 870, "tid": 870, "ts": 7429770622163.651,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770622164.603, "dur": 54.743,
    "args": {
      "External id": 4207,            "bytes": 2097152, "cid": 108, "correlation": 1390, "ptr": "0x7ffd80420410"
    }
  },
  {
    "ph": "s", "id": 1390, "pid": 870, "tid": 870, "ts": 7429770622164.603,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770622236.501, "dur": 0.297,
    "args": {
      "External id": 4207,      "cid": 370, "correlation": 1391
    }
  },
  {
    "ph": "f", "id": 1391, "pid": 870, "tid": 870, "ts": 7429770622236.501,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770622241.499, "dur": 0.190,
    "args": {
      "External id": 4207,      "cid": 370, "correlation": 1392
    }
  },
  {
    "ph": "f", "id": 1392, "pid": 870, "tid": 870, "ts": 7429770622241.499,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770622245.228, "dur": 0.206,
    "args": {
      "External id": 4207,      "cid": 370, "correlation": 1393
    }
  },
  {
    "ph": "f", "id": 1393, "pid": 870, "tid": 870, "ts": 7429770622245.228,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622249.637, "dur": 6.315,
    "args": {
      "External id": 4207,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1394,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1394, "pid": 870, "tid": 870, "ts": 7429770622249.637,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622272.407, "dur": 4.608,
    "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": 1395,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1395, "pid": 870, "tid": 870, "ts": 7429770622272.407,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622294.039, "dur": 5.071,
    "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": 1396,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1396, "pid": 870, "tid": 870, "ts": 7429770622294.039,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622320.015, "dur": 6.024,
    "args": {
      "External id": 4220,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 1397,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1397, "pid": 870, "tid": 870, "ts": 7429770622320.015,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770622376.646, "dur": 0.251,
    "args": {
      "External id": 4232,      "cid": 370, "correlation": 1398
    }
  },
  {
    "ph": "f", "id": 1398, "pid": 870, "tid": 870, "ts": 7429770622376.646,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770622380.494, "dur": 0.141,
    "args": {
      "External id": 4232,      "cid": 370, "correlation": 1399
    }
  },
  {
    "ph": "f", "id": 1399, "pid": 870, "tid": 870, "ts": 7429770622380.494,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770622383.651, "dur": 0.153,
    "args": {
      "External id": 4232,      "cid": 370, "correlation": 1400
    }
  },
  {
    "ph": "f", "id": 1400, "pid": 870, "tid": 870, "ts": 7429770622383.651,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622386.577, "dur": 6.500,
    "args": {
      "External id": 4232,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1401,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1401, "pid": 870, "tid": 870, "ts": 7429770622386.577,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622414.316, "dur": 4.003,
    "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": 1402,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1402, "pid": 870, "tid": 870, "ts": 7429770622414.316,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770622467.870, "dur": 0.246,
    "args": {
      "External id": 4251,      "cid": 370, "correlation": 1403
    }
  },
  {
    "ph": "f", "id": 1403, "pid": 870, "tid": 870, "ts": 7429770622467.870,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770622471.533, "dur": 0.186,
    "args": {
      "External id": 4251,      "cid": 370, "correlation": 1404
    }
  },
  {
    "ph": "f", "id": 1404, "pid": 870, "tid": 870, "ts": 7429770622471.533,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770622474.115, "dur": 0.374,
    "args": {
      "External id": 4251,      "cid": 370, "correlation": 1405
    }
  },
  {
    "ph": "f", "id": 1405, "pid": 870, "tid": 870, "ts": 7429770622474.115,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622476.414, "dur": 5.029,
    "args": {
      "External id": 4251,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1406,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1406, "pid": 870, "tid": 870, "ts": 7429770622476.414,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622495.687, "dur": 4.423,
    "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": 1407,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1407, "pid": 870, "tid": 870, "ts": 7429770622495.687,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622520.266, "dur": 5.735,
    "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": 1408,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1408, "pid": 870, "tid": 870, "ts": 7429770622520.266,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622537.230, "dur": 6.927,
    "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": 1409,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1409, "pid": 870, "tid": 870, "ts": 7429770622537.230,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622553.379, "dur": 4.702,
    "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": 1410,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1410, "pid": 870, "tid": 870, "ts": 7429770622553.379,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622565.113, "dur": 4.370,
    "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": 1411,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1411, "pid": 870, "tid": 870, "ts": 7429770622565.113,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622577.143, "dur": 5.615,
    "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": 1412,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1412, "pid": 870, "tid": 870, "ts": 7429770622577.143,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622592.263, "dur": 4.101,
    "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": 1413,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1413, "pid": 870, "tid": 870, "ts": 7429770622592.263,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770622633.472, "dur": 0.182,
    "args": {
      "External id": 4272,      "cid": 370, "correlation": 1414
    }
  },
  {
    "ph": "f", "id": 1414, "pid": 870, "tid": 870, "ts": 7429770622633.472,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770622637.292, "dur": 0.172,
    "args": {
      "External id": 4272,      "cid": 370, "correlation": 1415
    }
  },
  {
    "ph": "f", "id": 1415, "pid": 870, "tid": 870, "ts": 7429770622637.292,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770622640.198, "dur": 0.139,
    "args": {
      "External id": 4272,      "cid": 370, "correlation": 1416
    }
  },
  {
    "ph": "f", "id": 1416, "pid": 870, "tid": 870, "ts": 7429770622640.198,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622642.301, "dur": 6.759,
    "args": {
      "External id": 4272,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1417,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1417, "pid": 870, "tid": 870, "ts": 7429770622642.301,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622665.104, "dur": 5.113,
    "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": 1418,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1418, "pid": 870, "tid": 870, "ts": 7429770622665.104,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770622700.993, "dur": 0.153,
    "args": {
      "External id": 4283,      "cid": 370, "correlation": 1419
    }
  },
  {
    "ph": "f", "id": 1419, "pid": 870, "tid": 870, "ts": 7429770622700.993,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770622704.132, "dur": 0.139,
    "args": {
      "External id": 4283,      "cid": 370, "correlation": 1420
    }
  },
  {
    "ph": "f", "id": 1420, "pid": 870, "tid": 870, "ts": 7429770622704.132,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770622706.388, "dur": 0.139,
    "args": {
      "External id": 4283,      "cid": 370, "correlation": 1421
    }
  },
  {
    "ph": "f", "id": 1421, "pid": 870, "tid": 870, "ts": 7429770622706.388,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622708.223, "dur": 7.170,
    "args": {
      "External id": 4283,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1422,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1422, "pid": 870, "tid": 870, "ts": 7429770622708.223,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622724.683, "dur": 4.200,
    "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": 1423,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1423, "pid": 870, "tid": 870, "ts": 7429770622724.683,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770622754.805, "dur": 0.200,
    "args": {
      "External id": 4294,      "cid": 370, "correlation": 1424
    }
  },
  {
    "ph": "f", "id": 1424, "pid": 870, "tid": 870, "ts": 7429770622754.805,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770622758.392, "dur": 0.139,
    "args": {
      "External id": 4294,      "cid": 370, "correlation": 1425
    }
  },
  {
    "ph": "f", "id": 1425, "pid": 870, "tid": 870, "ts": 7429770622758.392,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770622760.888, "dur": 0.138,
    "args": {
      "External id": 4294,      "cid": 370, "correlation": 1426
    }
  },
  {
    "ph": "f", "id": 1426, "pid": 870, "tid": 870, "ts": 7429770622760.888,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622769.783, "dur": 5.940,
    "args": {
      "External id": 4294,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1427,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1427, "pid": 870, "tid": 870, "ts": 7429770622769.783,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622776.371, "dur": 4.423,
    "args": {
      "External id": 4294,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1428,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1428, "pid": 870, "tid": 870, "ts": 7429770622776.371,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622789.875, "dur": 4.450,
    "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": 1429,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1429, "pid": 870, "tid": 870, "ts": 7429770622789.875,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622819.825, "dur": 3.564,
    "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": 1430,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1430, "pid": 870, "tid": 870, "ts": 7429770622819.825,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622831.036, "dur": 6.897,
    "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": 1431,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1431, "pid": 870, "tid": 870, "ts": 7429770622831.036,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622846.128, "dur": 4.069,
    "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": 1432,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1432, "pid": 870, "tid": 870, "ts": 7429770622846.128,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622855.122, "dur": 3.639,
    "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": 1433,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1433, "pid": 870, "tid": 870, "ts": 7429770622855.122,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622865.417, "dur": 4.078,
    "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": 1434,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1434, "pid": 870, "tid": 870, "ts": 7429770622865.417,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622876.707, "dur": 6.529,
    "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": 1435,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1435, "pid": 870, "tid": 870, "ts": 7429770622876.707,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770622923.086, "dur": 0.289,
    "args": {
      "External id": 4315,      "cid": 370, "correlation": 1436
    }
  },
  {
    "ph": "f", "id": 1436, "pid": 870, "tid": 870, "ts": 7429770622923.086,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770622926.691, "dur": 0.148,
    "args": {
      "External id": 4315,      "cid": 370, "correlation": 1437
    }
  },
  {
    "ph": "f", "id": 1437, "pid": 870, "tid": 870, "ts": 7429770622926.691,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770622929.128, "dur": 0.148,
    "args": {
      "External id": 4315,      "cid": 370, "correlation": 1438
    }
  },
  {
    "ph": "f", "id": 1438, "pid": 870, "tid": 870, "ts": 7429770622929.128,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622931.102, "dur": 4.896,
    "args": {
      "External id": 4315,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1439,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1439, "pid": 870, "tid": 870, "ts": 7429770622931.102,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770622969.842, "dur": 0.209,
    "args": {
      "External id": 4328,      "cid": 370, "correlation": 1440
    }
  },
  {
    "ph": "f", "id": 1440, "pid": 870, "tid": 870, "ts": 7429770622969.842,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770622972.926, "dur": 0.170,
    "args": {
      "External id": 4328,      "cid": 370, "correlation": 1441
    }
  },
  {
    "ph": "f", "id": 1441, "pid": 870, "tid": 870, "ts": 7429770622972.926,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770622975.375, "dur": 0.153,
    "args": {
      "External id": 4328,      "cid": 370, "correlation": 1442
    }
  },
  {
    "ph": "f", "id": 1442, "pid": 870, "tid": 870, "ts": 7429770622975.375,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770622976.872, "dur": 6.815,
    "args": {
      "External id": 4328,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1443,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1443, "pid": 870, "tid": 870, "ts": 7429770622976.872,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770623018.302, "dur": 0.170,
    "args": {
      "External id": 4341,      "cid": 370, "correlation": 1444
    }
  },
  {
    "ph": "f", "id": 1444, "pid": 870, "tid": 870, "ts": 7429770623018.302,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770623021.392, "dur": 0.157,
    "args": {
      "External id": 4341,      "cid": 370, "correlation": 1445
    }
  },
  {
    "ph": "f", "id": 1445, "pid": 870, "tid": 870, "ts": 7429770623021.392,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770623023.637, "dur": 0.152,
    "args": {
      "External id": 4341,      "cid": 370, "correlation": 1446
    }
  },
  {
    "ph": "f", "id": 1446, "pid": 870, "tid": 870, "ts": 7429770623023.637,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623025.093, "dur": 5.012,
    "args": {
      "External id": 4341,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1447,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1447, "pid": 870, "tid": 870, "ts": 7429770623025.093,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623052.780, "dur": 5.061,
    "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": 1448,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1448, "pid": 870, "tid": 870, "ts": 7429770623052.780,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623079.488, "dur": 5.059,
    "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": 1449,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1449, "pid": 870, "tid": 870, "ts": 7429770623079.488,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623098.302, "dur": 7.489,
    "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": 1450,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1450, "pid": 870, "tid": 870, "ts": 7429770623098.302,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623112.613, "dur": 5.354,
    "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": 1451,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1451, "pid": 870, "tid": 870, "ts": 7429770623112.613,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623124.492, "dur": 3.789,
    "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": 1452,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1452, "pid": 870, "tid": 870, "ts": 7429770623124.492,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623133.909, "dur": 3.322,
    "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": 1453,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1453, "pid": 870, "tid": 870, "ts": 7429770623133.909,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623151.937, "dur": 3.855,
    "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": 1454,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1454, "pid": 870, "tid": 870, "ts": 7429770623151.937,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623161.591, "dur": 5.027,
    "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": 1455,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1455, "pid": 870, "tid": 870, "ts": 7429770623161.591,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623172.254, "dur": 3.294,
    "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": 1456,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1456, "pid": 870, "tid": 870, "ts": 7429770623172.254,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623180.014, "dur": 3.600,
    "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": 1457,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1457, "pid": 870, "tid": 870, "ts": 7429770623180.014,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623239.180, "dur": 3.571,
    "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": 1458,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1458, "pid": 870, "tid": 870, "ts": 7429770623239.180,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623271.815, "dur": 4.146,
    "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": 1459,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1459, "pid": 870, "tid": 870, "ts": 7429770623271.815,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770623300.015, "dur": 0.409,
    "args": {
      "External id": 4416,      "cid": 290, "correlation": 1460
    }
  },
  {
    "ph": "f", "id": 1460, "pid": 870, "tid": 870, "ts": 7429770623300.015,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770623301.014, "dur": 55.679,
    "args": {
      "External id": 4416,            "bytes": 2097152, "cid": 108, "correlation": 1461, "ptr": "0x7ffd80420410"
    }
  },
  {
    "ph": "s", "id": 1461, "pid": 870, "tid": 870, "ts": 7429770623301.014,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770623372.212, "dur": 0.332,
    "args": {
      "External id": 4416,      "cid": 370, "correlation": 1462
    }
  },
  {
    "ph": "f", "id": 1462, "pid": 870, "tid": 870, "ts": 7429770623372.212,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770623377.113, "dur": 0.155,
    "args": {
      "External id": 4416,      "cid": 370, "correlation": 1463
    }
  },
  {
    "ph": "f", "id": 1463, "pid": 870, "tid": 870, "ts": 7429770623377.113,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770623380.472, "dur": 0.180,
    "args": {
      "External id": 4416,      "cid": 370, "correlation": 1464
    }
  },
  {
    "ph": "f", "id": 1464, "pid": 870, "tid": 870, "ts": 7429770623380.472,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623384.249, "dur": 6.461,
    "args": {
      "External id": 4416,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1465,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1465, "pid": 870, "tid": 870, "ts": 7429770623384.249,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623405.642, "dur": 5.634,
    "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": 1466,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1466, "pid": 870, "tid": 870, "ts": 7429770623405.642,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623427.941, "dur": 7.910,
    "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": 1467,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1467, "pid": 870, "tid": 870, "ts": 7429770623427.941,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623449.232, "dur": 5.690,
    "args": {
      "External id": 4429,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 1468,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1468, "pid": 870, "tid": 870, "ts": 7429770623449.232,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770623498.743, "dur": 0.171,
    "args": {
      "External id": 4441,      "cid": 370, "correlation": 1469
    }
  },
  {
    "ph": "f", "id": 1469, "pid": 870, "tid": 870, "ts": 7429770623498.743,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770623502.360, "dur": 0.140,
    "args": {
      "External id": 4441,      "cid": 370, "correlation": 1470
    }
  },
  {
    "ph": "f", "id": 1470, "pid": 870, "tid": 870, "ts": 7429770623502.360,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770623505.281, "dur": 0.159,
    "args": {
      "External id": 4441,      "cid": 370, "correlation": 1471
    }
  },
  {
    "ph": "f", "id": 1471, "pid": 870, "tid": 870, "ts": 7429770623505.281,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623507.729, "dur": 6.965,
    "args": {
      "External id": 4441,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1472,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1472, "pid": 870, "tid": 870, "ts": 7429770623507.729,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623532.588, "dur": 4.215,
    "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": 1473,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1473, "pid": 870, "tid": 870, "ts": 7429770623532.588,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770623586.817, "dur": 0.185,
    "args": {
      "External id": 4460,      "cid": 370, "correlation": 1474
    }
  },
  {
    "ph": "f", "id": 1474, "pid": 870, "tid": 870, "ts": 7429770623586.817,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770623589.861, "dur": 0.189,
    "args": {
      "External id": 4460,      "cid": 370, "correlation": 1475
    }
  },
  {
    "ph": "f", "id": 1475, "pid": 870, "tid": 870, "ts": 7429770623589.861,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770623592.271, "dur": 0.356,
    "args": {
      "External id": 4460,      "cid": 370, "correlation": 1476
    }
  },
  {
    "ph": "f", "id": 1476, "pid": 870, "tid": 870, "ts": 7429770623592.271,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623594.349, "dur": 5.154,
    "args": {
      "External id": 4460,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1477,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1477, "pid": 870, "tid": 870, "ts": 7429770623594.349,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623613.283, "dur": 4.110,
    "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": 1478,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1478, "pid": 870, "tid": 870, "ts": 7429770623613.283,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623635.893, "dur": 4.101,
    "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": 1479,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1479, "pid": 870, "tid": 870, "ts": 7429770623635.893,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623650.532, "dur": 6.352,
    "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": 1480,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1480, "pid": 870, "tid": 870, "ts": 7429770623650.532,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623664.910, "dur": 4.300,
    "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": 1481,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1481, "pid": 870, "tid": 870, "ts": 7429770623664.910,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623675.951, "dur": 6.367,
    "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": 1482,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1482, "pid": 870, "tid": 870, "ts": 7429770623675.951,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623689.711, "dur": 4.033,
    "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": 1483,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1483, "pid": 870, "tid": 870, "ts": 7429770623689.711,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623702.321, "dur": 4.486,
    "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": 1484,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1484, "pid": 870, "tid": 870, "ts": 7429770623702.321,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770623742.335, "dur": 0.176,
    "args": {
      "External id": 4481,      "cid": 370, "correlation": 1485
    }
  },
  {
    "ph": "f", "id": 1485, "pid": 870, "tid": 870, "ts": 7429770623742.335,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770623745.994, "dur": 0.179,
    "args": {
      "External id": 4481,      "cid": 370, "correlation": 1486
    }
  },
  {
    "ph": "f", "id": 1486, "pid": 870, "tid": 870, "ts": 7429770623745.994,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770623748.783, "dur": 0.142,
    "args": {
      "External id": 4481,      "cid": 370, "correlation": 1487
    }
  },
  {
    "ph": "f", "id": 1487, "pid": 870, "tid": 870, "ts": 7429770623748.783,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623750.851, "dur": 6.360,
    "args": {
      "External id": 4481,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1488,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1488, "pid": 870, "tid": 870, "ts": 7429770623750.851,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623774.346, "dur": 5.806,
    "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": 1489,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1489, "pid": 870, "tid": 870, "ts": 7429770623774.346,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770623810.138, "dur": 0.331,
    "args": {
      "External id": 4493,      "cid": 290, "correlation": 1490
    }
  },
  {
    "ph": "f", "id": 1490, "pid": 870, "tid": 870, "ts": 7429770623810.138,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770623810.913, "dur": 46.519,
    "args": {
      "External id": 4493,            "bytes": 79691776, "cid": 108, "correlation": 1491, "ptr": "0x7ffd8041f0c0"
    }
  },
  {
    "ph": "s", "id": 1491, "pid": 870, "tid": 870, "ts": 7429770623810.913,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770623859.982, "dur": 0.205,
    "args": {
      "External id": 4492,      "cid": 370, "correlation": 1492
    }
  },
  {
    "ph": "f", "id": 1492, "pid": 870, "tid": 870, "ts": 7429770623859.982,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770623864.177, "dur": 0.140,
    "args": {
      "External id": 4492,      "cid": 370, "correlation": 1493
    }
  },
  {
    "ph": "f", "id": 1493, "pid": 870, "tid": 870, "ts": 7429770623864.177,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770623866.961, "dur": 0.137,
    "args": {
      "External id": 4492,      "cid": 370, "correlation": 1494
    }
  },
  {
    "ph": "f", "id": 1494, "pid": 870, "tid": 870, "ts": 7429770623866.961,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623869.473, "dur": 5.707,
    "args": {
      "External id": 4492,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1495,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1495, "pid": 870, "tid": 870, "ts": 7429770623869.473,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623886.510, "dur": 4.468,
    "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": 1496,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1496, "pid": 870, "tid": 870, "ts": 7429770623886.510,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770623918.163, "dur": 0.193,
    "args": {
      "External id": 4503,      "cid": 370, "correlation": 1497
    }
  },
  {
    "ph": "f", "id": 1497, "pid": 870, "tid": 870, "ts": 7429770623918.163,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770623921.434, "dur": 0.164,
    "args": {
      "External id": 4503,      "cid": 370, "correlation": 1498
    }
  },
  {
    "ph": "f", "id": 1498, "pid": 870, "tid": 870, "ts": 7429770623921.434,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770623924.251, "dur": 0.139,
    "args": {
      "External id": 4503,      "cid": 370, "correlation": 1499
    }
  },
  {
    "ph": "f", "id": 1499, "pid": 870, "tid": 870, "ts": 7429770623924.251,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623933.801, "dur": 5.870,
    "args": {
      "External id": 4503,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1500,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1500, "pid": 870, "tid": 870, "ts": 7429770623933.801,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623940.020, "dur": 6.206,
    "args": {
      "External id": 4503,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1501,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1501, "pid": 870, "tid": 870, "ts": 7429770623940.020,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623956.741, "dur": 5.813,
    "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": 1502,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1502, "pid": 870, "tid": 870, "ts": 7429770623956.741,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770623990.566, "dur": 4.767,
    "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": 1503,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1503, "pid": 870, "tid": 870, "ts": 7429770623990.566,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624005.311, "dur": 5.043,
    "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": 1504,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1504, "pid": 870, "tid": 870, "ts": 7429770624005.311,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624018.392, "dur": 3.667,
    "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": 1505,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1505, "pid": 870, "tid": 870, "ts": 7429770624018.392,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624029.089, "dur": 3.788,
    "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": 1506,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1506, "pid": 870, "tid": 870, "ts": 7429770624029.089,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624039.586, "dur": 5.817,
    "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": 1507,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1507, "pid": 870, "tid": 870, "ts": 7429770624039.586,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624053.472, "dur": 5.805,
    "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": 1508,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1508, "pid": 870, "tid": 870, "ts": 7429770624053.472,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770624103.335, "dur": 0.224,
    "args": {
      "External id": 4524,      "cid": 370, "correlation": 1509
    }
  },
  {
    "ph": "f", "id": 1509, "pid": 870, "tid": 870, "ts": 7429770624103.335,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770624107.142, "dur": 0.159,
    "args": {
      "External id": 4524,      "cid": 370, "correlation": 1510
    }
  },
  {
    "ph": "f", "id": 1510, "pid": 870, "tid": 870, "ts": 7429770624107.142,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770624109.475, "dur": 0.149,
    "args": {
      "External id": 4524,      "cid": 370, "correlation": 1511
    }
  },
  {
    "ph": "f", "id": 1511, "pid": 870, "tid": 870, "ts": 7429770624109.475,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624111.587, "dur": 6.213,
    "args": {
      "External id": 4524,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1512,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1512, "pid": 870, "tid": 870, "ts": 7429770624111.587,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770624158.243, "dur": 0.191,
    "args": {
      "External id": 4537,      "cid": 370, "correlation": 1513
    }
  },
  {
    "ph": "f", "id": 1513, "pid": 870, "tid": 870, "ts": 7429770624158.243,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770624161.416, "dur": 0.156,
    "args": {
      "External id": 4537,      "cid": 370, "correlation": 1514
    }
  },
  {
    "ph": "f", "id": 1514, "pid": 870, "tid": 870, "ts": 7429770624161.416,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770624163.453, "dur": 0.148,
    "args": {
      "External id": 4537,      "cid": 370, "correlation": 1515
    }
  },
  {
    "ph": "f", "id": 1515, "pid": 870, "tid": 870, "ts": 7429770624163.453,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624165.286, "dur": 6.977,
    "args": {
      "External id": 4537,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1516,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1516, "pid": 870, "tid": 870, "ts": 7429770624165.286,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770624208.912, "dur": 0.167,
    "args": {
      "External id": 4550,      "cid": 370, "correlation": 1517
    }
  },
  {
    "ph": "f", "id": 1517, "pid": 870, "tid": 870, "ts": 7429770624208.912,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770624212.181, "dur": 0.142,
    "args": {
      "External id": 4550,      "cid": 370, "correlation": 1518
    }
  },
  {
    "ph": "f", "id": 1518, "pid": 870, "tid": 870, "ts": 7429770624212.181,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770624213.989, "dur": 0.152,
    "args": {
      "External id": 4550,      "cid": 370, "correlation": 1519
    }
  },
  {
    "ph": "f", "id": 1519, "pid": 870, "tid": 870, "ts": 7429770624213.989,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624215.500, "dur": 5.613,
    "args": {
      "External id": 4550,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1520,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1520, "pid": 870, "tid": 870, "ts": 7429770624215.500,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624244.885, "dur": 7.210,
    "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": 1521,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1521, "pid": 870, "tid": 870, "ts": 7429770624244.885,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624274.661, "dur": 5.657,
    "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": 1522,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1522, "pid": 870, "tid": 870, "ts": 7429770624274.661,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624295.289, "dur": 7.168,
    "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": 1523,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1523, "pid": 870, "tid": 870, "ts": 7429770624295.289,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624309.723, "dur": 3.771,
    "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": 1524,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1524, "pid": 870, "tid": 870, "ts": 7429770624309.723,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624319.467, "dur": 4.144,
    "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": 1525,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1525, "pid": 870, "tid": 870, "ts": 7429770624319.467,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624329.660, "dur": 3.350,
    "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": 1526,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1526, "pid": 870, "tid": 870, "ts": 7429770624329.660,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624348.554, "dur": 3.456,
    "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": 1527,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1527, "pid": 870, "tid": 870, "ts": 7429770624348.554,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624359.296, "dur": 4.735,
    "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": 1528,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1528, "pid": 870, "tid": 870, "ts": 7429770624359.296,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624370.120, "dur": 4.090,
    "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": 1529,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1529, "pid": 870, "tid": 870, "ts": 7429770624370.120,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624378.751, "dur": 3.205,
    "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": 1530,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1530, "pid": 870, "tid": 870, "ts": 7429770624378.751,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624443.673, "dur": 4.439,
    "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": 1531,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1531, "pid": 870, "tid": 870, "ts": 7429770624443.673,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624511.520, "dur": 4.671,
    "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": 1532,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1532, "pid": 870, "tid": 870, "ts": 7429770624511.520,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770624540.170, "dur": 0.339,
    "args": {
      "External id": 4625,      "cid": 290, "correlation": 1533
    }
  },
  {
    "ph": "f", "id": 1533, "pid": 870, "tid": 870, "ts": 7429770624540.170,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770624540.982, "dur": 48.960,
    "args": {
      "External id": 4625,            "bytes": 2097152, "cid": 108, "correlation": 1534, "ptr": "0x7ffd80420410"
    }
  },
  {
    "ph": "s", "id": 1534, "pid": 870, "tid": 870, "ts": 7429770624540.982,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770624605.482, "dur": 0.308,
    "args": {
      "External id": 4625,      "cid": 370, "correlation": 1535
    }
  },
  {
    "ph": "f", "id": 1535, "pid": 870, "tid": 870, "ts": 7429770624605.482,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770624610.360, "dur": 0.152,
    "args": {
      "External id": 4625,      "cid": 370, "correlation": 1536
    }
  },
  {
    "ph": "f", "id": 1536, "pid": 870, "tid": 870, "ts": 7429770624610.360,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770624613.515, "dur": 0.180,
    "args": {
      "External id": 4625,      "cid": 370, "correlation": 1537
    }
  },
  {
    "ph": "f", "id": 1537, "pid": 870, "tid": 870, "ts": 7429770624613.515,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624616.857, "dur": 5.596,
    "args": {
      "External id": 4625,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1538,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1538, "pid": 870, "tid": 870, "ts": 7429770624616.857,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624638.143, "dur": 4.215,
    "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": 1539,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1539, "pid": 870, "tid": 870, "ts": 7429770624638.143,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624658.456, "dur": 5.844,
    "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": 1540,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1540, "pid": 870, "tid": 870, "ts": 7429770624658.456,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624678.123, "dur": 6.893,
    "args": {
      "External id": 4638,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 1541,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1541, "pid": 870, "tid": 870, "ts": 7429770624678.123,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770624729.437, "dur": 0.184,
    "args": {
      "External id": 4650,      "cid": 370, "correlation": 1542
    }
  },
  {
    "ph": "f", "id": 1542, "pid": 870, "tid": 870, "ts": 7429770624729.437,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770624732.718, "dur": 0.142,
    "args": {
      "External id": 4650,      "cid": 370, "correlation": 1543
    }
  },
  {
    "ph": "f", "id": 1543, "pid": 870, "tid": 870, "ts": 7429770624732.718,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770624735.592, "dur": 0.166,
    "args": {
      "External id": 4650,      "cid": 370, "correlation": 1544
    }
  },
  {
    "ph": "f", "id": 1544, "pid": 870, "tid": 870, "ts": 7429770624735.592,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624737.909, "dur": 5.363,
    "args": {
      "External id": 4650,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1545,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1545, "pid": 870, "tid": 870, "ts": 7429770624737.909,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624759.905, "dur": 4.592,
    "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": 1546,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1546, "pid": 870, "tid": 870, "ts": 7429770624759.905,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770624809.005, "dur": 0.216,
    "args": {
      "External id": 4669,      "cid": 370, "correlation": 1547
    }
  },
  {
    "ph": "f", "id": 1547, "pid": 870, "tid": 870, "ts": 7429770624809.005,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770624812.408, "dur": 0.181,
    "args": {
      "External id": 4669,      "cid": 370, "correlation": 1548
    }
  },
  {
    "ph": "f", "id": 1548, "pid": 870, "tid": 870, "ts": 7429770624812.408,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770624814.596, "dur": 0.359,
    "args": {
      "External id": 4669,      "cid": 370, "correlation": 1549
    }
  },
  {
    "ph": "f", "id": 1549, "pid": 870, "tid": 870, "ts": 7429770624814.596,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624816.667, "dur": 5.048,
    "args": {
      "External id": 4669,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1550,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1550, "pid": 870, "tid": 870, "ts": 7429770624816.667,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624836.263, "dur": 4.317,
    "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": 1551,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1551, "pid": 870, "tid": 870, "ts": 7429770624836.263,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624859.565, "dur": 5.801,
    "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": 1552,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1552, "pid": 870, "tid": 870, "ts": 7429770624859.565,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624875.433, "dur": 5.242,
    "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": 1553,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1553, "pid": 870, "tid": 870, "ts": 7429770624875.433,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624888.957, "dur": 4.459,
    "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": 1554,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1554, "pid": 870, "tid": 870, "ts": 7429770624888.957,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624899.942, "dur": 5.642,
    "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": 1555,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1555, "pid": 870, "tid": 870, "ts": 7429770624899.942,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624912.171, "dur": 3.950,
    "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": 1556,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1556, "pid": 870, "tid": 870, "ts": 7429770624912.171,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624924.788, "dur": 3.786,
    "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": 1557,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1557, "pid": 870, "tid": 870, "ts": 7429770624924.788,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770624970.659, "dur": 0.204,
    "args": {
      "External id": 4690,      "cid": 370, "correlation": 1558
    }
  },
  {
    "ph": "f", "id": 1558, "pid": 870, "tid": 870, "ts": 7429770624970.659,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770624974.073, "dur": 0.140,
    "args": {
      "External id": 4690,      "cid": 370, "correlation": 1559
    }
  },
  {
    "ph": "f", "id": 1559, "pid": 870, "tid": 870, "ts": 7429770624974.073,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770624976.694, "dur": 0.141,
    "args": {
      "External id": 4690,      "cid": 370, "correlation": 1560
    }
  },
  {
    "ph": "f", "id": 1560, "pid": 870, "tid": 870, "ts": 7429770624976.694,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770624978.707, "dur": 5.748,
    "args": {
      "External id": 4690,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1561,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1561, "pid": 870, "tid": 870, "ts": 7429770624978.707,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625000.759, "dur": 5.546,
    "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": 1562,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1562, "pid": 870, "tid": 870, "ts": 7429770625000.759,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770625030.809, "dur": 0.165,
    "args": {
      "External id": 4701,      "cid": 370, "correlation": 1563
    }
  },
  {
    "ph": "f", "id": 1563, "pid": 870, "tid": 870, "ts": 7429770625030.809,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770625033.932, "dur": 0.136,
    "args": {
      "External id": 4701,      "cid": 370, "correlation": 1564
    }
  },
  {
    "ph": "f", "id": 1564, "pid": 870, "tid": 870, "ts": 7429770625033.932,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770625035.874, "dur": 0.139,
    "args": {
      "External id": 4701,      "cid": 370, "correlation": 1565
    }
  },
  {
    "ph": "f", "id": 1565, "pid": 870, "tid": 870, "ts": 7429770625035.874,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625037.582, "dur": 7.191,
    "args": {
      "External id": 4701,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1566,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1566, "pid": 870, "tid": 870, "ts": 7429770625037.582,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625053.404, "dur": 3.959,
    "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": 1567,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1567, "pid": 870, "tid": 870, "ts": 7429770625053.404,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770625079.767, "dur": 0.166,
    "args": {
      "External id": 4712,      "cid": 370, "correlation": 1568
    }
  },
  {
    "ph": "f", "id": 1568, "pid": 870, "tid": 870, "ts": 7429770625079.767,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770625083.395, "dur": 0.140,
    "args": {
      "External id": 4712,      "cid": 370, "correlation": 1569
    }
  },
  {
    "ph": "f", "id": 1569, "pid": 870, "tid": 870, "ts": 7429770625083.395,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770625086.361, "dur": 0.139,
    "args": {
      "External id": 4712,      "cid": 370, "correlation": 1570
    }
  },
  {
    "ph": "f", "id": 1570, "pid": 870, "tid": 870, "ts": 7429770625086.361,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625095.202, "dur": 5.578,
    "args": {
      "External id": 4712,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1571,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1571, "pid": 870, "tid": 870, "ts": 7429770625095.202,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625101.348, "dur": 4.456,
    "args": {
      "External id": 4712,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1572,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1572, "pid": 870, "tid": 870, "ts": 7429770625101.348,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625115.061, "dur": 4.729,
    "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": 1573,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1573, "pid": 870, "tid": 870, "ts": 7429770625115.061,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625144.378, "dur": 5.515,
    "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": 1574,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1574, "pid": 870, "tid": 870, "ts": 7429770625144.378,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625157.718, "dur": 4.447,
    "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": 1575,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1575, "pid": 870, "tid": 870, "ts": 7429770625157.718,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625169.807, "dur": 3.400,
    "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": 1576,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1576, "pid": 870, "tid": 870, "ts": 7429770625169.807,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625177.506, "dur": 3.919,
    "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": 1577,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1577, "pid": 870, "tid": 870, "ts": 7429770625177.506,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625188.212, "dur": 4.204,
    "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": 1578,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1578, "pid": 870, "tid": 870, "ts": 7429770625188.212,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625199.066, "dur": 5.984,
    "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": 1579,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1579, "pid": 870, "tid": 870, "ts": 7429770625199.066,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770625247.079, "dur": 0.303,
    "args": {
      "External id": 4733,      "cid": 370, "correlation": 1580
    }
  },
  {
    "ph": "f", "id": 1580, "pid": 870, "tid": 870, "ts": 7429770625247.079,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770625250.518, "dur": 0.153,
    "args": {
      "External id": 4733,      "cid": 370, "correlation": 1581
    }
  },
  {
    "ph": "f", "id": 1581, "pid": 870, "tid": 870, "ts": 7429770625250.518,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770625252.778, "dur": 0.148,
    "args": {
      "External id": 4733,      "cid": 370, "correlation": 1582
    }
  },
  {
    "ph": "f", "id": 1582, "pid": 870, "tid": 870, "ts": 7429770625252.778,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625254.856, "dur": 5.834,
    "args": {
      "External id": 4733,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1583,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1583, "pid": 870, "tid": 870, "ts": 7429770625254.856,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770625296.936, "dur": 0.235,
    "args": {
      "External id": 4746,      "cid": 370, "correlation": 1584
    }
  },
  {
    "ph": "f", "id": 1584, "pid": 870, "tid": 870, "ts": 7429770625296.936,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770625299.898, "dur": 0.175,
    "args": {
      "External id": 4746,      "cid": 370, "correlation": 1585
    }
  },
  {
    "ph": "f", "id": 1585, "pid": 870, "tid": 870, "ts": 7429770625299.898,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770625302.298, "dur": 0.150,
    "args": {
      "External id": 4746,      "cid": 370, "correlation": 1586
    }
  },
  {
    "ph": "f", "id": 1586, "pid": 870, "tid": 870, "ts": 7429770625302.298,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625303.863, "dur": 5.292,
    "args": {
      "External id": 4746,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1587,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1587, "pid": 870, "tid": 870, "ts": 7429770625303.863,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770625343.268, "dur": 0.180,
    "args": {
      "External id": 4759,      "cid": 370, "correlation": 1588
    }
  },
  {
    "ph": "f", "id": 1588, "pid": 870, "tid": 870, "ts": 7429770625343.268,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770625345.968, "dur": 0.162,
    "args": {
      "External id": 4759,      "cid": 370, "correlation": 1589
    }
  },
  {
    "ph": "f", "id": 1589, "pid": 870, "tid": 870, "ts": 7429770625345.968,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770625347.929, "dur": 0.155,
    "args": {
      "External id": 4759,      "cid": 370, "correlation": 1590
    }
  },
  {
    "ph": "f", "id": 1590, "pid": 870, "tid": 870, "ts": 7429770625347.929,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625349.447, "dur": 4.993,
    "args": {
      "External id": 4759,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1591,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1591, "pid": 870, "tid": 870, "ts": 7429770625349.447,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625377.808, "dur": 3.714,
    "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": 1592,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1592, "pid": 870, "tid": 870, "ts": 7429770625377.808,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625400.807, "dur": 5.309,
    "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": 1593,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1593, "pid": 870, "tid": 870, "ts": 7429770625400.807,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625419.224, "dur": 8.035,
    "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": 1594,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1594, "pid": 870, "tid": 870, "ts": 7429770625419.224,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625435.154, "dur": 4.713,
    "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": 1595,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1595, "pid": 870, "tid": 870, "ts": 7429770625435.154,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625445.078, "dur": 3.940,
    "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": 1596,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1596, "pid": 870, "tid": 870, "ts": 7429770625445.078,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625454.512, "dur": 4.945,
    "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": 1597,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1597, "pid": 870, "tid": 870, "ts": 7429770625454.512,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625473.290, "dur": 3.683,
    "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": 1598,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1598, "pid": 870, "tid": 870, "ts": 7429770625473.290,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625484.143, "dur": 4.785,
    "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": 1599,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1599, "pid": 870, "tid": 870, "ts": 7429770625484.143,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625494.778, "dur": 4.030,
    "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": 1600,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1600, "pid": 870, "tid": 870, "ts": 7429770625494.778,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625504.156, "dur": 4.394,
    "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": 1601,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1601, "pid": 870, "tid": 870, "ts": 7429770625504.156,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625564.610, "dur": 4.282,
    "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": 1602,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1602, "pid": 870, "tid": 870, "ts": 7429770625564.610,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625598.170, "dur": 4.419,
    "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": 1603,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1603, "pid": 870, "tid": 870, "ts": 7429770625598.170,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770625626.451, "dur": 0.339,
    "args": {
      "External id": 4834,      "cid": 290, "correlation": 1604
    }
  },
  {
    "ph": "f", "id": 1604, "pid": 870, "tid": 870, "ts": 7429770625626.451,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770625627.346, "dur": 54.451,
    "args": {
      "External id": 4834,            "bytes": 2097152, "cid": 108, "correlation": 1605, "ptr": "0x7ffd80420410"
    }
  },
  {
    "ph": "s", "id": 1605, "pid": 870, "tid": 870, "ts": 7429770625627.346,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770625699.597, "dur": 0.317,
    "args": {
      "External id": 4834,      "cid": 370, "correlation": 1606
    }
  },
  {
    "ph": "f", "id": 1606, "pid": 870, "tid": 870, "ts": 7429770625699.597,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770625704.700, "dur": 0.156,
    "args": {
      "External id": 4834,      "cid": 370, "correlation": 1607
    }
  },
  {
    "ph": "f", "id": 1607, "pid": 870, "tid": 870, "ts": 7429770625704.700,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770625707.693, "dur": 0.163,
    "args": {
      "External id": 4834,      "cid": 370, "correlation": 1608
    }
  },
  {
    "ph": "f", "id": 1608, "pid": 870, "tid": 870, "ts": 7429770625707.693,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625711.340, "dur": 7.039,
    "args": {
      "External id": 4834,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1609,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1609, "pid": 870, "tid": 870, "ts": 7429770625711.340,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625733.477, "dur": 4.787,
    "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": 1610,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1610, "pid": 870, "tid": 870, "ts": 7429770625733.477,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625754.928, "dur": 6.893,
    "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": 1611,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1611, "pid": 870, "tid": 870, "ts": 7429770625754.928,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625774.349, "dur": 6.348,
    "args": {
      "External id": 4847,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 1612,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1612, "pid": 870, "tid": 870, "ts": 7429770625774.349,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770625826.939, "dur": 0.180,
    "args": {
      "External id": 4859,      "cid": 370, "correlation": 1613
    }
  },
  {
    "ph": "f", "id": 1613, "pid": 870, "tid": 870, "ts": 7429770625826.939,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770625830.739, "dur": 0.139,
    "args": {
      "External id": 4859,      "cid": 370, "correlation": 1614
    }
  },
  {
    "ph": "f", "id": 1614, "pid": 870, "tid": 870, "ts": 7429770625830.739,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770625833.495, "dur": 0.152,
    "args": {
      "External id": 4859,      "cid": 370, "correlation": 1615
    }
  },
  {
    "ph": "f", "id": 1615, "pid": 870, "tid": 870, "ts": 7429770625833.495,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625835.885, "dur": 6.935,
    "args": {
      "External id": 4859,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1616,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1616, "pid": 870, "tid": 870, "ts": 7429770625835.885,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625861.065, "dur": 4.325,
    "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": 1617,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1617, "pid": 870, "tid": 870, "ts": 7429770625861.065,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770625914.156, "dur": 0.256,
    "args": {
      "External id": 4878,      "cid": 370, "correlation": 1618
    }
  },
  {
    "ph": "f", "id": 1618, "pid": 870, "tid": 870, "ts": 7429770625914.156,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770625917.612, "dur": 0.185,
    "args": {
      "External id": 4878,      "cid": 370, "correlation": 1619
    }
  },
  {
    "ph": "f", "id": 1619, "pid": 870, "tid": 870, "ts": 7429770625917.612,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770625920.389, "dur": 0.355,
    "args": {
      "External id": 4878,      "cid": 370, "correlation": 1620
    }
  },
  {
    "ph": "f", "id": 1620, "pid": 870, "tid": 870, "ts": 7429770625920.389,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625922.522, "dur": 5.291,
    "args": {
      "External id": 4878,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1621,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1621, "pid": 870, "tid": 870, "ts": 7429770625922.522,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625941.761, "dur": 3.863,
    "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": 1622,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1622, "pid": 870, "tid": 870, "ts": 7429770625941.761,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625965.424, "dur": 4.664,
    "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": 1623,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1623, "pid": 870, "tid": 870, "ts": 7429770625965.424,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625980.924, "dur": 6.469,
    "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": 1624,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1624, "pid": 870, "tid": 870, "ts": 7429770625980.924,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770625997.027, "dur": 3.648,
    "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": 1625,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1625, "pid": 870, "tid": 870, "ts": 7429770625997.027,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626008.058, "dur": 4.792,
    "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": 1626,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1626, "pid": 870, "tid": 870, "ts": 7429770626008.058,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626019.367, "dur": 3.939,
    "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": 1627,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1627, "pid": 870, "tid": 870, "ts": 7429770626019.367,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626032.226, "dur": 3.245,
    "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": 1628,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1628, "pid": 870, "tid": 870, "ts": 7429770626032.226,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770626101.666, "dur": 0.397,
    "args": {
      "External id": 4899,      "cid": 370, "correlation": 1629
    }
  },
  {
    "ph": "f", "id": 1629, "pid": 870, "tid": 870, "ts": 7429770626101.666,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770626106.829, "dur": 0.183,
    "args": {
      "External id": 4899,      "cid": 370, "correlation": 1630
    }
  },
  {
    "ph": "f", "id": 1630, "pid": 870, "tid": 870, "ts": 7429770626106.829,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770626109.771, "dur": 0.139,
    "args": {
      "External id": 4899,      "cid": 370, "correlation": 1631
    }
  },
  {
    "ph": "f", "id": 1631, "pid": 870, "tid": 870, "ts": 7429770626109.771,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626112.717, "dur": 7.291,
    "args": {
      "External id": 4899,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1632,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1632, "pid": 870, "tid": 870, "ts": 7429770626112.717,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626142.660, "dur": 5.293,
    "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": 1633,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1633, "pid": 870, "tid": 870, "ts": 7429770626142.660,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770626182.766, "dur": 0.147,
    "args": {
      "External id": 4910,      "cid": 370, "correlation": 1634
    }
  },
  {
    "ph": "f", "id": 1634, "pid": 870, "tid": 870, "ts": 7429770626182.766,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770626185.751, "dur": 0.170,
    "args": {
      "External id": 4910,      "cid": 370, "correlation": 1635
    }
  },
  {
    "ph": "f", "id": 1635, "pid": 870, "tid": 870, "ts": 7429770626185.751,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770626187.990, "dur": 0.146,
    "args": {
      "External id": 4910,      "cid": 370, "correlation": 1636
    }
  },
  {
    "ph": "f", "id": 1636, "pid": 870, "tid": 870, "ts": 7429770626187.990,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626189.809, "dur": 4.888,
    "args": {
      "External id": 4910,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1637,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1637, "pid": 870, "tid": 870, "ts": 7429770626189.809,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626204.989, "dur": 4.155,
    "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": 1638,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1638, "pid": 870, "tid": 870, "ts": 7429770626204.989,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770626235.497, "dur": 0.157,
    "args": {
      "External id": 4921,      "cid": 370, "correlation": 1639
    }
  },
  {
    "ph": "f", "id": 1639, "pid": 870, "tid": 870, "ts": 7429770626235.497,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770626238.572, "dur": 0.138,
    "args": {
      "External id": 4921,      "cid": 370, "correlation": 1640
    }
  },
  {
    "ph": "f", "id": 1640, "pid": 870, "tid": 870, "ts": 7429770626238.572,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770626241.035, "dur": 0.139,
    "args": {
      "External id": 4921,      "cid": 370, "correlation": 1641
    }
  },
  {
    "ph": "f", "id": 1641, "pid": 870, "tid": 870, "ts": 7429770626241.035,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626250.345, "dur": 5.593,
    "args": {
      "External id": 4921,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1642,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1642, "pid": 870, "tid": 870, "ts": 7429770626250.345,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626256.376, "dur": 4.569,
    "args": {
      "External id": 4921,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1643,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1643, "pid": 870, "tid": 870, "ts": 7429770626256.376,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626270.903, "dur": 4.461,
    "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": 1644,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1644, "pid": 870, "tid": 870, "ts": 7429770626270.903,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626304.645, "dur": 5.298,
    "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": 1645,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1645, "pid": 870, "tid": 870, "ts": 7429770626304.645,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626320.499, "dur": 5.124,
    "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": 1646,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1646, "pid": 870, "tid": 870, "ts": 7429770626320.499,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626333.742, "dur": 3.736,
    "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": 1647,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1647, "pid": 870, "tid": 870, "ts": 7429770626333.742,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626344.526, "dur": 4.130,
    "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": 1648,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1648, "pid": 870, "tid": 870, "ts": 7429770626344.526,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626354.800, "dur": 3.745,
    "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": 1649,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1649, "pid": 870, "tid": 870, "ts": 7429770626354.800,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626367.356, "dur": 4.018,
    "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": 1650,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1650, "pid": 870, "tid": 870, "ts": 7429770626367.356,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770626411.500, "dur": 0.293,
    "args": {
      "External id": 4942,      "cid": 370, "correlation": 1651
    }
  },
  {
    "ph": "f", "id": 1651, "pid": 870, "tid": 870, "ts": 7429770626411.500,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770626415.161, "dur": 0.151,
    "args": {
      "External id": 4942,      "cid": 370, "correlation": 1652
    }
  },
  {
    "ph": "f", "id": 1652, "pid": 870, "tid": 870, "ts": 7429770626415.161,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770626417.352, "dur": 0.152,
    "args": {
      "External id": 4942,      "cid": 370, "correlation": 1653
    }
  },
  {
    "ph": "f", "id": 1653, "pid": 870, "tid": 870, "ts": 7429770626417.352,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626419.498, "dur": 6.336,
    "args": {
      "External id": 4942,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1654,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1654, "pid": 870, "tid": 870, "ts": 7429770626419.498,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770626463.700, "dur": 0.195,
    "args": {
      "External id": 4955,      "cid": 370, "correlation": 1655
    }
  },
  {
    "ph": "f", "id": 1655, "pid": 870, "tid": 870, "ts": 7429770626463.700,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770626466.599, "dur": 0.152,
    "args": {
      "External id": 4955,      "cid": 370, "correlation": 1656
    }
  },
  {
    "ph": "f", "id": 1656, "pid": 870, "tid": 870, "ts": 7429770626466.599,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770626468.848, "dur": 0.157,
    "args": {
      "External id": 4955,      "cid": 370, "correlation": 1657
    }
  },
  {
    "ph": "f", "id": 1657, "pid": 870, "tid": 870, "ts": 7429770626468.848,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626470.636, "dur": 5.290,
    "args": {
      "External id": 4955,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1658,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1658, "pid": 870, "tid": 870, "ts": 7429770626470.636,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770626510.205, "dur": 0.182,
    "args": {
      "External id": 4968,      "cid": 370, "correlation": 1659
    }
  },
  {
    "ph": "f", "id": 1659, "pid": 870, "tid": 870, "ts": 7429770626510.205,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770626513.217, "dur": 0.141,
    "args": {
      "External id": 4968,      "cid": 370, "correlation": 1660
    }
  },
  {
    "ph": "f", "id": 1660, "pid": 870, "tid": 870, "ts": 7429770626513.217,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770626515.286, "dur": 0.152,
    "args": {
      "External id": 4968,      "cid": 370, "correlation": 1661
    }
  },
  {
    "ph": "f", "id": 1661, "pid": 870, "tid": 870, "ts": 7429770626515.286,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626516.846, "dur": 5.245,
    "args": {
      "External id": 4968,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1662,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1662, "pid": 870, "tid": 870, "ts": 7429770626516.846,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626545.515, "dur": 4.085,
    "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": 1663,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1663, "pid": 870, "tid": 870, "ts": 7429770626545.515,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626572.043, "dur": 5.587,
    "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": 1664,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1664, "pid": 870, "tid": 870, "ts": 7429770626572.043,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626592.752, "dur": 7.548,
    "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": 1665,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1665, "pid": 870, "tid": 870, "ts": 7429770626592.752,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626607.940, "dur": 4.768,
    "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": 1666,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1666, "pid": 870, "tid": 870, "ts": 7429770626607.940,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626618.021, "dur": 6.504,
    "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": 1667,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1667, "pid": 870, "tid": 870, "ts": 7429770626618.021,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626631.527, "dur": 3.440,
    "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": 1668,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1668, "pid": 870, "tid": 870, "ts": 7429770626631.527,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626647.860, "dur": 3.825,
    "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": 1669,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1669, "pid": 870, "tid": 870, "ts": 7429770626647.860,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626657.802, "dur": 4.838,
    "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": 1670,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1670, "pid": 870, "tid": 870, "ts": 7429770626657.802,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626669.716, "dur": 5.171,
    "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": 1671,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1671, "pid": 870, "tid": 870, "ts": 7429770626669.716,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626679.363, "dur": 3.965,
    "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": 1672,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1672, "pid": 870, "tid": 870, "ts": 7429770626679.363,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626742.948, "dur": 4.647,
    "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": 1673,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1673, "pid": 870, "tid": 870, "ts": 7429770626742.948,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626779.536, "dur": 4.736,
    "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": 1674,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1674, "pid": 870, "tid": 870, "ts": 7429770626779.536,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770626809.567, "dur": 0.390,
    "args": {
      "External id": 5043,      "cid": 290, "correlation": 1675
    }
  },
  {
    "ph": "f", "id": 1675, "pid": 870, "tid": 870, "ts": 7429770626809.567,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770626810.670, "dur": 59.696,
    "args": {
      "External id": 5043,            "bytes": 2097152, "cid": 108, "correlation": 1676, "ptr": "0x7ffd80420410"
    }
  },
  {
    "ph": "s", "id": 1676, "pid": 870, "tid": 870, "ts": 7429770626810.670,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770626887.017, "dur": 0.234,
    "args": {
      "External id": 5044,      "cid": 290, "correlation": 1677
    }
  },
  {
    "ph": "f", "id": 1677, "pid": 870, "tid": 870, "ts": 7429770626887.017,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770626887.656, "dur": 30.605,
    "args": {
      "External id": 5044,            "bytes": 79691776, "cid": 108, "correlation": 1678, "ptr": "0x7"
    }
  },
  {
    "ph": "s", "id": 1678, "pid": 870, "tid": 870, "ts": 7429770626887.656,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770626919.887, "dur": 0.318,
    "args": {
      "External id": 5043,      "cid": 370, "correlation": 1679
    }
  },
  {
    "ph": "f", "id": 1679, "pid": 870, "tid": 870, "ts": 7429770626919.887,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770626924.815, "dur": 0.162,
    "args": {
      "External id": 5043,      "cid": 370, "correlation": 1680
    }
  },
  {
    "ph": "f", "id": 1680, "pid": 870, "tid": 870, "ts": 7429770626924.815,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770626928.958, "dur": 0.151,
    "args": {
      "External id": 5043,      "cid": 370, "correlation": 1681
    }
  },
  {
    "ph": "f", "id": 1681, "pid": 870, "tid": 870, "ts": 7429770626928.958,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626933.276, "dur": 6.490,
    "args": {
      "External id": 5043,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1682,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1682, "pid": 870, "tid": 870, "ts": 7429770626933.276,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626954.893, "dur": 5.443,
    "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": 1683,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1683, "pid": 870, "tid": 870, "ts": 7429770626954.893,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626977.778, "dur": 4.183,
    "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": 1684,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1684, "pid": 870, "tid": 870, "ts": 7429770626977.778,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770626995.393, "dur": 6.492,
    "args": {
      "External id": 5056,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 1685,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1685, "pid": 870, "tid": 870, "ts": 7429770626995.393,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770627047.114, "dur": 0.190,
    "args": {
      "External id": 5068,      "cid": 370, "correlation": 1686
    }
  },
  {
    "ph": "f", "id": 1686, "pid": 870, "tid": 870, "ts": 7429770627047.114,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770627051.149, "dur": 0.140,
    "args": {
      "External id": 5068,      "cid": 370, "correlation": 1687
    }
  },
  {
    "ph": "f", "id": 1687, "pid": 870, "tid": 870, "ts": 7429770627051.149,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770627053.755, "dur": 0.150,
    "args": {
      "External id": 5068,      "cid": 370, "correlation": 1688
    }
  },
  {
    "ph": "f", "id": 1688, "pid": 870, "tid": 870, "ts": 7429770627053.755,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627056.279, "dur": 5.777,
    "args": {
      "External id": 5068,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1689,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1689, "pid": 870, "tid": 870, "ts": 7429770627056.279,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627080.463, "dur": 4.957,
    "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": 1690,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1690, "pid": 870, "tid": 870, "ts": 7429770627080.463,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770627134.376, "dur": 0.248,
    "args": {
      "External id": 5087,      "cid": 370, "correlation": 1691
    }
  },
  {
    "ph": "f", "id": 1691, "pid": 870, "tid": 870, "ts": 7429770627134.376,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770627137.591, "dur": 0.214,
    "args": {
      "External id": 5087,      "cid": 370, "correlation": 1692
    }
  },
  {
    "ph": "f", "id": 1692, "pid": 870, "tid": 870, "ts": 7429770627137.591,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770627140.131, "dur": 0.357,
    "args": {
      "External id": 5087,      "cid": 370, "correlation": 1693
    }
  },
  {
    "ph": "f", "id": 1693, "pid": 870, "tid": 870, "ts": 7429770627140.131,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627142.421, "dur": 5.717,
    "args": {
      "External id": 5087,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1694,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1694, "pid": 870, "tid": 870, "ts": 7429770627142.421,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627161.948, "dur": 5.483,
    "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": 1695,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1695, "pid": 870, "tid": 870, "ts": 7429770627161.948,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627186.449, "dur": 4.935,
    "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": 1696,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1696, "pid": 870, "tid": 870, "ts": 7429770627186.449,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627201.410, "dur": 7.017,
    "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": 1697,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1697, "pid": 870, "tid": 870, "ts": 7429770627201.410,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627216.433, "dur": 4.063,
    "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": 1698,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1698, "pid": 870, "tid": 870, "ts": 7429770627216.433,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627226.673, "dur": 5.533,
    "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": 1699,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1699, "pid": 870, "tid": 870, "ts": 7429770627226.673,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627239.537, "dur": 4.174,
    "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": 1700,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1700, "pid": 870, "tid": 870, "ts": 7429770627239.537,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627251.507, "dur": 4.864,
    "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": 1701,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1701, "pid": 870, "tid": 870, "ts": 7429770627251.507,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770627294.921, "dur": 0.193,
    "args": {
      "External id": 5108,      "cid": 370, "correlation": 1702
    }
  },
  {
    "ph": "f", "id": 1702, "pid": 870, "tid": 870, "ts": 7429770627294.921,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770627298.492, "dur": 0.159,
    "args": {
      "External id": 5108,      "cid": 370, "correlation": 1703
    }
  },
  {
    "ph": "f", "id": 1703, "pid": 870, "tid": 870, "ts": 7429770627298.492,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770627300.860, "dur": 0.141,
    "args": {
      "External id": 5108,      "cid": 370, "correlation": 1704
    }
  },
  {
    "ph": "f", "id": 1704, "pid": 870, "tid": 870, "ts": 7429770627300.860,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627302.885, "dur": 5.688,
    "args": {
      "External id": 5108,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1705,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1705, "pid": 870, "tid": 870, "ts": 7429770627302.885,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627325.070, "dur": 5.250,
    "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": 1706,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1706, "pid": 870, "tid": 870, "ts": 7429770627325.070,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770627355.723, "dur": 0.158,
    "args": {
      "External id": 5119,      "cid": 370, "correlation": 1707
    }
  },
  {
    "ph": "f", "id": 1707, "pid": 870, "tid": 870, "ts": 7429770627355.723,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770627358.608, "dur": 0.159,
    "args": {
      "External id": 5119,      "cid": 370, "correlation": 1708
    }
  },
  {
    "ph": "f", "id": 1708, "pid": 870, "tid": 870, "ts": 7429770627358.608,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770627360.701, "dur": 0.141,
    "args": {
      "External id": 5119,      "cid": 370, "correlation": 1709
    }
  },
  {
    "ph": "f", "id": 1709, "pid": 870, "tid": 870, "ts": 7429770627360.701,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627362.395, "dur": 9.641,
    "args": {
      "External id": 5119,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1710,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1710, "pid": 870, "tid": 870, "ts": 7429770627362.395,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627416.846, "dur": 4.279,
    "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": 1711,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1711, "pid": 870, "tid": 870, "ts": 7429770627416.846,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770627441.898, "dur": 0.166,
    "args": {
      "External id": 5130,      "cid": 370, "correlation": 1712
    }
  },
  {
    "ph": "f", "id": 1712, "pid": 870, "tid": 870, "ts": 7429770627441.898,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770627445.057, "dur": 0.139,
    "args": {
      "External id": 5130,      "cid": 370, "correlation": 1713
    }
  },
  {
    "ph": "f", "id": 1713, "pid": 870, "tid": 870, "ts": 7429770627445.057,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770627447.909, "dur": 0.145,
    "args": {
      "External id": 5130,      "cid": 370, "correlation": 1714
    }
  },
  {
    "ph": "f", "id": 1714, "pid": 870, "tid": 870, "ts": 7429770627447.909,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627456.460, "dur": 6.109,
    "args": {
      "External id": 5130,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1715,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1715, "pid": 870, "tid": 870, "ts": 7429770627456.460,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627463.145, "dur": 4.173,
    "args": {
      "External id": 5130,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1716,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1716, "pid": 870, "tid": 870, "ts": 7429770627463.145,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627476.485, "dur": 4.985,
    "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": 1717,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1717, "pid": 870, "tid": 870, "ts": 7429770627476.485,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627506.146, "dur": 4.967,
    "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": 1718,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1718, "pid": 870, "tid": 870, "ts": 7429770627506.146,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627520.334, "dur": 5.549,
    "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": 1719,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1719, "pid": 870, "tid": 870, "ts": 7429770627520.334,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627532.920, "dur": 3.778,
    "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": 1720,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1720, "pid": 870, "tid": 870, "ts": 7429770627532.920,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627542.021, "dur": 4.724,
    "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": 1721,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1721, "pid": 870, "tid": 870, "ts": 7429770627542.021,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627552.357, "dur": 3.989,
    "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": 1722,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1722, "pid": 870, "tid": 870, "ts": 7429770627552.357,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627564.490, "dur": 4.006,
    "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": 1723,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1723, "pid": 870, "tid": 870, "ts": 7429770627564.490,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770627606.411, "dur": 0.216,
    "args": {
      "External id": 5151,      "cid": 370, "correlation": 1724
    }
  },
  {
    "ph": "f", "id": 1724, "pid": 870, "tid": 870, "ts": 7429770627606.411,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770627609.636, "dur": 0.155,
    "args": {
      "External id": 5151,      "cid": 370, "correlation": 1725
    }
  },
  {
    "ph": "f", "id": 1725, "pid": 870, "tid": 870, "ts": 7429770627609.636,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770627611.996, "dur": 0.150,
    "args": {
      "External id": 5151,      "cid": 370, "correlation": 1726
    }
  },
  {
    "ph": "f", "id": 1726, "pid": 870, "tid": 870, "ts": 7429770627611.996,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627614.015, "dur": 6.186,
    "args": {
      "External id": 5151,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1727,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1727, "pid": 870, "tid": 870, "ts": 7429770627614.015,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770627655.209, "dur": 0.186,
    "args": {
      "External id": 5164,      "cid": 370, "correlation": 1728
    }
  },
  {
    "ph": "f", "id": 1728, "pid": 870, "tid": 870, "ts": 7429770627655.209,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770627658.274, "dur": 0.151,
    "args": {
      "External id": 5164,      "cid": 370, "correlation": 1729
    }
  },
  {
    "ph": "f", "id": 1729, "pid": 870, "tid": 870, "ts": 7429770627658.274,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770627660.578, "dur": 0.147,
    "args": {
      "External id": 5164,      "cid": 370, "correlation": 1730
    }
  },
  {
    "ph": "f", "id": 1730, "pid": 870, "tid": 870, "ts": 7429770627660.578,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627662.129, "dur": 5.401,
    "args": {
      "External id": 5164,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1731,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1731, "pid": 870, "tid": 870, "ts": 7429770627662.129,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770627699.190, "dur": 0.159,
    "args": {
      "External id": 5177,      "cid": 370, "correlation": 1732
    }
  },
  {
    "ph": "f", "id": 1732, "pid": 870, "tid": 870, "ts": 7429770627699.190,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770627702.016, "dur": 0.159,
    "args": {
      "External id": 5177,      "cid": 370, "correlation": 1733
    }
  },
  {
    "ph": "f", "id": 1733, "pid": 870, "tid": 870, "ts": 7429770627702.016,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770627704.020, "dur": 0.152,
    "args": {
      "External id": 5177,      "cid": 370, "correlation": 1734
    }
  },
  {
    "ph": "f", "id": 1734, "pid": 870, "tid": 870, "ts": 7429770627704.020,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627705.472, "dur": 6.044,
    "args": {
      "External id": 5177,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1735,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1735, "pid": 870, "tid": 870, "ts": 7429770627705.472,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627733.961, "dur": 4.851,
    "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": 1736,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1736, "pid": 870, "tid": 870, "ts": 7429770627733.961,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627761.161, "dur": 5.750,
    "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": 1737,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1737, "pid": 870, "tid": 870, "ts": 7429770627761.161,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627781.107, "dur": 6.082,
    "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": 1738,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1738, "pid": 870, "tid": 870, "ts": 7429770627781.107,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627795.566, "dur": 4.713,
    "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": 1739,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1739, "pid": 870, "tid": 870, "ts": 7429770627795.566,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627805.394, "dur": 3.712,
    "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": 1740,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1740, "pid": 870, "tid": 870, "ts": 7429770627805.394,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627814.931, "dur": 4.616,
    "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": 1741,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1741, "pid": 870, "tid": 870, "ts": 7429770627814.931,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627832.845, "dur": 3.943,
    "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": 1742,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1742, "pid": 870, "tid": 870, "ts": 7429770627832.845,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627843.349, "dur": 4.609,
    "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": 1743,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1743, "pid": 870, "tid": 870, "ts": 7429770627843.349,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627853.917, "dur": 4.063,
    "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": 1744,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1744, "pid": 870, "tid": 870, "ts": 7429770627853.917,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627863.060, "dur": 4.843,
    "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": 1745,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1745, "pid": 870, "tid": 870, "ts": 7429770627863.060,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627924.586, "dur": 4.310,
    "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": 1746,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1746, "pid": 870, "tid": 870, "ts": 7429770627924.586,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770627957.649, "dur": 5.139,
    "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": 1747,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1747, "pid": 870, "tid": 870, "ts": 7429770627957.649,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770627985.564, "dur": 0.430,
    "args": {
      "External id": 5252,      "cid": 290, "correlation": 1748
    }
  },
  {
    "ph": "f", "id": 1748, "pid": 870, "tid": 870, "ts": 7429770627985.564,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770627986.543, "dur": 56.772,
    "args": {
      "External id": 5252,            "bytes": 2097152, "cid": 108, "correlation": 1749, "ptr": "0x7ffd80420410"
    }
  },
  {
    "ph": "s", "id": 1749, "pid": 870, "tid": 870, "ts": 7429770627986.543,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770628060.319, "dur": 0.391,
    "args": {
      "External id": 5252,      "cid": 370, "correlation": 1750
    }
  },
  {
    "ph": "f", "id": 1750, "pid": 870, "tid": 870, "ts": 7429770628060.319,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770628065.184, "dur": 0.153,
    "args": {
      "External id": 5252,      "cid": 370, "correlation": 1751
    }
  },
  {
    "ph": "f", "id": 1751, "pid": 870, "tid": 870, "ts": 7429770628065.184,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770628068.364, "dur": 0.153,
    "args": {
      "External id": 5252,      "cid": 370, "correlation": 1752
    }
  },
  {
    "ph": "f", "id": 1752, "pid": 870, "tid": 870, "ts": 7429770628068.364,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628072.094, "dur": 5.340,
    "args": {
      "External id": 5252,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1753,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1753, "pid": 870, "tid": 870, "ts": 7429770628072.094,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628092.818, "dur": 5.034,
    "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": 1754,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1754, "pid": 870, "tid": 870, "ts": 7429770628092.818,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628114.653, "dur": 6.720,
    "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": 1755,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1755, "pid": 870, "tid": 870, "ts": 7429770628114.653,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628134.528, "dur": 6.737,
    "args": {
      "External id": 5265,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 1756,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1756, "pid": 870, "tid": 870, "ts": 7429770628134.528,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770628186.449, "dur": 0.175,
    "args": {
      "External id": 5277,      "cid": 370, "correlation": 1757
    }
  },
  {
    "ph": "f", "id": 1757, "pid": 870, "tid": 870, "ts": 7429770628186.449,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770628190.245, "dur": 0.142,
    "args": {
      "External id": 5277,      "cid": 370, "correlation": 1758
    }
  },
  {
    "ph": "f", "id": 1758, "pid": 870, "tid": 870, "ts": 7429770628190.245,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770628193.036, "dur": 0.143,
    "args": {
      "External id": 5277,      "cid": 370, "correlation": 1759
    }
  },
  {
    "ph": "f", "id": 1759, "pid": 870, "tid": 870, "ts": 7429770628193.036,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628195.317, "dur": 6.605,
    "args": {
      "External id": 5277,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1760,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1760, "pid": 870, "tid": 870, "ts": 7429770628195.317,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628219.842, "dur": 4.467,
    "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": 1761,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1761, "pid": 870, "tid": 870, "ts": 7429770628219.842,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770628272.306, "dur": 0.225,
    "args": {
      "External id": 5296,      "cid": 370, "correlation": 1762
    }
  },
  {
    "ph": "f", "id": 1762, "pid": 870, "tid": 870, "ts": 7429770628272.306,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770628275.713, "dur": 0.192,
    "args": {
      "External id": 5296,      "cid": 370, "correlation": 1763
    }
  },
  {
    "ph": "f", "id": 1763, "pid": 870, "tid": 870, "ts": 7429770628275.713,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770628278.335, "dur": 0.190,
    "args": {
      "External id": 5296,      "cid": 370, "correlation": 1764
    }
  },
  {
    "ph": "f", "id": 1764, "pid": 870, "tid": 870, "ts": 7429770628278.335,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628280.210, "dur": 4.811,
    "args": {
      "External id": 5296,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1765,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1765, "pid": 870, "tid": 870, "ts": 7429770628280.210,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628300.131, "dur": 3.645,
    "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": 1766,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1766, "pid": 870, "tid": 870, "ts": 7429770628300.131,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628323.647, "dur": 5.469,
    "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": 1767,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1767, "pid": 870, "tid": 870, "ts": 7429770628323.647,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628341.017, "dur": 7.274,
    "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": 1768,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1768, "pid": 870, "tid": 870, "ts": 7429770628341.017,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628357.647, "dur": 4.537,
    "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": 1769,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1769, "pid": 870, "tid": 870, "ts": 7429770628357.647,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628368.790, "dur": 4.928,
    "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": 1770,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1770, "pid": 870, "tid": 870, "ts": 7429770628368.790,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628380.372, "dur": 6.167,
    "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": 1771,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1771, "pid": 870, "tid": 870, "ts": 7429770628380.372,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628394.561, "dur": 4.770,
    "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": 1772,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1772, "pid": 870, "tid": 870, "ts": 7429770628394.561,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770628436.823, "dur": 0.189,
    "args": {
      "External id": 5317,      "cid": 370, "correlation": 1773
    }
  },
  {
    "ph": "f", "id": 1773, "pid": 870, "tid": 870, "ts": 7429770628436.823,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770628440.064, "dur": 0.159,
    "args": {
      "External id": 5317,      "cid": 370, "correlation": 1774
    }
  },
  {
    "ph": "f", "id": 1774, "pid": 870, "tid": 870, "ts": 7429770628440.064,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770628442.728, "dur": 0.139,
    "args": {
      "External id": 5317,      "cid": 370, "correlation": 1775
    }
  },
  {
    "ph": "f", "id": 1775, "pid": 870, "tid": 870, "ts": 7429770628442.728,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628444.794, "dur": 5.411,
    "args": {
      "External id": 5317,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1776,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1776, "pid": 870, "tid": 870, "ts": 7429770628444.794,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628467.319, "dur": 5.954,
    "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": 1777,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1777, "pid": 870, "tid": 870, "ts": 7429770628467.319,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770628501.556, "dur": 0.180,
    "args": {
      "External id": 5328,      "cid": 370, "correlation": 1778
    }
  },
  {
    "ph": "f", "id": 1778, "pid": 870, "tid": 870, "ts": 7429770628501.556,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770628504.168, "dur": 0.136,
    "args": {
      "External id": 5328,      "cid": 370, "correlation": 1779
    }
  },
  {
    "ph": "f", "id": 1779, "pid": 870, "tid": 870, "ts": 7429770628504.168,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770628506.360, "dur": 0.141,
    "args": {
      "External id": 5328,      "cid": 370, "correlation": 1780
    }
  },
  {
    "ph": "f", "id": 1780, "pid": 870, "tid": 870, "ts": 7429770628506.360,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628508.011, "dur": 4.858,
    "args": {
      "External id": 5328,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1781,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1781, "pid": 870, "tid": 870, "ts": 7429770628508.011,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628522.231, "dur": 4.441,
    "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": 1782,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1782, "pid": 870, "tid": 870, "ts": 7429770628522.231,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770628549.428, "dur": 0.173,
    "args": {
      "External id": 5339,      "cid": 370, "correlation": 1783
    }
  },
  {
    "ph": "f", "id": 1783, "pid": 870, "tid": 870, "ts": 7429770628549.428,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770628552.583, "dur": 0.139,
    "args": {
      "External id": 5339,      "cid": 370, "correlation": 1784
    }
  },
  {
    "ph": "f", "id": 1784, "pid": 870, "tid": 870, "ts": 7429770628552.583,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770628555.091, "dur": 0.136,
    "args": {
      "External id": 5339,      "cid": 370, "correlation": 1785
    }
  },
  {
    "ph": "f", "id": 1785, "pid": 870, "tid": 870, "ts": 7429770628555.091,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628563.778, "dur": 5.313,
    "args": {
      "External id": 5339,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1786,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1786, "pid": 870, "tid": 870, "ts": 7429770628563.778,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628569.463, "dur": 3.681,
    "args": {
      "External id": 5339,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1787,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1787, "pid": 870, "tid": 870, "ts": 7429770628569.463,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628582.320, "dur": 5.349,
    "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": 1788,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1788, "pid": 870, "tid": 870, "ts": 7429770628582.320,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628613.622, "dur": 5.124,
    "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": 1789,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1789, "pid": 870, "tid": 870, "ts": 7429770628613.622,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628626.925, "dur": 4.813,
    "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": 1790,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1790, "pid": 870, "tid": 870, "ts": 7429770628626.925,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628639.084, "dur": 4.143,
    "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": 1791,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1791, "pid": 870, "tid": 870, "ts": 7429770628639.084,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628648.082, "dur": 3.551,
    "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": 1792,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1792, "pid": 870, "tid": 870, "ts": 7429770628648.082,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628657.065, "dur": 4.652,
    "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": 1793,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1793, "pid": 870, "tid": 870, "ts": 7429770628657.065,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628670.097, "dur": 3.839,
    "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": 1794,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1794, "pid": 870, "tid": 870, "ts": 7429770628670.097,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770628712.300, "dur": 0.226,
    "args": {
      "External id": 5360,      "cid": 370, "correlation": 1795
    }
  },
  {
    "ph": "f", "id": 1795, "pid": 870, "tid": 870, "ts": 7429770628712.300,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770628715.789, "dur": 0.144,
    "args": {
      "External id": 5360,      "cid": 370, "correlation": 1796
    }
  },
  {
    "ph": "f", "id": 1796, "pid": 870, "tid": 870, "ts": 7429770628715.789,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770628718.080, "dur": 0.151,
    "args": {
      "External id": 5360,      "cid": 370, "correlation": 1797
    }
  },
  {
    "ph": "f", "id": 1797, "pid": 870, "tid": 870, "ts": 7429770628718.080,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628720.232, "dur": 5.938,
    "args": {
      "External id": 5360,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1798,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1798, "pid": 870, "tid": 870, "ts": 7429770628720.232,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770628762.102, "dur": 0.193,
    "args": {
      "External id": 5373,      "cid": 370, "correlation": 1799
    }
  },
  {
    "ph": "f", "id": 1799, "pid": 870, "tid": 870, "ts": 7429770628762.102,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770628765.154, "dur": 0.140,
    "args": {
      "External id": 5373,      "cid": 370, "correlation": 1800
    }
  },
  {
    "ph": "f", "id": 1800, "pid": 870, "tid": 870, "ts": 7429770628765.154,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770628767.190, "dur": 0.155,
    "args": {
      "External id": 5373,      "cid": 370, "correlation": 1801
    }
  },
  {
    "ph": "f", "id": 1801, "pid": 870, "tid": 870, "ts": 7429770628767.190,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628769.006, "dur": 6.474,
    "args": {
      "External id": 5373,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1802,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1802, "pid": 870, "tid": 870, "ts": 7429770628769.006,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770628810.875, "dur": 0.170,
    "args": {
      "External id": 5386,      "cid": 370, "correlation": 1803
    }
  },
  {
    "ph": "f", "id": 1803, "pid": 870, "tid": 870, "ts": 7429770628810.875,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770628814.046, "dur": 0.135,
    "args": {
      "External id": 5386,      "cid": 370, "correlation": 1804
    }
  },
  {
    "ph": "f", "id": 1804, "pid": 870, "tid": 870, "ts": 7429770628814.046,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770628816.540, "dur": 0.149,
    "args": {
      "External id": 5386,      "cid": 370, "correlation": 1805
    }
  },
  {
    "ph": "f", "id": 1805, "pid": 870, "tid": 870, "ts": 7429770628816.540,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628818.216, "dur": 5.493,
    "args": {
      "External id": 5386,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1806,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1806, "pid": 870, "tid": 870, "ts": 7429770628818.216,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628845.743, "dur": 4.008,
    "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": 1807,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1807, "pid": 870, "tid": 870, "ts": 7429770628845.743,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628872.417, "dur": 5.971,
    "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": 1808,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1808, "pid": 870, "tid": 870, "ts": 7429770628872.417,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628891.086, "dur": 6.522,
    "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": 1809,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1809, "pid": 870, "tid": 870, "ts": 7429770628891.086,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628904.924, "dur": 5.362,
    "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": 1810,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1810, "pid": 870, "tid": 870, "ts": 7429770628904.924,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628916.749, "dur": 6.466,
    "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": 1811,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1811, "pid": 870, "tid": 870, "ts": 7429770628916.749,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628928.904, "dur": 3.467,
    "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": 1812,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1812, "pid": 870, "tid": 870, "ts": 7429770628928.904,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628946.140, "dur": 4.405,
    "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": 1813,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1813, "pid": 870, "tid": 870, "ts": 7429770628946.140,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628957.551, "dur": 8.004,
    "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": 1814,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1814, "pid": 870, "tid": 870, "ts": 7429770628957.551,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628971.719, "dur": 4.063,
    "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": 1815,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1815, "pid": 870, "tid": 870, "ts": 7429770628971.719,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770628980.457, "dur": 5.800,
    "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": 1816,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1816, "pid": 870, "tid": 870, "ts": 7429770628980.457,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629043.773, "dur": 4.364,
    "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": 1817,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1817, "pid": 870, "tid": 870, "ts": 7429770629043.773,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629077.513, "dur": 4.436,
    "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": 1818,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1818, "pid": 870, "tid": 870, "ts": 7429770629077.513,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770629106.501, "dur": 0.445,
    "args": {
      "External id": 5461,      "cid": 290, "correlation": 1819
    }
  },
  {
    "ph": "f", "id": 1819, "pid": 870, "tid": 870, "ts": 7429770629106.501,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770629107.697, "dur": 55.151,
    "args": {
      "External id": 5461,            "bytes": 2097152, "cid": 108, "correlation": 1820, "ptr": "0x7ffd80420410"
    }
  },
  {
    "ph": "s", "id": 1820, "pid": 870, "tid": 870, "ts": 7429770629107.697,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770629177.729, "dur": 0.353,
    "args": {
      "External id": 5461,      "cid": 370, "correlation": 1821
    }
  },
  {
    "ph": "f", "id": 1821, "pid": 870, "tid": 870, "ts": 7429770629177.729,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770629182.762, "dur": 0.145,
    "args": {
      "External id": 5461,      "cid": 370, "correlation": 1822
    }
  },
  {
    "ph": "f", "id": 1822, "pid": 870, "tid": 870, "ts": 7429770629182.762,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770629185.811, "dur": 0.181,
    "args": {
      "External id": 5461,      "cid": 370, "correlation": 1823
    }
  },
  {
    "ph": "f", "id": 1823, "pid": 870, "tid": 870, "ts": 7429770629185.811,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629189.253, "dur": 5.851,
    "args": {
      "External id": 5461,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1824,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1824, "pid": 870, "tid": 870, "ts": 7429770629189.253,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629209.475, "dur": 4.185,
    "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": 1825,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1825, "pid": 870, "tid": 870, "ts": 7429770629209.475,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629230.008, "dur": 4.141,
    "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": 1826,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1826, "pid": 870, "tid": 870, "ts": 7429770629230.008,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629246.330, "dur": 6.194,
    "args": {
      "External id": 5474,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 1827,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1827, "pid": 870, "tid": 870, "ts": 7429770629246.330,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770629297.725, "dur": 0.229,
    "args": {
      "External id": 5486,      "cid": 370, "correlation": 1828
    }
  },
  {
    "ph": "f", "id": 1828, "pid": 870, "tid": 870, "ts": 7429770629297.725,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770629301.453, "dur": 0.141,
    "args": {
      "External id": 5486,      "cid": 370, "correlation": 1829
    }
  },
  {
    "ph": "f", "id": 1829, "pid": 870, "tid": 870, "ts": 7429770629301.453,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770629304.439, "dur": 0.142,
    "args": {
      "External id": 5486,      "cid": 370, "correlation": 1830
    }
  },
  {
    "ph": "f", "id": 1830, "pid": 870, "tid": 870, "ts": 7429770629304.439,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629306.684, "dur": 5.266,
    "args": {
      "External id": 5486,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1831,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1831, "pid": 870, "tid": 870, "ts": 7429770629306.684,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629328.485, "dur": 4.233,
    "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": 1832,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1832, "pid": 870, "tid": 870, "ts": 7429770629328.485,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770629376.964, "dur": 0.264,
    "args": {
      "External id": 5505,      "cid": 370, "correlation": 1833
    }
  },
  {
    "ph": "f", "id": 1833, "pid": 870, "tid": 870, "ts": 7429770629376.964,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770629380.544, "dur": 0.183,
    "args": {
      "External id": 5505,      "cid": 370, "correlation": 1834
    }
  },
  {
    "ph": "f", "id": 1834, "pid": 870, "tid": 870, "ts": 7429770629380.544,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770629382.921, "dur": 0.190,
    "args": {
      "External id": 5505,      "cid": 370, "correlation": 1835
    }
  },
  {
    "ph": "f", "id": 1835, "pid": 870, "tid": 870, "ts": 7429770629382.921,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629384.772, "dur": 4.748,
    "args": {
      "External id": 5505,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1836,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1836, "pid": 870, "tid": 870, "ts": 7429770629384.772,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629401.849, "dur": 4.592,
    "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": 1837,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1837, "pid": 870, "tid": 870, "ts": 7429770629401.849,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629424.313, "dur": 5.124,
    "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": 1838,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1838, "pid": 870, "tid": 870, "ts": 7429770629424.313,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629439.225, "dur": 6.940,
    "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": 1839,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1839, "pid": 870, "tid": 870, "ts": 7429770629439.225,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629453.695, "dur": 4.576,
    "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": 1840,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1840, "pid": 870, "tid": 870, "ts": 7429770629453.695,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629463.979, "dur": 5.308,
    "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": 1841,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1841, "pid": 870, "tid": 870, "ts": 7429770629463.979,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629477.153, "dur": 3.763,
    "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": 1842,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1842, "pid": 870, "tid": 870, "ts": 7429770629477.153,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629488.693, "dur": 4.010,
    "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": 1843,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1843, "pid": 870, "tid": 870, "ts": 7429770629488.693,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770629528.659, "dur": 0.162,
    "args": {
      "External id": 5526,      "cid": 370, "correlation": 1844
    }
  },
  {
    "ph": "f", "id": 1844, "pid": 870, "tid": 870, "ts": 7429770629528.659,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770629531.620, "dur": 0.198,
    "args": {
      "External id": 5526,      "cid": 370, "correlation": 1845
    }
  },
  {
    "ph": "f", "id": 1845, "pid": 870, "tid": 870, "ts": 7429770629531.620,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770629534.452, "dur": 0.140,
    "args": {
      "External id": 5526,      "cid": 370, "correlation": 1846
    }
  },
  {
    "ph": "f", "id": 1846, "pid": 870, "tid": 870, "ts": 7429770629534.452,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629536.352, "dur": 4.805,
    "args": {
      "External id": 5526,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1847,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1847, "pid": 870, "tid": 870, "ts": 7429770629536.352,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629556.944, "dur": 5.901,
    "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": 1848,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1848, "pid": 870, "tid": 870, "ts": 7429770629556.944,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770629588.124, "dur": 0.318,
    "args": {
      "External id": 5538,      "cid": 290, "correlation": 1849
    }
  },
  {
    "ph": "f", "id": 1849, "pid": 870, "tid": 870, "ts": 7429770629588.124,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770629588.910, "dur": 44.631,
    "args": {
      "External id": 5538,            "bytes": 79691776, "cid": 108, "correlation": 1850, "ptr": "0x7ffd8041f0c0"
    }
  },
  {
    "ph": "s", "id": 1850, "pid": 870, "tid": 870, "ts": 7429770629588.910,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770629635.282, "dur": 0.155,
    "args": {
      "External id": 5537,      "cid": 370, "correlation": 1851
    }
  },
  {
    "ph": "f", "id": 1851, "pid": 870, "tid": 870, "ts": 7429770629635.282,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770629638.926, "dur": 0.139,
    "args": {
      "External id": 5537,      "cid": 370, "correlation": 1852
    }
  },
  {
    "ph": "f", "id": 1852, "pid": 870, "tid": 870, "ts": 7429770629638.926,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770629641.176, "dur": 0.140,
    "args": {
      "External id": 5537,      "cid": 370, "correlation": 1853
    }
  },
  {
    "ph": "f", "id": 1853, "pid": 870, "tid": 870, "ts": 7429770629641.176,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629643.645, "dur": 8.422,
    "args": {
      "External id": 5537,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1854,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1854, "pid": 870, "tid": 870, "ts": 7429770629643.645,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629662.938, "dur": 4.158,
    "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": 1855,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1855, "pid": 870, "tid": 870, "ts": 7429770629662.938,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770629696.251, "dur": 0.154,
    "args": {
      "External id": 5548,      "cid": 370, "correlation": 1856
    }
  },
  {
    "ph": "f", "id": 1856, "pid": 870, "tid": 870, "ts": 7429770629696.251,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770629699.406, "dur": 0.137,
    "args": {
      "External id": 5548,      "cid": 370, "correlation": 1857
    }
  },
  {
    "ph": "f", "id": 1857, "pid": 870, "tid": 870, "ts": 7429770629699.406,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770629702.127, "dur": 0.140,
    "args": {
      "External id": 5548,      "cid": 370, "correlation": 1858
    }
  },
  {
    "ph": "f", "id": 1858, "pid": 870, "tid": 870, "ts": 7429770629702.127,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629711.279, "dur": 5.180,
    "args": {
      "External id": 5548,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1859,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1859, "pid": 870, "tid": 870, "ts": 7429770629711.279,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629717.036, "dur": 4.538,
    "args": {
      "External id": 5548,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1860,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1860, "pid": 870, "tid": 870, "ts": 7429770629717.036,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629731.518, "dur": 3.637,
    "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": 1861,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1861, "pid": 870, "tid": 870, "ts": 7429770629731.518,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629762.497, "dur": 5.015,
    "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": 1862,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1862, "pid": 870, "tid": 870, "ts": 7429770629762.497,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629776.474, "dur": 4.596,
    "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": 1863,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1863, "pid": 870, "tid": 870, "ts": 7429770629776.474,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629788.356, "dur": 3.718,
    "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": 1864,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1864, "pid": 870, "tid": 870, "ts": 7429770629788.356,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629797.577, "dur": 4.710,
    "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": 1865,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1865, "pid": 870, "tid": 870, "ts": 7429770629797.577,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629808.361, "dur": 3.687,
    "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": 1866,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1866, "pid": 870, "tid": 870, "ts": 7429770629808.361,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629819.779, "dur": 4.026,
    "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": 1867,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1867, "pid": 870, "tid": 870, "ts": 7429770629819.779,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770629862.357, "dur": 0.223,
    "args": {
      "External id": 5569,      "cid": 370, "correlation": 1868
    }
  },
  {
    "ph": "f", "id": 1868, "pid": 870, "tid": 870, "ts": 7429770629862.357,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770629865.702, "dur": 0.154,
    "args": {
      "External id": 5569,      "cid": 370, "correlation": 1869
    }
  },
  {
    "ph": "f", "id": 1869, "pid": 870, "tid": 870, "ts": 7429770629865.702,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770629867.928, "dur": 0.148,
    "args": {
      "External id": 5569,      "cid": 370, "correlation": 1870
    }
  },
  {
    "ph": "f", "id": 1870, "pid": 870, "tid": 870, "ts": 7429770629867.928,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629869.855, "dur": 6.028,
    "args": {
      "External id": 5569,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1871,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1871, "pid": 870, "tid": 870, "ts": 7429770629869.855,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770629910.956, "dur": 0.273,
    "args": {
      "External id": 5582,      "cid": 370, "correlation": 1872
    }
  },
  {
    "ph": "f", "id": 1872, "pid": 870, "tid": 870, "ts": 7429770629910.956,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770629913.810, "dur": 0.149,
    "args": {
      "External id": 5582,      "cid": 370, "correlation": 1873
    }
  },
  {
    "ph": "f", "id": 1873, "pid": 870, "tid": 870, "ts": 7429770629913.810,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770629915.865, "dur": 0.148,
    "args": {
      "External id": 5582,      "cid": 370, "correlation": 1874
    }
  },
  {
    "ph": "f", "id": 1874, "pid": 870, "tid": 870, "ts": 7429770629915.865,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629917.519, "dur": 5.584,
    "args": {
      "External id": 5582,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1875,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1875, "pid": 870, "tid": 870, "ts": 7429770629917.519,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770629958.231, "dur": 0.183,
    "args": {
      "External id": 5595,      "cid": 370, "correlation": 1876
    }
  },
  {
    "ph": "f", "id": 1876, "pid": 870, "tid": 870, "ts": 7429770629958.231,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770629961.020, "dur": 0.138,
    "args": {
      "External id": 5595,      "cid": 370, "correlation": 1877
    }
  },
  {
    "ph": "f", "id": 1877, "pid": 870, "tid": 870, "ts": 7429770629961.020,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770629962.907, "dur": 0.150,
    "args": {
      "External id": 5595,      "cid": 370, "correlation": 1878
    }
  },
  {
    "ph": "f", "id": 1878, "pid": 870, "tid": 870, "ts": 7429770629962.907,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629964.432, "dur": 5.277,
    "args": {
      "External id": 5595,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1879,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1879, "pid": 870, "tid": 870, "ts": 7429770629964.432,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770629992.409, "dur": 3.851,
    "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": 1880,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1880, "pid": 870, "tid": 870, "ts": 7429770629992.409,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630017.414, "dur": 5.551,
    "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": 1881,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1881, "pid": 870, "tid": 870, "ts": 7429770630017.414,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630036.954, "dur": 8.061,
    "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": 1882,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1882, "pid": 870, "tid": 870, "ts": 7429770630036.954,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630052.237, "dur": 5.055,
    "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": 1883,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1883, "pid": 870, "tid": 870, "ts": 7429770630052.237,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630062.464, "dur": 5.288,
    "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": 1884,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1884, "pid": 870, "tid": 870, "ts": 7429770630062.464,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630096.657, "dur": 4.813,
    "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": 1885,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1885, "pid": 870, "tid": 870, "ts": 7429770630096.657,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630118.032, "dur": 4.308,
    "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": 1886,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1886, "pid": 870, "tid": 870, "ts": 7429770630118.032,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630130.302, "dur": 5.005,
    "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": 1887,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1887, "pid": 870, "tid": 870, "ts": 7429770630130.302,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630141.799, "dur": 4.240,
    "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": 1888,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1888, "pid": 870, "tid": 870, "ts": 7429770630141.799,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630150.858, "dur": 4.796,
    "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": 1889,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1889, "pid": 870, "tid": 870, "ts": 7429770630150.858,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630242.523, "dur": 5.433,
    "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": 1890,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1890, "pid": 870, "tid": 870, "ts": 7429770630242.523,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630275.716, "dur": 3.273,
    "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": 1891,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1891, "pid": 870, "tid": 870, "ts": 7429770630275.716,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770630304.156, "dur": 0.415,
    "args": {
      "External id": 5670,      "cid": 290, "correlation": 1892
    }
  },
  {
    "ph": "f", "id": 1892, "pid": 870, "tid": 870, "ts": 7429770630304.156,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770630305.241, "dur": 51.088,
    "args": {
      "External id": 5670,            "bytes": 2097152, "cid": 108, "correlation": 1893, "ptr": "0x7ffd80420410"
    }
  },
  {
    "ph": "s", "id": 1893, "pid": 870, "tid": 870, "ts": 7429770630305.241,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770630375.578, "dur": 0.441,
    "args": {
      "External id": 5670,      "cid": 370, "correlation": 1894
    }
  },
  {
    "ph": "f", "id": 1894, "pid": 870, "tid": 870, "ts": 7429770630375.578,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770630381.888, "dur": 0.212,
    "args": {
      "External id": 5670,      "cid": 370, "correlation": 1895
    }
  },
  {
    "ph": "f", "id": 1895, "pid": 870, "tid": 870, "ts": 7429770630381.888,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770630385.421, "dur": 0.184,
    "args": {
      "External id": 5670,      "cid": 370, "correlation": 1896
    }
  },
  {
    "ph": "f", "id": 1896, "pid": 870, "tid": 870, "ts": 7429770630385.421,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630391.132, "dur": 7.123,
    "args": {
      "External id": 5670,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1897,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1897, "pid": 870, "tid": 870, "ts": 7429770630391.132,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630415.603, "dur": 5.279,
    "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": 1898,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1898, "pid": 870, "tid": 870, "ts": 7429770630415.603,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630436.944, "dur": 6.430,
    "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": 1899,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1899, "pid": 870, "tid": 870, "ts": 7429770630436.944,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630457.633, "dur": 6.766,
    "args": {
      "External id": 5683,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 1900,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1900, "pid": 870, "tid": 870, "ts": 7429770630457.633,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770630509.938, "dur": 0.238,
    "args": {
      "External id": 5695,      "cid": 370, "correlation": 1901
    }
  },
  {
    "ph": "f", "id": 1901, "pid": 870, "tid": 870, "ts": 7429770630509.938,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770630513.857, "dur": 0.141,
    "args": {
      "External id": 5695,      "cid": 370, "correlation": 1902
    }
  },
  {
    "ph": "f", "id": 1902, "pid": 870, "tid": 870, "ts": 7429770630513.857,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770630516.685, "dur": 0.166,
    "args": {
      "External id": 5695,      "cid": 370, "correlation": 1903
    }
  },
  {
    "ph": "f", "id": 1903, "pid": 870, "tid": 870, "ts": 7429770630516.685,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630519.547, "dur": 5.895,
    "args": {
      "External id": 5695,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1904,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1904, "pid": 870, "tid": 870, "ts": 7429770630519.547,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630542.817, "dur": 4.133,
    "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": 1905,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1905, "pid": 870, "tid": 870, "ts": 7429770630542.817,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770630597.048, "dur": 0.237,
    "args": {
      "External id": 5714,      "cid": 370, "correlation": 1906
    }
  },
  {
    "ph": "f", "id": 1906, "pid": 870, "tid": 870, "ts": 7429770630597.048,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770630600.297, "dur": 0.193,
    "args": {
      "External id": 5714,      "cid": 370, "correlation": 1907
    }
  },
  {
    "ph": "f", "id": 1907, "pid": 870, "tid": 870, "ts": 7429770630600.297,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770630602.892, "dur": 0.195,
    "args": {
      "External id": 5714,      "cid": 370, "correlation": 1908
    }
  },
  {
    "ph": "f", "id": 1908, "pid": 870, "tid": 870, "ts": 7429770630602.892,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630604.824, "dur": 5.682,
    "args": {
      "External id": 5714,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1909,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1909, "pid": 870, "tid": 870, "ts": 7429770630604.824,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630626.976, "dur": 5.054,
    "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": 1910,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1910, "pid": 870, "tid": 870, "ts": 7429770630626.976,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630653.088, "dur": 6.999,
    "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": 1911,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1911, "pid": 870, "tid": 870, "ts": 7429770630653.088,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630672.035, "dur": 7.113,
    "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": 1912,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1912, "pid": 870, "tid": 870, "ts": 7429770630672.035,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630688.125, "dur": 4.366,
    "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": 1913,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1913, "pid": 870, "tid": 870, "ts": 7429770630688.125,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630699.320, "dur": 4.379,
    "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": 1914,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1914, "pid": 870, "tid": 870, "ts": 7429770630699.320,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630711.342, "dur": 4.340,
    "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": 1915,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1915, "pid": 870, "tid": 870, "ts": 7429770630711.342,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630724.467, "dur": 5.110,
    "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": 1916,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1916, "pid": 870, "tid": 870, "ts": 7429770630724.467,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770630765.194, "dur": 0.231,
    "args": {
      "External id": 5735,      "cid": 370, "correlation": 1917
    }
  },
  {
    "ph": "f", "id": 1917, "pid": 870, "tid": 870, "ts": 7429770630765.194,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770630768.933, "dur": 0.181,
    "args": {
      "External id": 5735,      "cid": 370, "correlation": 1918
    }
  },
  {
    "ph": "f", "id": 1918, "pid": 870, "tid": 870, "ts": 7429770630768.933,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770630771.469, "dur": 0.170,
    "args": {
      "External id": 5735,      "cid": 370, "correlation": 1919
    }
  },
  {
    "ph": "f", "id": 1919, "pid": 870, "tid": 870, "ts": 7429770630771.469,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630773.719, "dur": 5.431,
    "args": {
      "External id": 5735,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1920,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1920, "pid": 870, "tid": 870, "ts": 7429770630773.719,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630796.695, "dur": 5.660,
    "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": 1921,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1921, "pid": 870, "tid": 870, "ts": 7429770630796.695,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770630828.015, "dur": 0.209,
    "args": {
      "External id": 5746,      "cid": 370, "correlation": 1922
    }
  },
  {
    "ph": "f", "id": 1922, "pid": 870, "tid": 870, "ts": 7429770630828.015,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770630830.755, "dur": 0.141,
    "args": {
      "External id": 5746,      "cid": 370, "correlation": 1923
    }
  },
  {
    "ph": "f", "id": 1923, "pid": 870, "tid": 870, "ts": 7429770630830.755,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770630832.769, "dur": 0.139,
    "args": {
      "External id": 5746,      "cid": 370, "correlation": 1924
    }
  },
  {
    "ph": "f", "id": 1924, "pid": 870, "tid": 870, "ts": 7429770630832.769,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630834.377, "dur": 5.357,
    "args": {
      "External id": 5746,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1925,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1925, "pid": 870, "tid": 870, "ts": 7429770630834.377,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630848.599, "dur": 4.314,
    "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": 1926,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1926, "pid": 870, "tid": 870, "ts": 7429770630848.599,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770630874.722, "dur": 0.186,
    "args": {
      "External id": 5757,      "cid": 370, "correlation": 1927
    }
  },
  {
    "ph": "f", "id": 1927, "pid": 870, "tid": 870, "ts": 7429770630874.722,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770630877.564, "dur": 0.139,
    "args": {
      "External id": 5757,      "cid": 370, "correlation": 1928
    }
  },
  {
    "ph": "f", "id": 1928, "pid": 870, "tid": 870, "ts": 7429770630877.564,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770630879.941, "dur": 0.141,
    "args": {
      "External id": 5757,      "cid": 370, "correlation": 1929
    }
  },
  {
    "ph": "f", "id": 1929, "pid": 870, "tid": 870, "ts": 7429770630879.941,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630888.848, "dur": 5.324,
    "args": {
      "External id": 5757,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1930,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1930, "pid": 870, "tid": 870, "ts": 7429770630888.848,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630894.553, "dur": 3.818,
    "args": {
      "External id": 5757,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 1931,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1931, "pid": 870, "tid": 870, "ts": 7429770630894.553,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630908.212, "dur": 4.631,
    "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": 1932,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1932, "pid": 870, "tid": 870, "ts": 7429770630908.212,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630939.380, "dur": 5.233,
    "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": 1933,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1933, "pid": 870, "tid": 870, "ts": 7429770630939.380,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630952.914, "dur": 5.183,
    "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": 1934,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1934, "pid": 870, "tid": 870, "ts": 7429770630952.914,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630965.671, "dur": 3.693,
    "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": 1935,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1935, "pid": 870, "tid": 870, "ts": 7429770630965.671,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630974.885, "dur": 3.937,
    "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": 1936,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1936, "pid": 870, "tid": 870, "ts": 7429770630974.885,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630984.434, "dur": 3.964,
    "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": 1937,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1937, "pid": 870, "tid": 870, "ts": 7429770630984.434,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770630996.171, "dur": 3.456,
    "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": 1938,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1938, "pid": 870, "tid": 870, "ts": 7429770630996.171,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770631035.863, "dur": 0.306,
    "args": {
      "External id": 5778,      "cid": 370, "correlation": 1939
    }
  },
  {
    "ph": "f", "id": 1939, "pid": 870, "tid": 870, "ts": 7429770631035.863,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770631039.490, "dur": 0.155,
    "args": {
      "External id": 5778,      "cid": 370, "correlation": 1940
    }
  },
  {
    "ph": "f", "id": 1940, "pid": 870, "tid": 870, "ts": 7429770631039.490,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770631041.698, "dur": 0.151,
    "args": {
      "External id": 5778,      "cid": 370, "correlation": 1941
    }
  },
  {
    "ph": "f", "id": 1941, "pid": 870, "tid": 870, "ts": 7429770631041.698,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631043.974, "dur": 6.805,
    "args": {
      "External id": 5778,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1942,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1942, "pid": 870, "tid": 870, "ts": 7429770631043.974,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770631084.488, "dur": 0.189,
    "args": {
      "External id": 5791,      "cid": 370, "correlation": 1943
    }
  },
  {
    "ph": "f", "id": 1943, "pid": 870, "tid": 870, "ts": 7429770631084.488,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770631087.235, "dur": 0.173,
    "args": {
      "External id": 5791,      "cid": 370, "correlation": 1944
    }
  },
  {
    "ph": "f", "id": 1944, "pid": 870, "tid": 870, "ts": 7429770631087.235,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770631089.486, "dur": 0.137,
    "args": {
      "External id": 5791,      "cid": 370, "correlation": 1945
    }
  },
  {
    "ph": "f", "id": 1945, "pid": 870, "tid": 870, "ts": 7429770631089.486,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631091.195, "dur": 5.545,
    "args": {
      "External id": 5791,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1946,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1946, "pid": 870, "tid": 870, "ts": 7429770631091.195,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770631130.785, "dur": 0.162,
    "args": {
      "External id": 5804,      "cid": 370, "correlation": 1947
    }
  },
  {
    "ph": "f", "id": 1947, "pid": 870, "tid": 870, "ts": 7429770631130.785,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770631133.604, "dur": 0.158,
    "args": {
      "External id": 5804,      "cid": 370, "correlation": 1948
    }
  },
  {
    "ph": "f", "id": 1948, "pid": 870, "tid": 870, "ts": 7429770631133.604,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770631135.394, "dur": 0.140,
    "args": {
      "External id": 5804,      "cid": 370, "correlation": 1949
    }
  },
  {
    "ph": "f", "id": 1949, "pid": 870, "tid": 870, "ts": 7429770631135.394,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631136.975, "dur": 5.292,
    "args": {
      "External id": 5804,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1950,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1950, "pid": 870, "tid": 870, "ts": 7429770631136.975,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631164.874, "dur": 4.617,
    "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": 1951,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1951, "pid": 870, "tid": 870, "ts": 7429770631164.874,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631188.272, "dur": 6.208,
    "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": 1952,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1952, "pid": 870, "tid": 870, "ts": 7429770631188.272,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631207.770, "dur": 6.981,
    "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": 1953,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1953, "pid": 870, "tid": 870, "ts": 7429770631207.770,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631222.028, "dur": 5.181,
    "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": 1954,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1954, "pid": 870, "tid": 870, "ts": 7429770631222.028,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631232.181, "dur": 5.743,
    "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": 1955,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1955, "pid": 870, "tid": 870, "ts": 7429770631232.181,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631243.768, "dur": 3.733,
    "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": 1956,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1956, "pid": 870, "tid": 870, "ts": 7429770631243.768,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631260.191, "dur": 3.391,
    "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": 1957,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1957, "pid": 870, "tid": 870, "ts": 7429770631260.191,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631270.247, "dur": 4.634,
    "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": 1958,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1958, "pid": 870, "tid": 870, "ts": 7429770631270.247,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631280.404, "dur": 3.647,
    "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": 1959,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1959, "pid": 870, "tid": 870, "ts": 7429770631280.404,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631288.611, "dur": 4.706,
    "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": 1960,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1960, "pid": 870, "tid": 870, "ts": 7429770631288.611,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631347.764, "dur": 5.283,
    "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": 1961,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1961, "pid": 870, "tid": 870, "ts": 7429770631347.764,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631381.414, "dur": 5.298,
    "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": 1962,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1962, "pid": 870, "tid": 870, "ts": 7429770631381.414,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770631409.958, "dur": 0.442,
    "args": {
      "External id": 5879,      "cid": 290, "correlation": 1963
    }
  },
  {
    "ph": "f", "id": 1963, "pid": 870, "tid": 870, "ts": 7429770631409.958,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770631411.043, "dur": 52.739,
    "args": {
      "External id": 5879,            "bytes": 2097152, "cid": 108, "correlation": 1964, "ptr": "0x7ffd80420410"
    }
  },
  {
    "ph": "s", "id": 1964, "pid": 870, "tid": 870, "ts": 7429770631411.043,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770631478.637, "dur": 0.273,
    "args": {
      "External id": 5879,      "cid": 370, "correlation": 1965
    }
  },
  {
    "ph": "f", "id": 1965, "pid": 870, "tid": 870, "ts": 7429770631478.637,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770631483.566, "dur": 0.159,
    "args": {
      "External id": 5879,      "cid": 370, "correlation": 1966
    }
  },
  {
    "ph": "f", "id": 1966, "pid": 870, "tid": 870, "ts": 7429770631483.566,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770631486.670, "dur": 0.176,
    "args": {
      "External id": 5879,      "cid": 370, "correlation": 1967
    }
  },
  {
    "ph": "f", "id": 1967, "pid": 870, "tid": 870, "ts": 7429770631486.670,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631490.508, "dur": 5.621,
    "args": {
      "External id": 5879,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1968,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1968, "pid": 870, "tid": 870, "ts": 7429770631490.508,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631510.438, "dur": 4.500,
    "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": 1969,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1969, "pid": 870, "tid": 870, "ts": 7429770631510.438,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631531.091, "dur": 3.754,
    "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": 1970,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1970, "pid": 870, "tid": 870, "ts": 7429770631531.091,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631546.404, "dur": 6.965,
    "args": {
      "External id": 5892,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 1971,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1971, "pid": 870, "tid": 870, "ts": 7429770631546.404,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770631593.879, "dur": 0.189,
    "args": {
      "External id": 5904,      "cid": 370, "correlation": 1972
    }
  },
  {
    "ph": "f", "id": 1972, "pid": 870, "tid": 870, "ts": 7429770631593.879,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770631597.104, "dur": 0.143,
    "args": {
      "External id": 5904,      "cid": 370, "correlation": 1973
    }
  },
  {
    "ph": "f", "id": 1973, "pid": 870, "tid": 870, "ts": 7429770631597.104,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770631599.801, "dur": 0.158,
    "args": {
      "External id": 5904,      "cid": 370, "correlation": 1974
    }
  },
  {
    "ph": "f", "id": 1974, "pid": 870, "tid": 870, "ts": 7429770631599.801,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631602.123, "dur": 6.325,
    "args": {
      "External id": 5904,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1975,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1975, "pid": 870, "tid": 870, "ts": 7429770631602.123,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631624.915, "dur": 4.364,
    "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": 1976,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1976, "pid": 870, "tid": 870, "ts": 7429770631624.915,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770631671.222, "dur": 0.233,
    "args": {
      "External id": 5923,      "cid": 370, "correlation": 1977
    }
  },
  {
    "ph": "f", "id": 1977, "pid": 870, "tid": 870, "ts": 7429770631671.222,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770631674.344, "dur": 0.190,
    "args": {
      "External id": 5923,      "cid": 370, "correlation": 1978
    }
  },
  {
    "ph": "f", "id": 1978, "pid": 870, "tid": 870, "ts": 7429770631674.344,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770631676.614, "dur": 0.191,
    "args": {
      "External id": 5923,      "cid": 370, "correlation": 1979
    }
  },
  {
    "ph": "f", "id": 1979, "pid": 870, "tid": 870, "ts": 7429770631676.614,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631678.512, "dur": 5.653,
    "args": {
      "External id": 5923,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1980,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1980, "pid": 870, "tid": 870, "ts": 7429770631678.512,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631696.507, "dur": 4.185,
    "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": 1981,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1981, "pid": 870, "tid": 870, "ts": 7429770631696.507,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631719.000, "dur": 5.368,
    "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": 1982,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1982, "pid": 870, "tid": 870, "ts": 7429770631719.000,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631734.250, "dur": 6.653,
    "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": 1983,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1983, "pid": 870, "tid": 870, "ts": 7429770631734.250,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631748.533, "dur": 3.940,
    "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": 1984,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1984, "pid": 870, "tid": 870, "ts": 7429770631748.533,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631759.868, "dur": 4.141,
    "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": 1985,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1985, "pid": 870, "tid": 870, "ts": 7429770631759.868,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631770.705, "dur": 5.709,
    "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": 1986,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1986, "pid": 870, "tid": 870, "ts": 7429770631770.705,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631784.166, "dur": 3.649,
    "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": 1987,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1987, "pid": 870, "tid": 870, "ts": 7429770631784.166,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770631822.543, "dur": 0.217,
    "args": {
      "External id": 5944,      "cid": 370, "correlation": 1988
    }
  },
  {
    "ph": "f", "id": 1988, "pid": 870, "tid": 870, "ts": 7429770631822.543,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770631825.748, "dur": 0.181,
    "args": {
      "External id": 5944,      "cid": 370, "correlation": 1989
    }
  },
  {
    "ph": "f", "id": 1989, "pid": 870, "tid": 870, "ts": 7429770631825.748,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770631828.424, "dur": 0.150,
    "args": {
      "External id": 5944,      "cid": 370, "correlation": 1990
    }
  },
  {
    "ph": "f", "id": 1990, "pid": 870, "tid": 870, "ts": 7429770631828.424,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631830.359, "dur": 5.105,
    "args": {
      "External id": 5944,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1991,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1991, "pid": 870, "tid": 870, "ts": 7429770631830.359,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631851.780, "dur": 5.535,
    "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": 1992,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1992, "pid": 870, "tid": 870, "ts": 7429770631851.780,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770631882.530, "dur": 0.189,
    "args": {
      "External id": 5955,      "cid": 370, "correlation": 1993
    }
  },
  {
    "ph": "f", "id": 1993, "pid": 870, "tid": 870, "ts": 7429770631882.530,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770631885.323, "dur": 0.142,
    "args": {
      "External id": 5955,      "cid": 370, "correlation": 1994
    }
  },
  {
    "ph": "f", "id": 1994, "pid": 870, "tid": 870, "ts": 7429770631885.323,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770631887.353, "dur": 0.151,
    "args": {
      "External id": 5955,      "cid": 370, "correlation": 1995
    }
  },
  {
    "ph": "f", "id": 1995, "pid": 870, "tid": 870, "ts": 7429770631887.353,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631889.117, "dur": 5.017,
    "args": {
      "External id": 5955,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 1996,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1996, "pid": 870, "tid": 870, "ts": 7429770631889.117,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631902.905, "dur": 4.211,
    "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": 1997,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 1997, "pid": 870, "tid": 870, "ts": 7429770631902.905,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770631929.383, "dur": 0.343,
    "args": {
      "External id": 5967,      "cid": 290, "correlation": 1998
    }
  },
  {
    "ph": "f", "id": 1998, "pid": 870, "tid": 870, "ts": 7429770631929.383,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770631930.159, "dur": 50.379,
    "args": {
      "External id": 5967,            "bytes": 79691776, "cid": 108, "correlation": 1999, "ptr": "0x7ffd8041f0c0"
    }
  },
  {
    "ph": "s", "id": 1999, "pid": 870, "tid": 870, "ts": 7429770631930.159,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770631982.702, "dur": 0.181,
    "args": {
      "External id": 5966,      "cid": 370, "correlation": 2000
    }
  },
  {
    "ph": "f", "id": 2000, "pid": 870, "tid": 870, "ts": 7429770631982.702,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770631986.432, "dur": 0.142,
    "args": {
      "External id": 5966,      "cid": 370, "correlation": 2001
    }
  },
  {
    "ph": "f", "id": 2001, "pid": 870, "tid": 870, "ts": 7429770631986.432,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770631989.250, "dur": 0.143,
    "args": {
      "External id": 5966,      "cid": 370, "correlation": 2002
    }
  },
  {
    "ph": "f", "id": 2002, "pid": 870, "tid": 870, "ts": 7429770631989.250,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770631999.311, "dur": 5.854,
    "args": {
      "External id": 5966,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2003,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2003, "pid": 870, "tid": 870, "ts": 7429770631999.311,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632005.757, "dur": 3.750,
    "args": {
      "External id": 5966,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 2004,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2004, "pid": 870, "tid": 870, "ts": 7429770632005.757,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632021.961, "dur": 4.021,
    "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": 2005,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2005, "pid": 870, "tid": 870, "ts": 7429770632021.961,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632054.554, "dur": 4.937,
    "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": 2006,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2006, "pid": 870, "tid": 870, "ts": 7429770632054.554,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632068.594, "dur": 5.227,
    "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": 2007,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2007, "pid": 870, "tid": 870, "ts": 7429770632068.594,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632082.160, "dur": 4.325,
    "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": 2008,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2008, "pid": 870, "tid": 870, "ts": 7429770632082.160,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632092.551, "dur": 3.927,
    "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": 2009,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2009, "pid": 870, "tid": 870, "ts": 7429770632092.551,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632102.547, "dur": 3.642,
    "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": 2010,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2010, "pid": 870, "tid": 870, "ts": 7429770632102.547,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632113.675, "dur": 3.651,
    "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": 2011,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2011, "pid": 870, "tid": 870, "ts": 7429770632113.675,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770632156.264, "dur": 0.282,
    "args": {
      "External id": 5987,      "cid": 370, "correlation": 2012
    }
  },
  {
    "ph": "f", "id": 2012, "pid": 870, "tid": 870, "ts": 7429770632156.264,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770632159.666, "dur": 0.152,
    "args": {
      "External id": 5987,      "cid": 370, "correlation": 2013
    }
  },
  {
    "ph": "f", "id": 2013, "pid": 870, "tid": 870, "ts": 7429770632159.666,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770632162.144, "dur": 0.147,
    "args": {
      "External id": 5987,      "cid": 370, "correlation": 2014
    }
  },
  {
    "ph": "f", "id": 2014, "pid": 870, "tid": 870, "ts": 7429770632162.144,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632164.273, "dur": 6.153,
    "args": {
      "External id": 5987,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2015,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2015, "pid": 870, "tid": 870, "ts": 7429770632164.273,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770632206.927, "dur": 0.196,
    "args": {
      "External id": 6000,      "cid": 370, "correlation": 2016
    }
  },
  {
    "ph": "f", "id": 2016, "pid": 870, "tid": 870, "ts": 7429770632206.927,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770632210.094, "dur": 0.151,
    "args": {
      "External id": 6000,      "cid": 370, "correlation": 2017
    }
  },
  {
    "ph": "f", "id": 2017, "pid": 870, "tid": 870, "ts": 7429770632210.094,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770632212.283, "dur": 0.149,
    "args": {
      "External id": 6000,      "cid": 370, "correlation": 2018
    }
  },
  {
    "ph": "f", "id": 2018, "pid": 870, "tid": 870, "ts": 7429770632212.283,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632213.890, "dur": 6.790,
    "args": {
      "External id": 6000,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2019,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2019, "pid": 870, "tid": 870, "ts": 7429770632213.890,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770632254.430, "dur": 0.178,
    "args": {
      "External id": 6013,      "cid": 370, "correlation": 2020
    }
  },
  {
    "ph": "f", "id": 2020, "pid": 870, "tid": 870, "ts": 7429770632254.430,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770632257.378, "dur": 0.158,
    "args": {
      "External id": 6013,      "cid": 370, "correlation": 2021
    }
  },
  {
    "ph": "f", "id": 2021, "pid": 870, "tid": 870, "ts": 7429770632257.378,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770632259.252, "dur": 0.139,
    "args": {
      "External id": 6013,      "cid": 370, "correlation": 2022
    }
  },
  {
    "ph": "f", "id": 2022, "pid": 870, "tid": 870, "ts": 7429770632259.252,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632260.719, "dur": 5.101,
    "args": {
      "External id": 6013,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2023,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2023, "pid": 870, "tid": 870, "ts": 7429770632260.719,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632287.982, "dur": 3.926,
    "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": 2024,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2024, "pid": 870, "tid": 870, "ts": 7429770632287.982,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632323.976, "dur": 6.653,
    "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": 2025,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2025, "pid": 870, "tid": 870, "ts": 7429770632323.976,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632346.611, "dur": 7.112,
    "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": 2026,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2026, "pid": 870, "tid": 870, "ts": 7429770632346.611,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632362.302, "dur": 5.225,
    "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": 2027,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2027, "pid": 870, "tid": 870, "ts": 7429770632362.302,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632373.285, "dur": 4.746,
    "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": 2028,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2028, "pid": 870, "tid": 870, "ts": 7429770632373.285,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632384.773, "dur": 3.712,
    "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": 2029,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2029, "pid": 870, "tid": 870, "ts": 7429770632384.773,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632402.100, "dur": 4.246,
    "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": 2030,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2030, "pid": 870, "tid": 870, "ts": 7429770632402.100,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632412.266, "dur": 4.844,
    "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": 2031,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2031, "pid": 870, "tid": 870, "ts": 7429770632412.266,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632423.219, "dur": 4.608,
    "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": 2032,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2032, "pid": 870, "tid": 870, "ts": 7429770632423.219,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632432.221, "dur": 3.917,
    "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": 2033,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2033, "pid": 870, "tid": 870, "ts": 7429770632432.221,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632495.897, "dur": 4.371,
    "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": 2034,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2034, "pid": 870, "tid": 870, "ts": 7429770632495.897,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632528.861, "dur": 4.498,
    "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": 2035,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2035, "pid": 870, "tid": 870, "ts": 7429770632528.861,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770632558.073, "dur": 0.342,
    "args": {
      "External id": 6088,      "cid": 290, "correlation": 2036
    }
  },
  {
    "ph": "f", "id": 2036, "pid": 870, "tid": 870, "ts": 7429770632558.073,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770632558.919, "dur": 47.971,
    "args": {
      "External id": 6088,            "bytes": 2097152, "cid": 108, "correlation": 2037, "ptr": "0x7ffd80420410"
    }
  },
  {
    "ph": "s", "id": 2037, "pid": 870, "tid": 870, "ts": 7429770632558.919,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770632624.397, "dur": 0.526,
    "args": {
      "External id": 6088,      "cid": 370, "correlation": 2038
    }
  },
  {
    "ph": "f", "id": 2038, "pid": 870, "tid": 870, "ts": 7429770632624.397,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770632630.298, "dur": 0.140,
    "args": {
      "External id": 6088,      "cid": 370, "correlation": 2039
    }
  },
  {
    "ph": "f", "id": 2039, "pid": 870, "tid": 870, "ts": 7429770632630.298,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770632633.702, "dur": 0.183,
    "args": {
      "External id": 6088,      "cid": 370, "correlation": 2040
    }
  },
  {
    "ph": "f", "id": 2040, "pid": 870, "tid": 870, "ts": 7429770632633.702,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632638.593, "dur": 6.382,
    "args": {
      "External id": 6088,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2041,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2041, "pid": 870, "tid": 870, "ts": 7429770632638.593,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632660.533, "dur": 4.647,
    "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": 2042,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2042, "pid": 870, "tid": 870, "ts": 7429770632660.533,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632681.285, "dur": 7.085,
    "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": 2043,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2043, "pid": 870, "tid": 870, "ts": 7429770632681.285,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632701.587, "dur": 6.286,
    "args": {
      "External id": 6101,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 2044,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2044, "pid": 870, "tid": 870, "ts": 7429770632701.587,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770632753.943, "dur": 0.250,
    "args": {
      "External id": 6113,      "cid": 370, "correlation": 2045
    }
  },
  {
    "ph": "f", "id": 2045, "pid": 870, "tid": 870, "ts": 7429770632753.943,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770632757.646, "dur": 0.143,
    "args": {
      "External id": 6113,      "cid": 370, "correlation": 2046
    }
  },
  {
    "ph": "f", "id": 2046, "pid": 870, "tid": 870, "ts": 7429770632757.646,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770632760.581, "dur": 0.155,
    "args": {
      "External id": 6113,      "cid": 370, "correlation": 2047
    }
  },
  {
    "ph": "f", "id": 2047, "pid": 870, "tid": 870, "ts": 7429770632760.581,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632763.161, "dur": 6.878,
    "args": {
      "External id": 6113,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2048,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2048, "pid": 870, "tid": 870, "ts": 7429770632763.161,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632787.492, "dur": 4.180,
    "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": 2049,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2049, "pid": 870, "tid": 870, "ts": 7429770632787.492,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770632846.619, "dur": 0.245,
    "args": {
      "External id": 6132,      "cid": 370, "correlation": 2050
    }
  },
  {
    "ph": "f", "id": 2050, "pid": 870, "tid": 870, "ts": 7429770632846.619,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770632849.686, "dur": 0.212,
    "args": {
      "External id": 6132,      "cid": 370, "correlation": 2051
    }
  },
  {
    "ph": "f", "id": 2051, "pid": 870, "tid": 870, "ts": 7429770632849.686,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770632851.944, "dur": 0.199,
    "args": {
      "External id": 6132,      "cid": 370, "correlation": 2052
    }
  },
  {
    "ph": "f", "id": 2052, "pid": 870, "tid": 870, "ts": 7429770632851.944,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632854.103, "dur": 5.065,
    "args": {
      "External id": 6132,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2053,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2053, "pid": 870, "tid": 870, "ts": 7429770632854.103,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632872.797, "dur": 4.290,
    "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": 2054,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2054, "pid": 870, "tid": 870, "ts": 7429770632872.797,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632897.263, "dur": 5.551,
    "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": 2055,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2055, "pid": 870, "tid": 870, "ts": 7429770632897.263,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632914.372, "dur": 5.511,
    "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": 2056,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2056, "pid": 870, "tid": 870, "ts": 7429770632914.372,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632928.563, "dur": 4.632,
    "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": 2057,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2057, "pid": 870, "tid": 870, "ts": 7429770632928.563,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632939.901, "dur": 4.060,
    "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": 2058,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2058, "pid": 870, "tid": 870, "ts": 7429770632939.901,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632950.808, "dur": 5.183,
    "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": 2059,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2059, "pid": 870, "tid": 870, "ts": 7429770632950.808,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770632987.579, "dur": 4.027,
    "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": 2060,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2060, "pid": 870, "tid": 870, "ts": 7429770632987.579,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770633026.467, "dur": 0.173,
    "args": {
      "External id": 6153,      "cid": 370, "correlation": 2061
    }
  },
  {
    "ph": "f", "id": 2061, "pid": 870, "tid": 870, "ts": 7429770633026.467,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770633030.195, "dur": 0.147,
    "args": {
      "External id": 6153,      "cid": 370, "correlation": 2062
    }
  },
  {
    "ph": "f", "id": 2062, "pid": 870, "tid": 870, "ts": 7429770633030.195,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770633033.117, "dur": 0.163,
    "args": {
      "External id": 6153,      "cid": 370, "correlation": 2063
    }
  },
  {
    "ph": "f", "id": 2063, "pid": 870, "tid": 870, "ts": 7429770633033.117,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633035.398, "dur": 5.129,
    "args": {
      "External id": 6153,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2064,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2064, "pid": 870, "tid": 870, "ts": 7429770633035.398,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633057.539, "dur": 5.271,
    "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": 2065,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2065, "pid": 870, "tid": 870, "ts": 7429770633057.539,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770633088.314, "dur": 0.150,
    "args": {
      "External id": 6164,      "cid": 370, "correlation": 2066
    }
  },
  {
    "ph": "f", "id": 2066, "pid": 870, "tid": 870, "ts": 7429770633088.314,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770633091.017, "dur": 0.139,
    "args": {
      "External id": 6164,      "cid": 370, "correlation": 2067
    }
  },
  {
    "ph": "f", "id": 2067, "pid": 870, "tid": 870, "ts": 7429770633091.017,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770633093.211, "dur": 0.140,
    "args": {
      "External id": 6164,      "cid": 370, "correlation": 2068
    }
  },
  {
    "ph": "f", "id": 2068, "pid": 870, "tid": 870, "ts": 7429770633093.211,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633094.801, "dur": 5.060,
    "args": {
      "External id": 6164,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2069,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2069, "pid": 870, "tid": 870, "ts": 7429770633094.801,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633108.464, "dur": 4.310,
    "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": 2070,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2070, "pid": 870, "tid": 870, "ts": 7429770633108.464,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770633135.009, "dur": 0.181,
    "args": {
      "External id": 6175,      "cid": 370, "correlation": 2071
    }
  },
  {
    "ph": "f", "id": 2071, "pid": 870, "tid": 870, "ts": 7429770633135.009,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770633138.009, "dur": 0.137,
    "args": {
      "External id": 6175,      "cid": 370, "correlation": 2072
    }
  },
  {
    "ph": "f", "id": 2072, "pid": 870, "tid": 870, "ts": 7429770633138.009,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770633140.345, "dur": 0.139,
    "args": {
      "External id": 6175,      "cid": 370, "correlation": 2073
    }
  },
  {
    "ph": "f", "id": 2073, "pid": 870, "tid": 870, "ts": 7429770633140.345,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633149.883, "dur": 4.818,
    "args": {
      "External id": 6175,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2074,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2074, "pid": 870, "tid": 870, "ts": 7429770633149.883,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633155.061, "dur": 4.789,
    "args": {
      "External id": 6175,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 2075,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2075, "pid": 870, "tid": 870, "ts": 7429770633155.061,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633169.735, "dur": 5.177,
    "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": 2076,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2076, "pid": 870, "tid": 870, "ts": 7429770633169.735,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633201.868, "dur": 4.345,
    "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": 2077,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2077, "pid": 870, "tid": 870, "ts": 7429770633201.868,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633214.918, "dur": 4.675,
    "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": 2078,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2078, "pid": 870, "tid": 870, "ts": 7429770633214.918,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633226.666, "dur": 3.526,
    "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": 2079,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2079, "pid": 870, "tid": 870, "ts": 7429770633226.666,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633235.912, "dur": 3.391,
    "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": 2080,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2080, "pid": 870, "tid": 870, "ts": 7429770633235.912,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633244.971, "dur": 3.510,
    "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": 2081,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2081, "pid": 870, "tid": 870, "ts": 7429770633244.971,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633256.264, "dur": 3.611,
    "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": 2082,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2082, "pid": 870, "tid": 870, "ts": 7429770633256.264,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770633296.603, "dur": 0.298,
    "args": {
      "External id": 6196,      "cid": 370, "correlation": 2083
    }
  },
  {
    "ph": "f", "id": 2083, "pid": 870, "tid": 870, "ts": 7429770633296.603,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770633300.069, "dur": 0.142,
    "args": {
      "External id": 6196,      "cid": 370, "correlation": 2084
    }
  },
  {
    "ph": "f", "id": 2084, "pid": 870, "tid": 870, "ts": 7429770633300.069,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770633302.361, "dur": 0.150,
    "args": {
      "External id": 6196,      "cid": 370, "correlation": 2085
    }
  },
  {
    "ph": "f", "id": 2085, "pid": 870, "tid": 870, "ts": 7429770633302.361,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633304.437, "dur": 5.284,
    "args": {
      "External id": 6196,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2086,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2086, "pid": 870, "tid": 870, "ts": 7429770633304.437,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770633344.645, "dur": 0.155,
    "args": {
      "External id": 6209,      "cid": 370, "correlation": 2087
    }
  },
  {
    "ph": "f", "id": 2087, "pid": 870, "tid": 870, "ts": 7429770633344.645,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770633347.957, "dur": 0.157,
    "args": {
      "External id": 6209,      "cid": 370, "correlation": 2088
    }
  },
  {
    "ph": "f", "id": 2088, "pid": 870, "tid": 870, "ts": 7429770633347.957,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770633350.023, "dur": 0.151,
    "args": {
      "External id": 6209,      "cid": 370, "correlation": 2089
    }
  },
  {
    "ph": "f", "id": 2089, "pid": 870, "tid": 870, "ts": 7429770633350.023,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633351.628, "dur": 6.468,
    "args": {
      "External id": 6209,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2090,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2090, "pid": 870, "tid": 870, "ts": 7429770633351.628,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770633389.839, "dur": 0.175,
    "args": {
      "External id": 6222,      "cid": 370, "correlation": 2091
    }
  },
  {
    "ph": "f", "id": 2091, "pid": 870, "tid": 870, "ts": 7429770633389.839,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770633392.578, "dur": 0.156,
    "args": {
      "External id": 6222,      "cid": 370, "correlation": 2092
    }
  },
  {
    "ph": "f", "id": 2092, "pid": 870, "tid": 870, "ts": 7429770633392.578,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770633394.520, "dur": 0.154,
    "args": {
      "External id": 6222,      "cid": 370, "correlation": 2093
    }
  },
  {
    "ph": "f", "id": 2093, "pid": 870, "tid": 870, "ts": 7429770633394.520,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633396.017, "dur": 4.670,
    "args": {
      "External id": 6222,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2094,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2094, "pid": 870, "tid": 870, "ts": 7429770633396.017,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633422.490, "dur": 4.036,
    "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": 2095,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2095, "pid": 870, "tid": 870, "ts": 7429770633422.490,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633446.426, "dur": 5.874,
    "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": 2096,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2096, "pid": 870, "tid": 870, "ts": 7429770633446.426,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633466.080, "dur": 7.052,
    "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": 2097,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2097, "pid": 870, "tid": 870, "ts": 7429770633466.080,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633479.996, "dur": 4.419,
    "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": 2098,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2098, "pid": 870, "tid": 870, "ts": 7429770633479.996,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633489.890, "dur": 7.907,
    "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": 2099,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2099, "pid": 870, "tid": 870, "ts": 7429770633489.890,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633504.320, "dur": 4.788,
    "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": 2100,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2100, "pid": 870, "tid": 870, "ts": 7429770633504.320,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633522.021, "dur": 3.761,
    "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": 2101,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2101, "pid": 870, "tid": 870, "ts": 7429770633522.021,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633532.393, "dur": 5.179,
    "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": 2102,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2102, "pid": 870, "tid": 870, "ts": 7429770633532.393,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633543.334, "dur": 4.843,
    "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": 2103,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2103, "pid": 870, "tid": 870, "ts": 7429770633543.334,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633553.093, "dur": 4.660,
    "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": 2104,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2104, "pid": 870, "tid": 870, "ts": 7429770633553.093,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633612.118, "dur": 5.602,
    "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": 2105,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2105, "pid": 870, "tid": 870, "ts": 7429770633612.118,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633645.065, "dur": 4.582,
    "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": 2106,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2106, "pid": 870, "tid": 870, "ts": 7429770633645.065,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770633672.562, "dur": 0.441,
    "args": {
      "External id": 6297,      "cid": 290, "correlation": 2107
    }
  },
  {
    "ph": "f", "id": 2107, "pid": 870, "tid": 870, "ts": 7429770633672.562,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770633673.625, "dur": 53.975,
    "args": {
      "External id": 6297,            "bytes": 2097152, "cid": 108, "correlation": 2108, "ptr": "0x7ffd80420410"
    }
  },
  {
    "ph": "s", "id": 2108, "pid": 870, "tid": 870, "ts": 7429770633673.625,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770633743.320, "dur": 0.308,
    "args": {
      "External id": 6297,      "cid": 370, "correlation": 2109
    }
  },
  {
    "ph": "f", "id": 2109, "pid": 870, "tid": 870, "ts": 7429770633743.320,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770633748.359, "dur": 0.155,
    "args": {
      "External id": 6297,      "cid": 370, "correlation": 2110
    }
  },
  {
    "ph": "f", "id": 2110, "pid": 870, "tid": 870, "ts": 7429770633748.359,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770633751.389, "dur": 0.183,
    "args": {
      "External id": 6297,      "cid": 370, "correlation": 2111
    }
  },
  {
    "ph": "f", "id": 2111, "pid": 870, "tid": 870, "ts": 7429770633751.389,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633755.065, "dur": 4.907,
    "args": {
      "External id": 6297,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2112,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2112, "pid": 870, "tid": 870, "ts": 7429770633755.065,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633774.348, "dur": 4.142,
    "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": 2113,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2113, "pid": 870, "tid": 870, "ts": 7429770633774.348,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633794.245, "dur": 3.876,
    "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": 2114,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2114, "pid": 870, "tid": 870, "ts": 7429770633794.245,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633810.570, "dur": 5.712,
    "args": {
      "External id": 6310,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 2115,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2115, "pid": 870, "tid": 870, "ts": 7429770633810.570,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770633858.931, "dur": 0.182,
    "args": {
      "External id": 6322,      "cid": 370, "correlation": 2116
    }
  },
  {
    "ph": "f", "id": 2116, "pid": 870, "tid": 870, "ts": 7429770633858.931,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770633862.460, "dur": 0.151,
    "args": {
      "External id": 6322,      "cid": 370, "correlation": 2117
    }
  },
  {
    "ph": "f", "id": 2117, "pid": 870, "tid": 870, "ts": 7429770633862.460,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770633865.101, "dur": 0.147,
    "args": {
      "External id": 6322,      "cid": 370, "correlation": 2118
    }
  },
  {
    "ph": "f", "id": 2118, "pid": 870, "tid": 870, "ts": 7429770633865.101,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633867.478, "dur": 6.037,
    "args": {
      "External id": 6322,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2119,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2119, "pid": 870, "tid": 870, "ts": 7429770633867.478,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633890.783, "dur": 3.819,
    "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": 2120,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2120, "pid": 870, "tid": 870, "ts": 7429770633890.783,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770633939.462, "dur": 0.225,
    "args": {
      "External id": 6341,      "cid": 370, "correlation": 2121
    }
  },
  {
    "ph": "f", "id": 2121, "pid": 870, "tid": 870, "ts": 7429770633939.462,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770633942.507, "dur": 0.179,
    "args": {
      "External id": 6341,      "cid": 370, "correlation": 2122
    }
  },
  {
    "ph": "f", "id": 2122, "pid": 870, "tid": 870, "ts": 7429770633942.507,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770633944.630, "dur": 0.188,
    "args": {
      "External id": 6341,      "cid": 370, "correlation": 2123
    }
  },
  {
    "ph": "f", "id": 2123, "pid": 870, "tid": 870, "ts": 7429770633944.630,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633946.625, "dur": 4.936,
    "args": {
      "External id": 6341,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2124,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2124, "pid": 870, "tid": 870, "ts": 7429770633946.625,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633964.348, "dur": 3.952,
    "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": 2125,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2125, "pid": 870, "tid": 870, "ts": 7429770633964.348,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770633986.228, "dur": 5.982,
    "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": 2126,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2126, "pid": 870, "tid": 870, "ts": 7429770633986.228,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634002.033, "dur": 5.716,
    "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": 2127,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2127, "pid": 870, "tid": 870, "ts": 7429770634002.033,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634015.399, "dur": 5.478,
    "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": 2128,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2128, "pid": 870, "tid": 870, "ts": 7429770634015.399,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634026.707, "dur": 5.383,
    "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": 2129,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2129, "pid": 870, "tid": 870, "ts": 7429770634026.707,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634038.667, "dur": 4.888,
    "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": 2130,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2130, "pid": 870, "tid": 870, "ts": 7429770634038.667,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634051.379, "dur": 4.271,
    "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": 2131,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2131, "pid": 870, "tid": 870, "ts": 7429770634051.379,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770634112.514, "dur": 0.304,
    "args": {
      "External id": 6362,      "cid": 370, "correlation": 2132
    }
  },
  {
    "ph": "f", "id": 2132, "pid": 870, "tid": 870, "ts": 7429770634112.514,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770634116.221, "dur": 0.197,
    "args": {
      "External id": 6362,      "cid": 370, "correlation": 2133
    }
  },
  {
    "ph": "f", "id": 2133, "pid": 870, "tid": 870, "ts": 7429770634116.221,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770634119.064, "dur": 0.207,
    "args": {
      "External id": 6362,      "cid": 370, "correlation": 2134
    }
  },
  {
    "ph": "f", "id": 2134, "pid": 870, "tid": 870, "ts": 7429770634119.064,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634121.803, "dur": 5.336,
    "args": {
      "External id": 6362,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2135,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2135, "pid": 870, "tid": 870, "ts": 7429770634121.803,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634146.719, "dur": 4.866,
    "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": 2136,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2136, "pid": 870, "tid": 870, "ts": 7429770634146.719,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770634177.755, "dur": 0.185,
    "args": {
      "External id": 6373,      "cid": 370, "correlation": 2137
    }
  },
  {
    "ph": "f", "id": 2137, "pid": 870, "tid": 870, "ts": 7429770634177.755,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770634180.376, "dur": 0.180,
    "args": {
      "External id": 6373,      "cid": 370, "correlation": 2138
    }
  },
  {
    "ph": "f", "id": 2138, "pid": 870, "tid": 870, "ts": 7429770634180.376,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770634182.517, "dur": 0.140,
    "args": {
      "External id": 6373,      "cid": 370, "correlation": 2139
    }
  },
  {
    "ph": "f", "id": 2139, "pid": 870, "tid": 870, "ts": 7429770634182.517,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634184.323, "dur": 5.551,
    "args": {
      "External id": 6373,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2140,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2140, "pid": 870, "tid": 870, "ts": 7429770634184.323,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634199.200, "dur": 4.240,
    "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": 2141,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2141, "pid": 870, "tid": 870, "ts": 7429770634199.200,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770634225.672, "dur": 0.158,
    "args": {
      "External id": 6384,      "cid": 370, "correlation": 2142
    }
  },
  {
    "ph": "f", "id": 2142, "pid": 870, "tid": 870, "ts": 7429770634225.672,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770634228.389, "dur": 0.138,
    "args": {
      "External id": 6384,      "cid": 370, "correlation": 2143
    }
  },
  {
    "ph": "f", "id": 2143, "pid": 870, "tid": 870, "ts": 7429770634228.389,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770634230.898, "dur": 0.141,
    "args": {
      "External id": 6384,      "cid": 370, "correlation": 2144
    }
  },
  {
    "ph": "f", "id": 2144, "pid": 870, "tid": 870, "ts": 7429770634230.898,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634239.768, "dur": 7.134,
    "args": {
      "External id": 6384,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2145,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2145, "pid": 870, "tid": 870, "ts": 7429770634239.768,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634247.270, "dur": 4.201,
    "args": {
      "External id": 6384,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 2146,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2146, "pid": 870, "tid": 870, "ts": 7429770634247.270,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634260.969, "dur": 5.991,
    "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": 2147,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2147, "pid": 870, "tid": 870, "ts": 7429770634260.969,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634293.600, "dur": 4.555,
    "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": 2148,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2148, "pid": 870, "tid": 870, "ts": 7429770634293.600,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634306.867, "dur": 5.748,
    "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": 2149,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2149, "pid": 870, "tid": 870, "ts": 7429770634306.867,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634320.989, "dur": 3.778,
    "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": 2150,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2150, "pid": 870, "tid": 870, "ts": 7429770634320.989,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634329.852, "dur": 4.052,
    "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": 2151,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2151, "pid": 870, "tid": 870, "ts": 7429770634329.852,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634340.611, "dur": 3.987,
    "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": 2152,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2152, "pid": 870, "tid": 870, "ts": 7429770634340.611,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634351.962, "dur": 4.272,
    "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": 2153,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2153, "pid": 870, "tid": 870, "ts": 7429770634351.962,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770634394.034, "dur": 0.209,
    "args": {
      "External id": 6405,      "cid": 370, "correlation": 2154
    }
  },
  {
    "ph": "f", "id": 2154, "pid": 870, "tid": 870, "ts": 7429770634394.034,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770634397.212, "dur": 0.152,
    "args": {
      "External id": 6405,      "cid": 370, "correlation": 2155
    }
  },
  {
    "ph": "f", "id": 2155, "pid": 870, "tid": 870, "ts": 7429770634397.212,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770634399.352, "dur": 0.171,
    "args": {
      "External id": 6405,      "cid": 370, "correlation": 2156
    }
  },
  {
    "ph": "f", "id": 2156, "pid": 870, "tid": 870, "ts": 7429770634399.352,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634401.381, "dur": 5.689,
    "args": {
      "External id": 6405,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2157,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2157, "pid": 870, "tid": 870, "ts": 7429770634401.381,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770634440.807, "dur": 0.160,
    "args": {
      "External id": 6418,      "cid": 370, "correlation": 2158
    }
  },
  {
    "ph": "f", "id": 2158, "pid": 870, "tid": 870, "ts": 7429770634440.807,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770634443.517, "dur": 0.155,
    "args": {
      "External id": 6418,      "cid": 370, "correlation": 2159
    }
  },
  {
    "ph": "f", "id": 2159, "pid": 870, "tid": 870, "ts": 7429770634443.517,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770634445.550, "dur": 0.150,
    "args": {
      "External id": 6418,      "cid": 370, "correlation": 2160
    }
  },
  {
    "ph": "f", "id": 2160, "pid": 870, "tid": 870, "ts": 7429770634445.550,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634447.105, "dur": 6.262,
    "args": {
      "External id": 6418,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2161,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2161, "pid": 870, "tid": 870, "ts": 7429770634447.105,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770634484.278, "dur": 0.169,
    "args": {
      "External id": 6431,      "cid": 370, "correlation": 2162
    }
  },
  {
    "ph": "f", "id": 2162, "pid": 870, "tid": 870, "ts": 7429770634484.278,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770634486.977, "dur": 0.158,
    "args": {
      "External id": 6431,      "cid": 370, "correlation": 2163
    }
  },
  {
    "ph": "f", "id": 2163, "pid": 870, "tid": 870, "ts": 7429770634486.977,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770634488.852, "dur": 0.151,
    "args": {
      "External id": 6431,      "cid": 370, "correlation": 2164
    }
  },
  {
    "ph": "f", "id": 2164, "pid": 870, "tid": 870, "ts": 7429770634488.852,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634490.281, "dur": 5.285,
    "args": {
      "External id": 6431,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2165,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2165, "pid": 870, "tid": 870, "ts": 7429770634490.281,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634518.148, "dur": 3.888,
    "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": 2166,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2166, "pid": 870, "tid": 870, "ts": 7429770634518.148,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634542.262, "dur": 4.996,
    "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": 2167,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2167, "pid": 870, "tid": 870, "ts": 7429770634542.262,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634562.024, "dur": 7.235,
    "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": 2168,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2168, "pid": 870, "tid": 870, "ts": 7429770634562.024,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634576.245, "dur": 4.038,
    "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": 2169,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2169, "pid": 870, "tid": 870, "ts": 7429770634576.245,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634585.784, "dur": 5.351,
    "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": 2170,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2170, "pid": 870, "tid": 870, "ts": 7429770634585.784,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634596.648, "dur": 3.546,
    "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": 2171,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2171, "pid": 870, "tid": 870, "ts": 7429770634596.648,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634613.062, "dur": 3.337,
    "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": 2172,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2172, "pid": 870, "tid": 870, "ts": 7429770634613.062,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634621.927, "dur": 4.919,
    "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": 2173,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2173, "pid": 870, "tid": 870, "ts": 7429770634621.927,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634633.447, "dur": 3.326,
    "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": 2174,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2174, "pid": 870, "tid": 870, "ts": 7429770634633.447,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634641.013, "dur": 3.594,
    "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": 2175,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2175, "pid": 870, "tid": 870, "ts": 7429770634641.013,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634700.628, "dur": 5.995,
    "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": 2176,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2176, "pid": 870, "tid": 870, "ts": 7429770634700.628,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634733.710, "dur": 3.474,
    "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": 2177,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2177, "pid": 870, "tid": 870, "ts": 7429770634733.710,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770634761.232, "dur": 0.498,
    "args": {
      "External id": 6506,      "cid": 290, "correlation": 2178
    }
  },
  {
    "ph": "f", "id": 2178, "pid": 870, "tid": 870, "ts": 7429770634761.232,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770634762.404, "dur": 56.761,
    "args": {
      "External id": 6506,            "bytes": 2097152, "cid": 108, "correlation": 2179, "ptr": "0x7ffd80420410"
    }
  },
  {
    "ph": "s", "id": 2179, "pid": 870, "tid": 870, "ts": 7429770634762.404,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770634833.274, "dur": 0.317,
    "args": {
      "External id": 6506,      "cid": 370, "correlation": 2180
    }
  },
  {
    "ph": "f", "id": 2180, "pid": 870, "tid": 870, "ts": 7429770634833.274,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770634837.792, "dur": 0.154,
    "args": {
      "External id": 6506,      "cid": 370, "correlation": 2181
    }
  },
  {
    "ph": "f", "id": 2181, "pid": 870, "tid": 870, "ts": 7429770634837.792,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770634840.774, "dur": 0.158,
    "args": {
      "External id": 6506,      "cid": 370, "correlation": 2182
    }
  },
  {
    "ph": "f", "id": 2182, "pid": 870, "tid": 870, "ts": 7429770634840.774,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634844.504, "dur": 5.825,
    "args": {
      "External id": 6506,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2183,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2183, "pid": 870, "tid": 870, "ts": 7429770634844.504,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634865.151, "dur": 4.293,
    "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": 2184,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2184, "pid": 870, "tid": 870, "ts": 7429770634865.151,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634885.603, "dur": 3.574,
    "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": 2185,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2185, "pid": 870, "tid": 870, "ts": 7429770634885.603,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634902.510, "dur": 5.747,
    "args": {
      "External id": 6519,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 2186,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2186, "pid": 870, "tid": 870, "ts": 7429770634902.510,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770634951.832, "dur": 0.176,
    "args": {
      "External id": 6531,      "cid": 370, "correlation": 2187
    }
  },
  {
    "ph": "f", "id": 2187, "pid": 870, "tid": 870, "ts": 7429770634951.832,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770634955.488, "dur": 0.137,
    "args": {
      "External id": 6531,      "cid": 370, "correlation": 2188
    }
  },
  {
    "ph": "f", "id": 2188, "pid": 870, "tid": 870, "ts": 7429770634955.488,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770634958.295, "dur": 0.141,
    "args": {
      "External id": 6531,      "cid": 370, "correlation": 2189
    }
  },
  {
    "ph": "f", "id": 2189, "pid": 870, "tid": 870, "ts": 7429770634958.295,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634960.668, "dur": 6.341,
    "args": {
      "External id": 6531,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2190,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2190, "pid": 870, "tid": 870, "ts": 7429770634960.668,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770634985.040, "dur": 4.094,
    "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": 2191,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2191, "pid": 870, "tid": 870, "ts": 7429770634985.040,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770635035.237, "dur": 0.211,
    "args": {
      "External id": 6550,      "cid": 370, "correlation": 2192
    }
  },
  {
    "ph": "f", "id": 2192, "pid": 870, "tid": 870, "ts": 7429770635035.237,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770635038.561, "dur": 0.188,
    "args": {
      "External id": 6550,      "cid": 370, "correlation": 2193
    }
  },
  {
    "ph": "f", "id": 2193, "pid": 870, "tid": 870, "ts": 7429770635038.561,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770635040.625, "dur": 0.186,
    "args": {
      "External id": 6550,      "cid": 370, "correlation": 2194
    }
  },
  {
    "ph": "f", "id": 2194, "pid": 870, "tid": 870, "ts": 7429770635040.625,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635042.579, "dur": 7.594,
    "args": {
      "External id": 6550,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2195,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2195, "pid": 870, "tid": 870, "ts": 7429770635042.579,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635063.400, "dur": 4.782,
    "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": 2196,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2196, "pid": 870, "tid": 870, "ts": 7429770635063.400,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635086.995, "dur": 5.932,
    "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": 2197,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2197, "pid": 870, "tid": 870, "ts": 7429770635086.995,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635103.620, "dur": 4.931,
    "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": 2198,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2198, "pid": 870, "tid": 870, "ts": 7429770635103.620,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635117.219, "dur": 5.041,
    "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": 2199,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2199, "pid": 870, "tid": 870, "ts": 7429770635117.219,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635128.790, "dur": 4.242,
    "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": 2200,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2200, "pid": 870, "tid": 870, "ts": 7429770635128.790,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635140.277, "dur": 5.272,
    "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": 2201,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2201, "pid": 870, "tid": 870, "ts": 7429770635140.277,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635153.145, "dur": 4.628,
    "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": 2202,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2202, "pid": 870, "tid": 870, "ts": 7429770635153.145,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770635193.781, "dur": 0.304,
    "args": {
      "External id": 6572,      "cid": 290, "correlation": 2203
    }
  },
  {
    "ph": "f", "id": 2203, "pid": 870, "tid": 870, "ts": 7429770635193.781,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770635194.488, "dur": 89.375,
    "args": {
      "External id": 6572,            "bytes": 79691776, "cid": 108, "correlation": 2204, "ptr": "0x7ffd8041f0a8"
    }
  },
  {
    "ph": "s", "id": 2204, "pid": 870, "tid": 870, "ts": 7429770635194.488,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770635287.501, "dur": 0.248,
    "args": {
      "External id": 6571,      "cid": 370, "correlation": 2205
    }
  },
  {
    "ph": "f", "id": 2205, "pid": 870, "tid": 870, "ts": 7429770635287.501,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770635291.854, "dur": 0.143,
    "args": {
      "External id": 6571,      "cid": 370, "correlation": 2206
    }
  },
  {
    "ph": "f", "id": 2206, "pid": 870, "tid": 870, "ts": 7429770635291.854,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770635295.149, "dur": 0.206,
    "args": {
      "External id": 6571,      "cid": 370, "correlation": 2207
    }
  },
  {
    "ph": "f", "id": 2207, "pid": 870, "tid": 870, "ts": 7429770635295.149,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635298.488, "dur": 6.138,
    "args": {
      "External id": 6571,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2208,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2208, "pid": 870, "tid": 870, "ts": 7429770635298.488,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635326.109, "dur": 4.611,
    "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": 2209,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2209, "pid": 870, "tid": 870, "ts": 7429770635326.109,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770635358.758, "dur": 0.149,
    "args": {
      "External id": 6582,      "cid": 370, "correlation": 2210
    }
  },
  {
    "ph": "f", "id": 2210, "pid": 870, "tid": 870, "ts": 7429770635358.758,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770635361.794, "dur": 0.147,
    "args": {
      "External id": 6582,      "cid": 370, "correlation": 2211
    }
  },
  {
    "ph": "f", "id": 2211, "pid": 870, "tid": 870, "ts": 7429770635361.794,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770635364.042, "dur": 0.139,
    "args": {
      "External id": 6582,      "cid": 370, "correlation": 2212
    }
  },
  {
    "ph": "f", "id": 2212, "pid": 870, "tid": 870, "ts": 7429770635364.042,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635365.901, "dur": 6.363,
    "args": {
      "External id": 6582,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2213,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2213, "pid": 870, "tid": 870, "ts": 7429770635365.901,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635382.130, "dur": 4.459,
    "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": 2214,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2214, "pid": 870, "tid": 870, "ts": 7429770635382.130,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770635408.909, "dur": 0.163,
    "args": {
      "External id": 6593,      "cid": 370, "correlation": 2215
    }
  },
  {
    "ph": "f", "id": 2215, "pid": 870, "tid": 870, "ts": 7429770635408.909,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770635412.271, "dur": 0.142,
    "args": {
      "External id": 6593,      "cid": 370, "correlation": 2216
    }
  },
  {
    "ph": "f", "id": 2216, "pid": 870, "tid": 870, "ts": 7429770635412.271,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770635414.621, "dur": 0.139,
    "args": {
      "External id": 6593,      "cid": 370, "correlation": 2217
    }
  },
  {
    "ph": "f", "id": 2217, "pid": 870, "tid": 870, "ts": 7429770635414.621,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635423.575, "dur": 6.303,
    "args": {
      "External id": 6593,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2218,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2218, "pid": 870, "tid": 870, "ts": 7429770635423.575,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635430.240, "dur": 4.216,
    "args": {
      "External id": 6593,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 2219,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2219, "pid": 870, "tid": 870, "ts": 7429770635430.240,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635444.279, "dur": 3.923,
    "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": 2220,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2220, "pid": 870, "tid": 870, "ts": 7429770635444.279,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635476.358, "dur": 4.154,
    "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": 2221,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2221, "pid": 870, "tid": 870, "ts": 7429770635476.358,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635490.142, "dur": 5.910,
    "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": 2222,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2222, "pid": 870, "tid": 870, "ts": 7429770635490.142,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635504.167, "dur": 3.646,
    "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": 2223,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2223, "pid": 870, "tid": 870, "ts": 7429770635504.167,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635513.205, "dur": 4.201,
    "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": 2224,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2224, "pid": 870, "tid": 870, "ts": 7429770635513.205,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635524.120, "dur": 3.693,
    "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": 2225,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2225, "pid": 870, "tid": 870, "ts": 7429770635524.120,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635535.161, "dur": 4.332,
    "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": 2226,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2226, "pid": 870, "tid": 870, "ts": 7429770635535.161,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770635576.912, "dur": 0.240,
    "args": {
      "External id": 6614,      "cid": 370, "correlation": 2227
    }
  },
  {
    "ph": "f", "id": 2227, "pid": 870, "tid": 870, "ts": 7429770635576.912,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770635580.616, "dur": 0.151,
    "args": {
      "External id": 6614,      "cid": 370, "correlation": 2228
    }
  },
  {
    "ph": "f", "id": 2228, "pid": 870, "tid": 870, "ts": 7429770635580.616,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770635582.773, "dur": 0.154,
    "args": {
      "External id": 6614,      "cid": 370, "correlation": 2229
    }
  },
  {
    "ph": "f", "id": 2229, "pid": 870, "tid": 870, "ts": 7429770635582.773,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635584.843, "dur": 6.684,
    "args": {
      "External id": 6614,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2230,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2230, "pid": 870, "tid": 870, "ts": 7429770635584.843,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770635626.213, "dur": 0.270,
    "args": {
      "External id": 6627,      "cid": 370, "correlation": 2231
    }
  },
  {
    "ph": "f", "id": 2231, "pid": 870, "tid": 870, "ts": 7429770635626.213,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770635629.562, "dur": 0.157,
    "args": {
      "External id": 6627,      "cid": 370, "correlation": 2232
    }
  },
  {
    "ph": "f", "id": 2232, "pid": 870, "tid": 870, "ts": 7429770635629.562,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770635631.572, "dur": 0.152,
    "args": {
      "External id": 6627,      "cid": 370, "correlation": 2233
    }
  },
  {
    "ph": "f", "id": 2233, "pid": 870, "tid": 870, "ts": 7429770635631.572,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635633.170, "dur": 6.309,
    "args": {
      "External id": 6627,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2234,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2234, "pid": 870, "tid": 870, "ts": 7429770635633.170,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770635671.937, "dur": 0.189,
    "args": {
      "External id": 6640,      "cid": 370, "correlation": 2235
    }
  },
  {
    "ph": "f", "id": 2235, "pid": 870, "tid": 870, "ts": 7429770635671.937,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770635675.099, "dur": 0.159,
    "args": {
      "External id": 6640,      "cid": 370, "correlation": 2236
    }
  },
  {
    "ph": "f", "id": 2236, "pid": 870, "tid": 870, "ts": 7429770635675.099,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770635676.964, "dur": 0.154,
    "args": {
      "External id": 6640,      "cid": 370, "correlation": 2237
    }
  },
  {
    "ph": "f", "id": 2237, "pid": 870, "tid": 870, "ts": 7429770635676.964,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635678.419, "dur": 5.549,
    "args": {
      "External id": 6640,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2238,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2238, "pid": 870, "tid": 870, "ts": 7429770635678.419,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635707.654, "dur": 4.483,
    "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": 2239,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2239, "pid": 870, "tid": 870, "ts": 7429770635707.654,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635735.532, "dur": 4.499,
    "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": 2240,      "grid": [672, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2240, "pid": 870, "tid": 870, "ts": 7429770635735.532,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635754.287, "dur": 9.552,
    "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": 2241,      "grid": [84, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2241, "pid": 870, "tid": 870, "ts": 7429770635754.287,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635796.365, "dur": 5.637,
    "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": 2242,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2242, "pid": 870, "tid": 870, "ts": 7429770635796.365,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635807.582, "dur": 4.907,
    "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": 2243,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2243, "pid": 870, "tid": 870, "ts": 7429770635807.582,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635818.472, "dur": 3.952,
    "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": 2244,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2244, "pid": 870, "tid": 870, "ts": 7429770635818.472,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635835.119, "dur": 3.668,
    "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": 2245,      "grid": [168, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2245, "pid": 870, "tid": 870, "ts": 7429770635835.119,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635845.160, "dur": 4.508,
    "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": 2246,      "grid": [21, 2, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2246, "pid": 870, "tid": 870, "ts": 7429770635845.160,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635855.602, "dur": 3.445,
    "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": 2247,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2247, "pid": 870, "tid": 870, "ts": 7429770635855.602,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635864.078, "dur": 3.490,
    "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": 2248,      "grid": [336, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2248, "pid": 870, "tid": 870, "ts": 7429770635864.078,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635927.624, "dur": 4.728,
    "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": 2249,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2249, "pid": 870, "tid": 870, "ts": 7429770635927.624,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770635960.247, "dur": 4.434,
    "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": 2250,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2250, "pid": 870, "tid": 870, "ts": 7429770635960.247,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770635988.466, "dur": 0.358,
    "args": {
      "External id": 6715,      "cid": 290, "correlation": 2251
    }
  },
  {
    "ph": "f", "id": 2251, "pid": 870, "tid": 870, "ts": 7429770635988.466,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770635989.279, "dur": 48.809,
    "args": {
      "External id": 6715,            "bytes": 2097152, "cid": 108, "correlation": 2252, "ptr": "0x7ffd80420410"
    }
  },
  {
    "ph": "s", "id": 2252, "pid": 870, "tid": 870, "ts": 7429770635989.279,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770636054.623, "dur": 0.375,
    "args": {
      "External id": 6715,      "cid": 370, "correlation": 2253
    }
  },
  {
    "ph": "f", "id": 2253, "pid": 870, "tid": 870, "ts": 7429770636054.623,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770636059.104, "dur": 0.151,
    "args": {
      "External id": 6715,      "cid": 370, "correlation": 2254
    }
  },
  {
    "ph": "f", "id": 2254, "pid": 870, "tid": 870, "ts": 7429770636059.104,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770636062.347, "dur": 0.252,
    "args": {
      "External id": 6715,      "cid": 370, "correlation": 2255
    }
  },
  {
    "ph": "f", "id": 2255, "pid": 870, "tid": 870, "ts": 7429770636062.347,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770636066.241, "dur": 5.490,
    "args": {
      "External id": 6715,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2256,      "grid": [24576, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2256, "pid": 870, "tid": 870, "ts": 7429770636066.241,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770636085.908, "dur": 4.273,
    "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": 2257,      "grid": [221, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2257, "pid": 870, "tid": 870, "ts": 7429770636085.908,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770636105.524, "dur": 3.612,
    "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": 2258,      "grid": [882, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2258, "pid": 870, "tid": 870, "ts": 7429770636105.524,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770636121.627, "dur": 5.840,
    "args": {
      "External id": 6728,          "kernel": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "cid": 107, "correlation": 2259,      "grid": [672, 1, 1],      "block": [64, 2, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2259, "pid": 870, "tid": 870, "ts": 7429770636121.627,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770636169.740, "dur": 0.195,
    "args": {
      "External id": 6740,      "cid": 370, "correlation": 2260
    }
  },
  {
    "ph": "f", "id": 2260, "pid": 870, "tid": 870, "ts": 7429770636169.740,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770636173.265, "dur": 0.139,
    "args": {
      "External id": 6740,      "cid": 370, "correlation": 2261
    }
  },
  {
    "ph": "f", "id": 2261, "pid": 870, "tid": 870, "ts": 7429770636173.265,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770636176.041, "dur": 0.152,
    "args": {
      "External id": 6740,      "cid": 370, "correlation": 2262
    }
  },
  {
    "ph": "f", "id": 2262, "pid": 870, "tid": 870, "ts": 7429770636176.041,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770636178.323, "dur": 6.962,
    "args": {
      "External id": 6740,          "kernel": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2263,      "grid": [16384, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2263, "pid": 870, "tid": 870, "ts": 7429770636178.323,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770636202.297, "dur": 4.193,
    "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": 2264,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2264, "pid": 870, "tid": 870, "ts": 7429770636202.297,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770636249.979, "dur": 0.207,
    "args": {
      "External id": 6759,      "cid": 370, "correlation": 2265
    }
  },
  {
    "ph": "f", "id": 2265, "pid": 870, "tid": 870, "ts": 7429770636249.979,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770636253.212, "dur": 0.184,
    "args": {
      "External id": 6759,      "cid": 370, "correlation": 2266
    }
  },
  {
    "ph": "f", "id": 2266, "pid": 870, "tid": 870, "ts": 7429770636253.212,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770636255.409, "dur": 0.190,
    "args": {
      "External id": 6759,      "cid": 370, "correlation": 2267
    }
  },
  {
    "ph": "f", "id": 2267, "pid": 870, "tid": 870, "ts": 7429770636255.409,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770636257.347, "dur": 4.634,
    "args": {
      "External id": 6759,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2268,      "grid": [49152, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2268, "pid": 870, "tid": 870, "ts": 7429770636257.347,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770636275.485, "dur": 4.793,
    "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": 2269,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2269, "pid": 870, "tid": 870, "ts": 7429770636275.485,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770636298.526, "dur": 5.090,
    "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": 2270,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2270, "pid": 870, "tid": 870, "ts": 7429770636298.526,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770636313.681, "dur": 5.794,
    "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": 2271,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2271, "pid": 870, "tid": 870, "ts": 7429770636313.681,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770636328.130, "dur": 5.255,
    "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": 2272,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2272, "pid": 870, "tid": 870, "ts": 7429770636328.130,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770636339.453, "dur": 4.879,
    "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": 2273,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2273, "pid": 870, "tid": 870, "ts": 7429770636339.453,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770636351.124, "dur": 5.709,
    "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": 2274,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2274, "pid": 870, "tid": 870, "ts": 7429770636351.124,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770636364.474, "dur": 3.708,
    "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": 2275,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2275, "pid": 870, "tid": 870, "ts": 7429770636364.474,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770636402.568, "dur": 0.206,
    "args": {
      "External id": 6780,      "cid": 370, "correlation": 2276
    }
  },
  {
    "ph": "f", "id": 2276, "pid": 870, "tid": 870, "ts": 7429770636402.568,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770636405.821, "dur": 0.141,
    "args": {
      "External id": 6780,      "cid": 370, "correlation": 2277
    }
  },
  {
    "ph": "f", "id": 2277, "pid": 870, "tid": 870, "ts": 7429770636405.821,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770636408.263, "dur": 0.143,
    "args": {
      "External id": 6780,      "cid": 370, "correlation": 2278
    }
  },
  {
    "ph": "f", "id": 2278, "pid": 870, "tid": 870, "ts": 7429770636408.263,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770636410.359, "dur": 4.723,
    "args": {
      "External id": 6780,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2279,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2279, "pid": 870, "tid": 870, "ts": 7429770636410.359,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770636431.265, "dur": 4.910,
    "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": 2280,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2280, "pid": 870, "tid": 870, "ts": 7429770636431.265,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770636463.680, "dur": 0.181,
    "args": {
      "External id": 6791,      "cid": 370, "correlation": 2281
    }
  },
  {
    "ph": "f", "id": 2281, "pid": 870, "tid": 870, "ts": 7429770636463.680,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770636466.280, "dur": 0.166,
    "args": {
      "External id": 6791,      "cid": 370, "correlation": 2282
    }
  },
  {
    "ph": "f", "id": 2282, "pid": 870, "tid": 870, "ts": 7429770636466.280,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770636468.366, "dur": 0.138,
    "args": {
      "External id": 6791,      "cid": 370, "correlation": 2283
    }
  },
  {
    "ph": "f", "id": 2283, "pid": 870, "tid": 870, "ts": 7429770636468.366,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770636470.152, "dur": 5.937,
    "args": {
      "External id": 6791,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2284,      "grid": [23040, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2284, "pid": 870, "tid": 870, "ts": 7429770636470.152,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770636484.521, "dur": 4.152,
    "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": 2285,      "grid": [1176, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2285, "pid": 870, "tid": 870, "ts": 7429770636484.521,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770636510.933, "dur": 0.158,
    "args": {
      "External id": 6802,      "cid": 370, "correlation": 2286
    }
  },
  {
    "ph": "f", "id": 2286, "pid": 870, "tid": 870, "ts": 7429770636510.933,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770636514.218, "dur": 0.145,
    "args": {
      "External id": 6802,      "cid": 370, "correlation": 2287
    }
  },
  {
    "ph": "f", "id": 2287, "pid": 870, "tid": 870, "ts": 7429770636514.218,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770636516.714, "dur": 0.137,
    "args": {
      "External id": 6802,      "cid": 370, "correlation": 2288
    }
  },
  {
    "ph": "f", "id": 2288, "pid": 870, "tid": 870, "ts": 7429770636516.714,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770636525.447, "dur": 7.841,
    "args": {
      "External id": 6802,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2289,      "grid": [165120, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2289, "pid": 870, "tid": 870, "ts": 7429770636525.447,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770636533.659, "dur": 5.251,
    "args": {
      "External id": 6802,          "kernel": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "cid": 65, "correlation": 2290,      "grid": [86016, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2290, "pid": 870, "tid": 870, "ts": 7429770636533.659,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770636547.919, "dur": 3.903,
    "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": 2291,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2291, "pid": 870, "tid": 870, "ts": 7429770636547.919,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770636575.131, "dur": 3.655,
    "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": 2292,      "grid": [336, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2292, "pid": 870, "tid": 870, "ts": 7429770636575.131,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770636586.549, "dur": 5.505,
    "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": 2293,      "grid": [11, 1, 1],      "block": [64, 8, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2293, "pid": 870, "tid": 870, "ts": 7429770636586.549,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770636598.815, "dur": 3.736,
    "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": 2294,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2294, "pid": 870, "tid": 870, "ts": 7429770636598.815,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770636607.550, "dur": 3.338,
    "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": 2295,      "grid": [1, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2295, "pid": 870, "tid": 870, "ts": 7429770636607.550,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770636616.930, "dur": 3.482,
    "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": 2296,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2296, "pid": 870, "tid": 870, "ts": 7429770636616.930,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770636628.526, "dur": 3.189,
    "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": 2297,      "grid": [1344, 1, 1],      "block": [128, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2297, "pid": 870, "tid": 870, "ts": 7429770636628.526,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipStreamIsCapturing", "pid": 870, "tid": 870,
    "ts": 7429770636766.107, "dur": 0.520,
    "args": {
      "External id": 6830,      "cid": 290, "correlation": 2298
    }
  },
  {
    "ph": "f", "id": 2298, "pid": 870, "tid": 870, "ts": 7429770636766.107,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipMalloc", "pid": 870, "tid": 870,
    "ts": 7429770636767.165, "dur": 54.711,
    "args": {
      "External id": 6830,            "bytes": 79691776, "cid": 108, "correlation": 2299, "ptr": "0x7f9a344826f0"
    }
  },
  {
    "ph": "s", "id": 2299, "pid": 870, "tid": 870, "ts": 7429770636767.165,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770636824.595, "dur": 0.311,
    "args": {
      "External id": 6829,      "cid": 370, "correlation": 2300
    }
  },
  {
    "ph": "f", "id": 2300, "pid": 870, "tid": 870, "ts": 7429770636824.595,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770636830.231, "dur": 0.190,
    "args": {
      "External id": 6829,      "cid": 370, "correlation": 2301
    }
  },
  {
    "ph": "f", "id": 2301, "pid": 870, "tid": 870, "ts": 7429770636830.231,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipGetDevicePropertiesR0600", "pid": 870, "tid": 870,
    "ts": 7429770636833.677, "dur": 0.244,
    "args": {
      "External id": 6829,      "cid": 370, "correlation": 2302
    }
  },
  {
    "ph": "f", "id": 2302, "pid": 870, "tid": 870, "ts": 7429770636833.677,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipExtModuleLaunchKernel", "pid": 870, "tid": 870,
    "ts": 7429770636837.576, "dur": 7.429,
    "args": {
      "External id": 6829,          "kernel": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_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_WG32_8_1_WGMXCC1_WGMXCCG0", "cid": 65, "correlation": 2303,      "grid": [513024, 1, 1],      "block": [256, 1, 1],      "shared memory": 0
    }
  },
  {
    "ph": "s", "id": 2303, "pid": 870, "tid": 870, "ts": 7429770636837.576,
    "cat": "ac2g", "name": "ac2g"
  },
  {
    "ph": "X", "cat": "cuda_runtime", "name": "hipDeviceSynchronize", "pid": 870, "tid": 870,
    "ts": 7429770637141.610, "dur": 39049.217,
    "args": {
            "cid": 48, "correlation": 2304
    }
  },
  {
    "ph": "f", "id": 2304, "pid": 870, "tid": 870, "ts": 7429770637141.610,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::(anonymous namespace)::indexSelectLargeIndex<float, long, unsigned int, 2, 2, -2, true>(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, unsigned int, long)", "pid": 2, "tid": 0,
    "ts": 7429770639545.962, "dur": 4.089,
    "args": {
      "External id": 4,      "device": 2, "stream": 0,      "correlation": 1, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [512, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1, "pid": 2, "tid": 0, "ts": 7429770639545.962,
    "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": 7429770639550.052, "dur": 4.649,
    "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": 7429770639550.052,
    "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": 7429770639554.702, "dur": 4.570,
    "args": {
      "External id": 16,      "device": 2, "stream": 0,      "correlation": 3, "kind": "Dispatch Kernel",      "grid":     [7, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 3, "pid": 2, "tid": 0, "ts": 7429770639554.702,
    "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": 7429770639559.273, "dur": 4.529,
    "args": {
      "External id": 17,      "device": 2, "stream": 0,      "correlation": 4, "kind": "Dispatch Kernel",      "grid":     [28, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 4, "pid": 2, "tid": 0, "ts": 7429770639559.273,
    "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": 7429770639563.803, "dur": 4.530,
    "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": 7429770639563.803,
    "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": 7429770639568.334, "dur": 4.690,
    "args": {
      "External id": 24,      "device": 2, "stream": 0,      "correlation": 6, "kind": "Dispatch Kernel",      "grid":     [14, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 6, "pid": 2, "tid": 0, "ts": 7429770639568.334,
    "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": 7429770639573.025, "dur": 5.010,
    "args": {
      "External id": 25,      "device": 2, "stream": 0,      "correlation": 7, "kind": "Dispatch Kernel",      "grid":     [14, 1, 1], "block":     [512, 1, 1]
    }
  },
  {
    "ph": "f", "id": 7, "pid": 2, "tid": 0, "ts": 7429770639573.025,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7429770639578.036, "dur": 4.770,
    "args": {
      "External id": 38,      "device": 2, "stream": 0,      "correlation": 8, "kind": "DtoD",      "bytes": 28224, "memory bandwidth (GB/s)": 5.9157409348145045
    }
  },
  {
    "ph": "f", "id": 8, "pid": 2, "tid": 0, "ts": 7429770639578.036,
    "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": 7429770639582.807, "dur": 8.819,
    "args": {
      "External id": 54,      "device": 2, "stream": 0,      "correlation": 9, "kind": "Dispatch Kernel",      "grid":     [14, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 9, "pid": 2, "tid": 0, "ts": 7429770639582.807,
    "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": 7429770639591.627, "dur": 8.499,
    "args": {
      "External id": 55,      "device": 2, "stream": 0,      "correlation": 10, "kind": "Dispatch Kernel",      "grid":     [2, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 10, "pid": 2, "tid": 0, "ts": 7429770639591.627,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7429770639600.127, "dur": 4.770,
    "args": {
      "External id": 66,      "device": 2, "stream": 0,      "correlation": 11, "kind": "DtoD",      "bytes": 28224, "memory bandwidth (GB/s)": 5.9157409348145045
    }
  },
  {
    "ph": "f", "id": 11, "pid": 2, "tid": 0, "ts": 7429770639600.127,
    "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": 7429770639604.898, "dur": 5.090,
    "args": {
      "External id": 67,      "device": 2, "stream": 0,      "correlation": 12, "kind": "Dispatch Kernel",      "grid":     [7, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 12, "pid": 2, "tid": 0, "ts": 7429770639604.898,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "gpu_memcpy", "name": "Memcpy DtoD (Device -> Device)", "pid": 2, "tid": 0,
    "ts": 7429770639609.989, "dur": 4.971,
    "args": {
      "External id": 74,      "device": 2, "stream": 0,      "correlation": 13, "kind": "DtoD",      "bytes": 28224, "memory bandwidth (GB/s)": 5.676588897827836
    }
  },
  {
    "ph": "f", "id": 13, "pid": 2, "tid": 0, "ts": 7429770639609.989,
    "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": 7429770639614.961, "dur": 4.850,
    "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": 7429770639614.961,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770639619.812, "dur": 5.612,
    "args": {
      "External id": 106,      "device": 2, "stream": 0,      "correlation": 18, "kind": "Dispatch Kernel",      "grid":     [1536, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 18, "pid": 2, "tid": 0, "ts": 7429770639619.812,
    "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": 7429770639625.425, "dur": 5.331,
    "args": {
      "External id": 111,      "device": 2, "stream": 0,      "correlation": 19, "kind": "Dispatch Kernel",      "grid":     [3, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 19, "pid": 2, "tid": 0, "ts": 7429770639625.425,
    "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": 7429770639630.757, "dur": 4.810,
    "args": {
      "External id": 112,      "device": 2, "stream": 0,      "correlation": 20, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 20, "pid": 2, "tid": 0, "ts": 7429770639630.757,
    "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": 7429770639635.568, "dur": 4.650,
    "args": {
      "External id": 113,      "device": 2, "stream": 0,      "correlation": 21, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 21, "pid": 2, "tid": 0, "ts": 7429770639635.568,
    "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": 7429770639640.219, "dur": 5.451,
    "args": {
      "External id": 114,      "device": 2, "stream": 0,      "correlation": 22, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 22, "pid": 2, "tid": 0, "ts": 7429770639640.219,
    "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": 7429770639645.671, "dur": 5.131,
    "args": {
      "External id": 115,      "device": 2, "stream": 0,      "correlation": 23, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 23, "pid": 2, "tid": 0, "ts": 7429770639645.671,
    "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": 7429770639650.803, "dur": 4.850,
    "args": {
      "External id": 119,      "device": 2, "stream": 0,      "correlation": 24, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 24, "pid": 2, "tid": 0, "ts": 7429770639650.803,
    "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": 7429770639655.654, "dur": 8.098,
    "args": {
      "External id": 122,      "device": 2, "stream": 0,      "correlation": 25, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 25, "pid": 2, "tid": 0, "ts": 7429770639655.654,
    "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": 7429770639663.753, "dur": 4.569,
    "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": 7429770639663.753,
    "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": 7429770639668.323, "dur": 4.570,
    "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": 7429770639668.323,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770639672.894, "dur": 4.850,
    "args": {
      "External id": 125,      "device": 2, "stream": 0,      "correlation": 28, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 28, "pid": 2, "tid": 0, "ts": 7429770639672.894,
    "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": 7429770639677.745, "dur": 7.777,
    "args": {
      "External id": 127,      "device": 2, "stream": 0,      "correlation": 29, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 29, "pid": 2, "tid": 0, "ts": 7429770639677.745,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770639685.523, "dur": 64.066,
    "args": {
      "External id": 135,      "device": 2, "stream": 0,      "correlation": 33, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 33, "pid": 2, "tid": 0, "ts": 7429770639685.523,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770639749.590, "dur": 51.036,
    "args": {
      "External id": 148,      "device": 2, "stream": 0,      "correlation": 37, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 37, "pid": 2, "tid": 0, "ts": 7429770639749.590,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770639800.627, "dur": 51.157,
    "args": {
      "External id": 161,      "device": 2, "stream": 0,      "correlation": 41, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 41, "pid": 2, "tid": 0, "ts": 7429770639800.627,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770639851.785, "dur": 4.850,
    "args": {
      "External id": 171,      "device": 2, "stream": 0,      "correlation": 42, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 42, "pid": 2, "tid": 0, "ts": 7429770639851.785,
    "cat": "ac2g", "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": 7429770639856.636, "dur": 4.810,
    "args": {
      "External id": 176,      "device": 2, "stream": 0,      "correlation": 43, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 43, "pid": 2, "tid": 0, "ts": 7429770639856.636,
    "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": 7429770639861.447, "dur": 6.293,
    "args": {
      "External id": 177,      "device": 2, "stream": 0,      "correlation": 44, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 44, "pid": 2, "tid": 0, "ts": 7429770639861.447,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770639867.741, "dur": 5.412,
    "args": {
      "External id": 178,      "device": 2, "stream": 0,      "correlation": 45, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 45, "pid": 2, "tid": 0, "ts": 7429770639867.741,
    "cat": "ac2g", "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": 7429770639873.154, "dur": 5.732,
    "args": {
      "External id": 179,      "device": 2, "stream": 0,      "correlation": 46, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 46, "pid": 2, "tid": 0, "ts": 7429770639873.154,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770639878.887, "dur": 4.970,
    "args": {
      "External id": 180,      "device": 2, "stream": 0,      "correlation": 47, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 47, "pid": 2, "tid": 0, "ts": 7429770639878.887,
    "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::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": 7429770639883.858, "dur": 5.292,
    "args": {
      "External id": 185,      "device": 2, "stream": 0,      "correlation": 48, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 48, "pid": 2, "tid": 0, "ts": 7429770639883.858,
    "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": 7429770639889.151, "dur": 6.453,
    "args": {
      "External id": 186,      "device": 2, "stream": 0,      "correlation": 49, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 49, "pid": 2, "tid": 0, "ts": 7429770639889.151,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770639895.605, "dur": 4.690,
    "args": {
      "External id": 187,      "device": 2, "stream": 0,      "correlation": 50, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 50, "pid": 2, "tid": 0, "ts": 7429770639895.605,
    "cat": "ac2g", "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": 7429770639900.296, "dur": 4.770,
    "args": {
      "External id": 188,      "device": 2, "stream": 0,      "correlation": 51, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 51, "pid": 2, "tid": 0, "ts": 7429770639900.296,
    "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::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::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": 7429770639905.067, "dur": 6.494,
    "args": {
      "External id": 205,      "device": 2, "stream": 0,      "correlation": 52, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 52, "pid": 2, "tid": 0, "ts": 7429770639905.067,
    "cat": "ac2g", "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": 7429770639911.562, "dur": 5.612,
    "args": {
      "External id": 223,      "device": 2, "stream": 0,      "correlation": 53, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 53, "pid": 2, "tid": 0, "ts": 7429770639911.562,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770639917.175, "dur": 10.343,
    "args": {
      "External id": 236,      "device": 2, "stream": 0,      "correlation": 57, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 57, "pid": 2, "tid": 0, "ts": 7429770639917.175,
    "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": 7429770639927.519, "dur": 5.010,
    "args": {
      "External id": 239,      "device": 2, "stream": 0,      "correlation": 58, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 58, "pid": 2, "tid": 0, "ts": 7429770639927.519,
    "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": 7429770639932.530, "dur": 4.850,
    "args": {
      "External id": 246,      "device": 2, "stream": 0,      "correlation": 59, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 59, "pid": 2, "tid": 0, "ts": 7429770639932.530,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770639937.381, "dur": 5.171,
    "args": {
      "External id": 249,      "device": 2, "stream": 0,      "correlation": 60, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 60, "pid": 2, "tid": 0, "ts": 7429770639937.381,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770639942.553, "dur": 11.666,
    "args": {
      "External id": 261,      "device": 2, "stream": 0,      "correlation": 64, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 64, "pid": 2, "tid": 0, "ts": 7429770639942.553,
    "cat": "ac2g", "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": 7429770639954.220, "dur": 5.091,
    "args": {
      "External id": 270,      "device": 2, "stream": 0,      "correlation": 65, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 65, "pid": 2, "tid": 0, "ts": 7429770639954.220,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770639959.312, "dur": 63.344,
    "args": {
      "External id": 280,      "device": 2, "stream": 0,      "correlation": 69, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 69, "pid": 2, "tid": 0, "ts": 7429770639959.312,
    "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": 7429770640022.657, "dur": 5.492,
    "args": {
      "External id": 283,      "device": 2, "stream": 0,      "correlation": 70, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 70, "pid": 2, "tid": 0, "ts": 7429770640022.657,
    "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": 7429770640028.150, "dur": 6.133,
    "args": {
      "External id": 285,      "device": 2, "stream": 0,      "correlation": 71, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 71, "pid": 2, "tid": 0, "ts": 7429770640028.150,
    "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": 7429770640034.284, "dur": 7.817,
    "args": {
      "External id": 288,      "device": 2, "stream": 0,      "correlation": 72, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 72, "pid": 2, "tid": 0, "ts": 7429770640034.284,
    "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": 7429770640042.102, "dur": 4.610,
    "args": {
      "External id": 289,      "device": 2, "stream": 0,      "correlation": 73, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 73, "pid": 2, "tid": 0, "ts": 7429770640042.102,
    "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": 7429770640046.713, "dur": 4.769,
    "args": {
      "External id": 290,      "device": 2, "stream": 0,      "correlation": 74, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 74, "pid": 2, "tid": 0, "ts": 7429770640046.713,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770640051.483, "dur": 4.931,
    "args": {
      "External id": 291,      "device": 2, "stream": 0,      "correlation": 75, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 75, "pid": 2, "tid": 0, "ts": 7429770640051.483,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770640056.415, "dur": 5.010,
    "args": {
      "External id": 293,      "device": 2, "stream": 0,      "correlation": 76, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 76, "pid": 2, "tid": 0, "ts": 7429770640056.415,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770640061.426, "dur": 282.809,
    "args": {
      "External id": 301,      "device": 2, "stream": 0,      "correlation": 80, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 80, "pid": 2, "tid": 0, "ts": 7429770640061.426,
    "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": 7429770640344.236, "dur": 5.892,
    "args": {
      "External id": 304,      "device": 2, "stream": 0,      "correlation": 81, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 81, "pid": 2, "tid": 0, "ts": 7429770640344.236,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770640350.129, "dur": 301.652,
    "args": {
      "External id": 312,      "device": 2, "stream": 0,      "correlation": 85, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 85, "pid": 2, "tid": 0, "ts": 7429770640350.129,
    "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": 7429770640651.782, "dur": 5.692,
    "args": {
      "External id": 315,      "device": 2, "stream": 0,      "correlation": 86, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 86, "pid": 2, "tid": 0, "ts": 7429770640651.782,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770640657.475, "dur": 165.379,
    "args": {
      "External id": 323,      "device": 2, "stream": 0,      "correlation": 90, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 90, "pid": 2, "tid": 0, "ts": 7429770640657.475,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770640822.855, "dur": 7.817,
    "args": {
      "External id": 323,      "device": 2, "stream": 0,      "correlation": 91, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 91, "pid": 2, "tid": 0, "ts": 7429770640822.855,
    "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": 7429770640830.673, "dur": 5.532,
    "args": {
      "External id": 326,      "device": 2, "stream": 0,      "correlation": 92, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 92, "pid": 2, "tid": 0, "ts": 7429770640830.673,
    "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": 7429770640836.206, "dur": 6.093,
    "args": {
      "External id": 328,      "device": 2, "stream": 0,      "correlation": 93, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 93, "pid": 2, "tid": 0, "ts": 7429770640836.206,
    "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": 7429770640842.300, "dur": 8.819,
    "args": {
      "External id": 331,      "device": 2, "stream": 0,      "correlation": 94, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 94, "pid": 2, "tid": 0, "ts": 7429770640842.300,
    "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": 7429770640851.120, "dur": 5.331,
    "args": {
      "External id": 332,      "device": 2, "stream": 0,      "correlation": 95, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 95, "pid": 2, "tid": 0, "ts": 7429770640851.120,
    "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": 7429770640856.452, "dur": 5.732,
    "args": {
      "External id": 333,      "device": 2, "stream": 0,      "correlation": 96, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 96, "pid": 2, "tid": 0, "ts": 7429770640856.452,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770640862.185, "dur": 5.733,
    "args": {
      "External id": 334,      "device": 2, "stream": 0,      "correlation": 97, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 97, "pid": 2, "tid": 0, "ts": 7429770640862.185,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770640867.919, "dur": 5.892,
    "args": {
      "External id": 336,      "device": 2, "stream": 0,      "correlation": 98, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 98, "pid": 2, "tid": 0, "ts": 7429770640867.919,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770640873.812, "dur": 69.118,
    "args": {
      "External id": 344,      "device": 2, "stream": 0,      "correlation": 102, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 102, "pid": 2, "tid": 0, "ts": 7429770640873.812,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770640942.931, "dur": 53.923,
    "args": {
      "External id": 357,      "device": 2, "stream": 0,      "correlation": 106, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 106, "pid": 2, "tid": 0, "ts": 7429770640942.931,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770640996.855, "dur": 54.123,
    "args": {
      "External id": 370,      "device": 2, "stream": 0,      "correlation": 110, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 110, "pid": 2, "tid": 0, "ts": 7429770640996.855,
    "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": 7429770641050.979, "dur": 5.131,
    "args": {
      "External id": 380,      "device": 2, "stream": 0,      "correlation": 111, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 111, "pid": 2, "tid": 0, "ts": 7429770641050.979,
    "cat": "ac2g", "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": 7429770641056.111, "dur": 5.251,
    "args": {
      "External id": 385,      "device": 2, "stream": 0,      "correlation": 112, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 112, "pid": 2, "tid": 0, "ts": 7429770641056.111,
    "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": 7429770641061.363, "dur": 6.935,
    "args": {
      "External id": 386,      "device": 2, "stream": 0,      "correlation": 113, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 113, "pid": 2, "tid": 0, "ts": 7429770641061.363,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770641068.299, "dur": 5.211,
    "args": {
      "External id": 387,      "device": 2, "stream": 0,      "correlation": 114, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 114, "pid": 2, "tid": 0, "ts": 7429770641068.299,
    "cat": "ac2g", "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": 7429770641073.511, "dur": 5.411,
    "args": {
      "External id": 388,      "device": 2, "stream": 0,      "correlation": 115, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 115, "pid": 2, "tid": 0, "ts": 7429770641073.511,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770641078.923, "dur": 4.971,
    "args": {
      "External id": 389,      "device": 2, "stream": 0,      "correlation": 116, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 116, "pid": 2, "tid": 0, "ts": 7429770641078.923,
    "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": 7429770641083.895, "dur": 4.930,
    "args": {
      "External id": 394,      "device": 2, "stream": 0,      "correlation": 117, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 117, "pid": 2, "tid": 0, "ts": 7429770641083.895,
    "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": 7429770641088.826, "dur": 6.895,
    "args": {
      "External id": 395,      "device": 2, "stream": 0,      "correlation": 118, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 118, "pid": 2, "tid": 0, "ts": 7429770641088.826,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770641095.722, "dur": 5.331,
    "args": {
      "External id": 396,      "device": 2, "stream": 0,      "correlation": 119, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 119, "pid": 2, "tid": 0, "ts": 7429770641095.722,
    "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": 7429770641101.054, "dur": 5.692,
    "args": {
      "External id": 397,      "device": 2, "stream": 0,      "correlation": 120, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 120, "pid": 2, "tid": 0, "ts": 7429770641101.054,
    "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::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::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": 7429770641106.747, "dur": 6.254,
    "args": {
      "External id": 414,      "device": 2, "stream": 0,      "correlation": 121, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 121, "pid": 2, "tid": 0, "ts": 7429770641106.747,
    "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::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::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": 7429770641113.002, "dur": 5.892,
    "args": {
      "External id": 432,      "device": 2, "stream": 0,      "correlation": 122, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 122, "pid": 2, "tid": 0, "ts": 7429770641113.002,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770641118.895, "dur": 10.784,
    "args": {
      "External id": 445,      "device": 2, "stream": 0,      "correlation": 126, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 126, "pid": 2, "tid": 0, "ts": 7429770641118.895,
    "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": 7429770641129.680, "dur": 5.050,
    "args": {
      "External id": 448,      "device": 2, "stream": 0,      "correlation": 127, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 127, "pid": 2, "tid": 0, "ts": 7429770641129.680,
    "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": 7429770641134.731, "dur": 5.292,
    "args": {
      "External id": 455,      "device": 2, "stream": 0,      "correlation": 128, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 128, "pid": 2, "tid": 0, "ts": 7429770641134.731,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770641140.024, "dur": 5.571,
    "args": {
      "External id": 458,      "device": 2, "stream": 0,      "correlation": 129, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 129, "pid": 2, "tid": 0, "ts": 7429770641140.024,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770641145.596, "dur": 12.628,
    "args": {
      "External id": 470,      "device": 2, "stream": 0,      "correlation": 133, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 133, "pid": 2, "tid": 0, "ts": 7429770641145.596,
    "cat": "ac2g", "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": 7429770641158.225, "dur": 5.773,
    "args": {
      "External id": 479,      "device": 2, "stream": 0,      "correlation": 134, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 134, "pid": 2, "tid": 0, "ts": 7429770641158.225,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770641163.999, "dur": 67.073,
    "args": {
      "External id": 489,      "device": 2, "stream": 0,      "correlation": 138, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 138, "pid": 2, "tid": 0, "ts": 7429770641163.999,
    "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": 7429770641231.073, "dur": 5.171,
    "args": {
      "External id": 492,      "device": 2, "stream": 0,      "correlation": 139, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 139, "pid": 2, "tid": 0, "ts": 7429770641231.073,
    "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": 7429770641236.245, "dur": 5.291,
    "args": {
      "External id": 494,      "device": 2, "stream": 0,      "correlation": 140, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 140, "pid": 2, "tid": 0, "ts": 7429770641236.245,
    "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": 7429770641241.537, "dur": 8.899,
    "args": {
      "External id": 497,      "device": 2, "stream": 0,      "correlation": 141, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 141, "pid": 2, "tid": 0, "ts": 7429770641241.537,
    "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": 7429770641250.437, "dur": 5.251,
    "args": {
      "External id": 498,      "device": 2, "stream": 0,      "correlation": 142, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 142, "pid": 2, "tid": 0, "ts": 7429770641250.437,
    "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": 7429770641255.689, "dur": 5.211,
    "args": {
      "External id": 499,      "device": 2, "stream": 0,      "correlation": 143, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 143, "pid": 2, "tid": 0, "ts": 7429770641255.689,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770641260.901, "dur": 5.131,
    "args": {
      "External id": 500,      "device": 2, "stream": 0,      "correlation": 144, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 144, "pid": 2, "tid": 0, "ts": 7429770641260.901,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770641266.033, "dur": 5.171,
    "args": {
      "External id": 502,      "device": 2, "stream": 0,      "correlation": 145, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 145, "pid": 2, "tid": 0, "ts": 7429770641266.033,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770641271.205, "dur": 268.095,
    "args": {
      "External id": 510,      "device": 2, "stream": 0,      "correlation": 149, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 149, "pid": 2, "tid": 0, "ts": 7429770641271.205,
    "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": 7429770641539.301, "dur": 6.053,
    "args": {
      "External id": 513,      "device": 2, "stream": 0,      "correlation": 150, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 150, "pid": 2, "tid": 0, "ts": 7429770641539.301,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770641545.355, "dur": 267.252,
    "args": {
      "External id": 521,      "device": 2, "stream": 0,      "correlation": 154, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 154, "pid": 2, "tid": 0, "ts": 7429770641545.355,
    "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": 7429770641812.608, "dur": 5.813,
    "args": {
      "External id": 524,      "device": 2, "stream": 0,      "correlation": 155, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 155, "pid": 2, "tid": 0, "ts": 7429770641812.608,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770641818.422, "dur": 165.419,
    "args": {
      "External id": 532,      "device": 2, "stream": 0,      "correlation": 159, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 159, "pid": 2, "tid": 0, "ts": 7429770641818.422,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770641983.842, "dur": 7.295,
    "args": {
      "External id": 532,      "device": 2, "stream": 0,      "correlation": 160, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 160, "pid": 2, "tid": 0, "ts": 7429770641983.842,
    "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": 7429770641991.138, "dur": 5.412,
    "args": {
      "External id": 535,      "device": 2, "stream": 0,      "correlation": 161, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 161, "pid": 2, "tid": 0, "ts": 7429770641991.138,
    "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": 7429770641996.551, "dur": 6.614,
    "args": {
      "External id": 537,      "device": 2, "stream": 0,      "correlation": 162, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 162, "pid": 2, "tid": 0, "ts": 7429770641996.551,
    "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": 7429770642003.166, "dur": 8.258,
    "args": {
      "External id": 540,      "device": 2, "stream": 0,      "correlation": 163, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 163, "pid": 2, "tid": 0, "ts": 7429770642003.166,
    "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": 7429770642011.425, "dur": 4.570,
    "args": {
      "External id": 541,      "device": 2, "stream": 0,      "correlation": 164, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 164, "pid": 2, "tid": 0, "ts": 7429770642011.425,
    "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": 7429770642015.996, "dur": 4.649,
    "args": {
      "External id": 542,      "device": 2, "stream": 0,      "correlation": 165, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 165, "pid": 2, "tid": 0, "ts": 7429770642015.996,
    "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": 7429770642020.646, "dur": 4.931,
    "args": {
      "External id": 543,      "device": 2, "stream": 0,      "correlation": 166, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 166, "pid": 2, "tid": 0, "ts": 7429770642020.646,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770642025.578, "dur": 5.170,
    "args": {
      "External id": 545,      "device": 2, "stream": 0,      "correlation": 167, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 167, "pid": 2, "tid": 0, "ts": 7429770642025.578,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770642030.749, "dur": 65.269,
    "args": {
      "External id": 553,      "device": 2, "stream": 0,      "correlation": 171, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 171, "pid": 2, "tid": 0, "ts": 7429770642030.749,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770642096.019, "dur": 51.838,
    "args": {
      "External id": 566,      "device": 2, "stream": 0,      "correlation": 175, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 175, "pid": 2, "tid": 0, "ts": 7429770642096.019,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770642147.858, "dur": 51.157,
    "args": {
      "External id": 579,      "device": 2, "stream": 0,      "correlation": 179, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 179, "pid": 2, "tid": 0, "ts": 7429770642147.858,
    "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": 7429770642199.016, "dur": 5.091,
    "args": {
      "External id": 589,      "device": 2, "stream": 0,      "correlation": 180, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 180, "pid": 2, "tid": 0, "ts": 7429770642199.016,
    "cat": "ac2g", "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": 7429770642204.108, "dur": 5.652,
    "args": {
      "External id": 594,      "device": 2, "stream": 0,      "correlation": 181, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 181, "pid": 2, "tid": 0, "ts": 7429770642204.108,
    "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": 7429770642209.761, "dur": 6.373,
    "args": {
      "External id": 595,      "device": 2, "stream": 0,      "correlation": 182, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 182, "pid": 2, "tid": 0, "ts": 7429770642209.761,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770642216.135, "dur": 5.131,
    "args": {
      "External id": 596,      "device": 2, "stream": 0,      "correlation": 183, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 183, "pid": 2, "tid": 0, "ts": 7429770642216.135,
    "cat": "ac2g", "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": 7429770642221.267, "dur": 5.251,
    "args": {
      "External id": 597,      "device": 2, "stream": 0,      "correlation": 184, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 184, "pid": 2, "tid": 0, "ts": 7429770642221.267,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770642226.519, "dur": 5.371,
    "args": {
      "External id": 598,      "device": 2, "stream": 0,      "correlation": 185, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 185, "pid": 2, "tid": 0, "ts": 7429770642226.519,
    "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::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": 7429770642231.891, "dur": 5.492,
    "args": {
      "External id": 603,      "device": 2, "stream": 0,      "correlation": 186, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 186, "pid": 2, "tid": 0, "ts": 7429770642231.891,
    "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": 7429770642237.384, "dur": 6.093,
    "args": {
      "External id": 604,      "device": 2, "stream": 0,      "correlation": 187, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 187, "pid": 2, "tid": 0, "ts": 7429770642237.384,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770642243.478, "dur": 4.970,
    "args": {
      "External id": 605,      "device": 2, "stream": 0,      "correlation": 188, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 188, "pid": 2, "tid": 0, "ts": 7429770642243.478,
    "cat": "ac2g", "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": 7429770642248.449, "dur": 5.011,
    "args": {
      "External id": 606,      "device": 2, "stream": 0,      "correlation": 189, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 189, "pid": 2, "tid": 0, "ts": 7429770642248.449,
    "cat": "ac2g", "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": 7429770642253.461, "dur": 6.494,
    "args": {
      "External id": 623,      "device": 2, "stream": 0,      "correlation": 190, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 190, "pid": 2, "tid": 0, "ts": 7429770642253.461,
    "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::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::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": 7429770642259.956, "dur": 5.451,
    "args": {
      "External id": 641,      "device": 2, "stream": 0,      "correlation": 191, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 191, "pid": 2, "tid": 0, "ts": 7429770642259.956,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770642265.408, "dur": 9.822,
    "args": {
      "External id": 654,      "device": 2, "stream": 0,      "correlation": 195, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 195, "pid": 2, "tid": 0, "ts": 7429770642265.408,
    "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": 7429770642275.231, "dur": 4.890,
    "args": {
      "External id": 657,      "device": 2, "stream": 0,      "correlation": 196, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 196, "pid": 2, "tid": 0, "ts": 7429770642275.231,
    "cat": "ac2g", "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": 7429770642280.122, "dur": 4.810,
    "args": {
      "External id": 664,      "device": 2, "stream": 0,      "correlation": 197, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 197, "pid": 2, "tid": 0, "ts": 7429770642280.122,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770642284.933, "dur": 5.532,
    "args": {
      "External id": 667,      "device": 2, "stream": 0,      "correlation": 198, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 198, "pid": 2, "tid": 0, "ts": 7429770642284.933,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770642290.466, "dur": 10.784,
    "args": {
      "External id": 679,      "device": 2, "stream": 0,      "correlation": 202, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 202, "pid": 2, "tid": 0, "ts": 7429770642290.466,
    "cat": "ac2g", "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": 7429770642301.251, "dur": 5.050,
    "args": {
      "External id": 688,      "device": 2, "stream": 0,      "correlation": 203, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 203, "pid": 2, "tid": 0, "ts": 7429770642301.251,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770642306.302, "dur": 64.748,
    "args": {
      "External id": 698,      "device": 2, "stream": 0,      "correlation": 207, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 207, "pid": 2, "tid": 0, "ts": 7429770642306.302,
    "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": 7429770642371.051, "dur": 5.251,
    "args": {
      "External id": 701,      "device": 2, "stream": 0,      "correlation": 208, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 208, "pid": 2, "tid": 0, "ts": 7429770642371.051,
    "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": 7429770642376.303, "dur": 6.694,
    "args": {
      "External id": 703,      "device": 2, "stream": 0,      "correlation": 209, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 209, "pid": 2, "tid": 0, "ts": 7429770642376.303,
    "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": 7429770642382.998, "dur": 8.018,
    "args": {
      "External id": 706,      "device": 2, "stream": 0,      "correlation": 210, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 210, "pid": 2, "tid": 0, "ts": 7429770642382.998,
    "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": 7429770642391.017, "dur": 5.091,
    "args": {
      "External id": 707,      "device": 2, "stream": 0,      "correlation": 211, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 211, "pid": 2, "tid": 0, "ts": 7429770642391.017,
    "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": 7429770642396.109, "dur": 5.130,
    "args": {
      "External id": 708,      "device": 2, "stream": 0,      "correlation": 212, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 212, "pid": 2, "tid": 0, "ts": 7429770642396.109,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770642401.240, "dur": 4.690,
    "args": {
      "External id": 709,      "device": 2, "stream": 0,      "correlation": 213, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 213, "pid": 2, "tid": 0, "ts": 7429770642401.240,
    "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": 7429770642405.931, "dur": 4.971,
    "args": {
      "External id": 711,      "device": 2, "stream": 0,      "correlation": 214, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 214, "pid": 2, "tid": 0, "ts": 7429770642405.931,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770642410.903, "dur": 252.940,
    "args": {
      "External id": 719,      "device": 2, "stream": 0,      "correlation": 218, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 218, "pid": 2, "tid": 0, "ts": 7429770642410.903,
    "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": 7429770642663.844, "dur": 5.451,
    "args": {
      "External id": 722,      "device": 2, "stream": 0,      "correlation": 219, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 219, "pid": 2, "tid": 0, "ts": 7429770642663.844,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770642669.296, "dur": 252.860,
    "args": {
      "External id": 730,      "device": 2, "stream": 0,      "correlation": 223, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 223, "pid": 2, "tid": 0, "ts": 7429770642669.296,
    "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": 7429770642922.157, "dur": 5.532,
    "args": {
      "External id": 733,      "device": 2, "stream": 0,      "correlation": 224, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 224, "pid": 2, "tid": 0, "ts": 7429770642922.157,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770642927.690, "dur": 164.657,
    "args": {
      "External id": 741,      "device": 2, "stream": 0,      "correlation": 228, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 228, "pid": 2, "tid": 0, "ts": 7429770642927.690,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770643092.348, "dur": 7.736,
    "args": {
      "External id": 741,      "device": 2, "stream": 0,      "correlation": 229, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 229, "pid": 2, "tid": 0, "ts": 7429770643092.348,
    "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": 7429770643100.085, "dur": 4.811,
    "args": {
      "External id": 744,      "device": 2, "stream": 0,      "correlation": 230, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 230, "pid": 2, "tid": 0, "ts": 7429770643100.085,
    "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": 7429770643104.897, "dur": 5.491,
    "args": {
      "External id": 746,      "device": 2, "stream": 0,      "correlation": 231, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 231, "pid": 2, "tid": 0, "ts": 7429770643104.897,
    "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": 7429770643110.389, "dur": 8.058,
    "args": {
      "External id": 749,      "device": 2, "stream": 0,      "correlation": 232, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 232, "pid": 2, "tid": 0, "ts": 7429770643110.389,
    "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": 7429770643118.448, "dur": 4.850,
    "args": {
      "External id": 750,      "device": 2, "stream": 0,      "correlation": 233, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 233, "pid": 2, "tid": 0, "ts": 7429770643118.448,
    "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": 7429770643123.299, "dur": 4.489,
    "args": {
      "External id": 751,      "device": 2, "stream": 0,      "correlation": 234, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 234, "pid": 2, "tid": 0, "ts": 7429770643123.299,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770643127.789, "dur": 4.810,
    "args": {
      "External id": 752,      "device": 2, "stream": 0,      "correlation": 235, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 235, "pid": 2, "tid": 0, "ts": 7429770643127.789,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770643132.600, "dur": 5.011,
    "args": {
      "External id": 754,      "device": 2, "stream": 0,      "correlation": 236, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 236, "pid": 2, "tid": 0, "ts": 7429770643132.600,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770643137.612, "dur": 63.705,
    "args": {
      "External id": 762,      "device": 2, "stream": 0,      "correlation": 240, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 240, "pid": 2, "tid": 0, "ts": 7429770643137.612,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770643201.318, "dur": 50.074,
    "args": {
      "External id": 775,      "device": 2, "stream": 0,      "correlation": 244, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 244, "pid": 2, "tid": 0, "ts": 7429770643201.318,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770643251.393, "dur": 50.996,
    "args": {
      "External id": 788,      "device": 2, "stream": 0,      "correlation": 248, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 248, "pid": 2, "tid": 0, "ts": 7429770643251.393,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770643302.390, "dur": 5.171,
    "args": {
      "External id": 798,      "device": 2, "stream": 0,      "correlation": 249, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 249, "pid": 2, "tid": 0, "ts": 7429770643302.390,
    "cat": "ac2g", "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": 7429770643307.562, "dur": 5.371,
    "args": {
      "External id": 803,      "device": 2, "stream": 0,      "correlation": 250, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 250, "pid": 2, "tid": 0, "ts": 7429770643307.562,
    "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": 7429770643312.934, "dur": 6.214,
    "args": {
      "External id": 804,      "device": 2, "stream": 0,      "correlation": 251, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 251, "pid": 2, "tid": 0, "ts": 7429770643312.934,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770643319.149, "dur": 4.729,
    "args": {
      "External id": 805,      "device": 2, "stream": 0,      "correlation": 252, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 252, "pid": 2, "tid": 0, "ts": 7429770643319.149,
    "cat": "ac2g", "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": 7429770643323.879, "dur": 5.372,
    "args": {
      "External id": 806,      "device": 2, "stream": 0,      "correlation": 253, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 253, "pid": 2, "tid": 0, "ts": 7429770643323.879,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770643329.252, "dur": 4.770,
    "args": {
      "External id": 807,      "device": 2, "stream": 0,      "correlation": 254, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 254, "pid": 2, "tid": 0, "ts": 7429770643329.252,
    "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": 7429770643334.023, "dur": 4.730,
    "args": {
      "External id": 812,      "device": 2, "stream": 0,      "correlation": 255, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 255, "pid": 2, "tid": 0, "ts": 7429770643334.023,
    "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": 7429770643338.754, "dur": 6.453,
    "args": {
      "External id": 813,      "device": 2, "stream": 0,      "correlation": 256, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 256, "pid": 2, "tid": 0, "ts": 7429770643338.754,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770643345.208, "dur": 5.332,
    "args": {
      "External id": 814,      "device": 2, "stream": 0,      "correlation": 257, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 257, "pid": 2, "tid": 0, "ts": 7429770643345.208,
    "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::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": 7429770643350.541, "dur": 5.090,
    "args": {
      "External id": 815,      "device": 2, "stream": 0,      "correlation": 258, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 258, "pid": 2, "tid": 0, "ts": 7429770643350.541,
    "cat": "ac2g", "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": 7429770643355.632, "dur": 5.171,
    "args": {
      "External id": 832,      "device": 2, "stream": 0,      "correlation": 259, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 259, "pid": 2, "tid": 0, "ts": 7429770643355.632,
    "cat": "ac2g", "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": 7429770643360.804, "dur": 5.893,
    "args": {
      "External id": 850,      "device": 2, "stream": 0,      "correlation": 260, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 260, "pid": 2, "tid": 0, "ts": 7429770643360.804,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770643366.698, "dur": 9.821,
    "args": {
      "External id": 863,      "device": 2, "stream": 0,      "correlation": 264, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 264, "pid": 2, "tid": 0, "ts": 7429770643366.698,
    "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": 7429770643376.520, "dur": 5.372,
    "args": {
      "External id": 866,      "device": 2, "stream": 0,      "correlation": 265, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 265, "pid": 2, "tid": 0, "ts": 7429770643376.520,
    "cat": "ac2g", "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": 7429770643381.893, "dur": 5.451,
    "args": {
      "External id": 873,      "device": 2, "stream": 0,      "correlation": 266, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 266, "pid": 2, "tid": 0, "ts": 7429770643381.893,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770643387.345, "dur": 5.572,
    "args": {
      "External id": 876,      "device": 2, "stream": 0,      "correlation": 267, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 267, "pid": 2, "tid": 0, "ts": 7429770643387.345,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770643392.918, "dur": 10.864,
    "args": {
      "External id": 888,      "device": 2, "stream": 0,      "correlation": 271, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 271, "pid": 2, "tid": 0, "ts": 7429770643392.918,
    "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::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::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": 7429770643403.783, "dur": 5.291,
    "args": {
      "External id": 897,      "device": 2, "stream": 0,      "correlation": 272, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 272, "pid": 2, "tid": 0, "ts": 7429770643403.783,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770643409.075, "dur": 63.786,
    "args": {
      "External id": 907,      "device": 2, "stream": 0,      "correlation": 276, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 276, "pid": 2, "tid": 0, "ts": 7429770643409.075,
    "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": 7429770643472.862, "dur": 5.251,
    "args": {
      "External id": 910,      "device": 2, "stream": 0,      "correlation": 277, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 277, "pid": 2, "tid": 0, "ts": 7429770643472.862,
    "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": 7429770643478.114, "dur": 6.053,
    "args": {
      "External id": 912,      "device": 2, "stream": 0,      "correlation": 278, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 278, "pid": 2, "tid": 0, "ts": 7429770643478.114,
    "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": 7429770643484.168, "dur": 8.017,
    "args": {
      "External id": 915,      "device": 2, "stream": 0,      "correlation": 279, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 279, "pid": 2, "tid": 0, "ts": 7429770643484.168,
    "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": 7429770643492.186, "dur": 5.371,
    "args": {
      "External id": 916,      "device": 2, "stream": 0,      "correlation": 280, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 280, "pid": 2, "tid": 0, "ts": 7429770643492.186,
    "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": 7429770643497.558, "dur": 5.452,
    "args": {
      "External id": 917,      "device": 2, "stream": 0,      "correlation": 281, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 281, "pid": 2, "tid": 0, "ts": 7429770643497.558,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770643503.011, "dur": 5.291,
    "args": {
      "External id": 918,      "device": 2, "stream": 0,      "correlation": 282, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 282, "pid": 2, "tid": 0, "ts": 7429770643503.011,
    "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": 7429770643508.303, "dur": 4.970,
    "args": {
      "External id": 920,      "device": 2, "stream": 0,      "correlation": 283, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 283, "pid": 2, "tid": 0, "ts": 7429770643508.303,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770643513.274, "dur": 252.900,
    "args": {
      "External id": 928,      "device": 2, "stream": 0,      "correlation": 287, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 287, "pid": 2, "tid": 0, "ts": 7429770643513.274,
    "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": 7429770643766.175, "dur": 5.572,
    "args": {
      "External id": 931,      "device": 2, "stream": 0,      "correlation": 288, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 288, "pid": 2, "tid": 0, "ts": 7429770643766.175,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770643771.748, "dur": 252.459,
    "args": {
      "External id": 939,      "device": 2, "stream": 0,      "correlation": 292, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 292, "pid": 2, "tid": 0, "ts": 7429770643771.748,
    "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": 7429770644024.208, "dur": 5.772,
    "args": {
      "External id": 942,      "device": 2, "stream": 0,      "correlation": 293, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 293, "pid": 2, "tid": 0, "ts": 7429770644024.208,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770644029.981, "dur": 164.858,
    "args": {
      "External id": 950,      "device": 2, "stream": 0,      "correlation": 297, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 297, "pid": 2, "tid": 0, "ts": 7429770644029.981,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770644194.840, "dur": 7.817,
    "args": {
      "External id": 950,      "device": 2, "stream": 0,      "correlation": 298, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 298, "pid": 2, "tid": 0, "ts": 7429770644194.840,
    "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": 7429770644202.658, "dur": 5.371,
    "args": {
      "External id": 953,      "device": 2, "stream": 0,      "correlation": 299, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 299, "pid": 2, "tid": 0, "ts": 7429770644202.658,
    "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": 7429770644208.030, "dur": 5.813,
    "args": {
      "External id": 955,      "device": 2, "stream": 0,      "correlation": 300, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 300, "pid": 2, "tid": 0, "ts": 7429770644208.030,
    "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": 7429770644213.844, "dur": 7.977,
    "args": {
      "External id": 958,      "device": 2, "stream": 0,      "correlation": 301, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 301, "pid": 2, "tid": 0, "ts": 7429770644213.844,
    "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": 7429770644221.822, "dur": 4.890,
    "args": {
      "External id": 959,      "device": 2, "stream": 0,      "correlation": 302, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 302, "pid": 2, "tid": 0, "ts": 7429770644221.822,
    "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": 7429770644226.713, "dur": 4.850,
    "args": {
      "External id": 960,      "device": 2, "stream": 0,      "correlation": 303, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 303, "pid": 2, "tid": 0, "ts": 7429770644226.713,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770644231.564, "dur": 4.770,
    "args": {
      "External id": 961,      "device": 2, "stream": 0,      "correlation": 304, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 304, "pid": 2, "tid": 0, "ts": 7429770644231.564,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770644236.335, "dur": 5.091,
    "args": {
      "External id": 963,      "device": 2, "stream": 0,      "correlation": 305, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 305, "pid": 2, "tid": 0, "ts": 7429770644236.335,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770644241.427, "dur": 63.625,
    "args": {
      "External id": 971,      "device": 2, "stream": 0,      "correlation": 311, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 311, "pid": 2, "tid": 0, "ts": 7429770644241.427,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770644305.053, "dur": 50.956,
    "args": {
      "External id": 984,      "device": 2, "stream": 0,      "correlation": 315, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 315, "pid": 2, "tid": 0, "ts": 7429770644305.053,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770644356.010, "dur": 50.395,
    "args": {
      "External id": 997,      "device": 2, "stream": 0,      "correlation": 319, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 319, "pid": 2, "tid": 0, "ts": 7429770644356.010,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770644406.406, "dur": 4.730,
    "args": {
      "External id": 1007,      "device": 2, "stream": 0,      "correlation": 320, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 320, "pid": 2, "tid": 0, "ts": 7429770644406.406,
    "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::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": 7429770644411.137, "dur": 4.730,
    "args": {
      "External id": 1012,      "device": 2, "stream": 0,      "correlation": 321, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 321, "pid": 2, "tid": 0, "ts": 7429770644411.137,
    "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": 7429770644415.868, "dur": 6.213,
    "args": {
      "External id": 1013,      "device": 2, "stream": 0,      "correlation": 322, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 322, "pid": 2, "tid": 0, "ts": 7429770644415.868,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770644422.082, "dur": 4.730,
    "args": {
      "External id": 1014,      "device": 2, "stream": 0,      "correlation": 323, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 323, "pid": 2, "tid": 0, "ts": 7429770644422.082,
    "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::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": 7429770644426.813, "dur": 5.612,
    "args": {
      "External id": 1015,      "device": 2, "stream": 0,      "correlation": 324, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 324, "pid": 2, "tid": 0, "ts": 7429770644426.813,
    "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": 7429770644432.426, "dur": 5.491,
    "args": {
      "External id": 1016,      "device": 2, "stream": 0,      "correlation": 325, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 325, "pid": 2, "tid": 0, "ts": 7429770644432.426,
    "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::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": 7429770644437.918, "dur": 5.091,
    "args": {
      "External id": 1021,      "device": 2, "stream": 0,      "correlation": 326, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 326, "pid": 2, "tid": 0, "ts": 7429770644437.918,
    "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": 7429770644443.010, "dur": 6.053,
    "args": {
      "External id": 1022,      "device": 2, "stream": 0,      "correlation": 327, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 327, "pid": 2, "tid": 0, "ts": 7429770644443.010,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770644449.064, "dur": 4.569,
    "args": {
      "External id": 1023,      "device": 2, "stream": 0,      "correlation": 328, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 328, "pid": 2, "tid": 0, "ts": 7429770644449.064,
    "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::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": 7429770644453.634, "dur": 5.412,
    "args": {
      "External id": 1024,      "device": 2, "stream": 0,      "correlation": 329, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 329, "pid": 2, "tid": 0, "ts": 7429770644453.634,
    "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::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::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": 7429770644459.047, "dur": 5.451,
    "args": {
      "External id": 1041,      "device": 2, "stream": 0,      "correlation": 330, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 330, "pid": 2, "tid": 0, "ts": 7429770644459.047,
    "cat": "ac2g", "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": 7429770644464.499, "dur": 5.372,
    "args": {
      "External id": 1059,      "device": 2, "stream": 0,      "correlation": 331, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 331, "pid": 2, "tid": 0, "ts": 7429770644464.499,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770644469.872, "dur": 9.581,
    "args": {
      "External id": 1072,      "device": 2, "stream": 0,      "correlation": 335, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 335, "pid": 2, "tid": 0, "ts": 7429770644469.872,
    "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": 7429770644479.454, "dur": 5.090,
    "args": {
      "External id": 1075,      "device": 2, "stream": 0,      "correlation": 336, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 336, "pid": 2, "tid": 0, "ts": 7429770644479.454,
    "cat": "ac2g", "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": 7429770644484.545, "dur": 5.412,
    "args": {
      "External id": 1082,      "device": 2, "stream": 0,      "correlation": 337, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 337, "pid": 2, "tid": 0, "ts": 7429770644484.545,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770644489.958, "dur": 5.090,
    "args": {
      "External id": 1085,      "device": 2, "stream": 0,      "correlation": 338, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 338, "pid": 2, "tid": 0, "ts": 7429770644489.958,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770644495.049, "dur": 11.145,
    "args": {
      "External id": 1097,      "device": 2, "stream": 0,      "correlation": 342, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 342, "pid": 2, "tid": 0, "ts": 7429770644495.049,
    "cat": "ac2g", "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": 7429770644506.195, "dur": 5.371,
    "args": {
      "External id": 1106,      "device": 2, "stream": 0,      "correlation": 343, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 343, "pid": 2, "tid": 0, "ts": 7429770644506.195,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770644511.567, "dur": 63.786,
    "args": {
      "External id": 1116,      "device": 2, "stream": 0,      "correlation": 347, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 347, "pid": 2, "tid": 0, "ts": 7429770644511.567,
    "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": 7429770644575.354, "dur": 4.650,
    "args": {
      "External id": 1119,      "device": 2, "stream": 0,      "correlation": 348, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 348, "pid": 2, "tid": 0, "ts": 7429770644575.354,
    "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": 7429770644580.005, "dur": 6.133,
    "args": {
      "External id": 1121,      "device": 2, "stream": 0,      "correlation": 349, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 349, "pid": 2, "tid": 0, "ts": 7429770644580.005,
    "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": 7429770644586.139, "dur": 7.656,
    "args": {
      "External id": 1124,      "device": 2, "stream": 0,      "correlation": 350, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 350, "pid": 2, "tid": 0, "ts": 7429770644586.139,
    "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": 7429770644593.796, "dur": 4.530,
    "args": {
      "External id": 1125,      "device": 2, "stream": 0,      "correlation": 351, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 351, "pid": 2, "tid": 0, "ts": 7429770644593.796,
    "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": 7429770644598.327, "dur": 4.569,
    "args": {
      "External id": 1126,      "device": 2, "stream": 0,      "correlation": 352, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 352, "pid": 2, "tid": 0, "ts": 7429770644598.327,
    "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": 7429770644602.897, "dur": 4.770,
    "args": {
      "External id": 1127,      "device": 2, "stream": 0,      "correlation": 353, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 353, "pid": 2, "tid": 0, "ts": 7429770644602.897,
    "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": 7429770644607.668, "dur": 4.890,
    "args": {
      "External id": 1129,      "device": 2, "stream": 0,      "correlation": 354, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 354, "pid": 2, "tid": 0, "ts": 7429770644607.668,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770644612.559, "dur": 252.499,
    "args": {
      "External id": 1137,      "device": 2, "stream": 0,      "correlation": 358, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 358, "pid": 2, "tid": 0, "ts": 7429770644612.559,
    "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": 7429770644865.059, "dur": 5.291,
    "args": {
      "External id": 1140,      "device": 2, "stream": 0,      "correlation": 359, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 359, "pid": 2, "tid": 0, "ts": 7429770644865.059,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770644870.351, "dur": 252.700,
    "args": {
      "External id": 1148,      "device": 2, "stream": 0,      "correlation": 363, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 363, "pid": 2, "tid": 0, "ts": 7429770644870.351,
    "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": 7429770645123.052, "dur": 5.652,
    "args": {
      "External id": 1151,      "device": 2, "stream": 0,      "correlation": 364, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 364, "pid": 2, "tid": 0, "ts": 7429770645123.052,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770645128.705, "dur": 165.298,
    "args": {
      "External id": 1159,      "device": 2, "stream": 0,      "correlation": 368, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 368, "pid": 2, "tid": 0, "ts": 7429770645128.705,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770645294.004, "dur": 7.938,
    "args": {
      "External id": 1159,      "device": 2, "stream": 0,      "correlation": 369, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 369, "pid": 2, "tid": 0, "ts": 7429770645294.004,
    "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": 7429770645301.943, "dur": 4.730,
    "args": {
      "External id": 1162,      "device": 2, "stream": 0,      "correlation": 370, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 370, "pid": 2, "tid": 0, "ts": 7429770645301.943,
    "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": 7429770645306.674, "dur": 5.531,
    "args": {
      "External id": 1164,      "device": 2, "stream": 0,      "correlation": 371, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 371, "pid": 2, "tid": 0, "ts": 7429770645306.674,
    "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": 7429770645312.206, "dur": 7.978,
    "args": {
      "External id": 1167,      "device": 2, "stream": 0,      "correlation": 372, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 372, "pid": 2, "tid": 0, "ts": 7429770645312.206,
    "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": 7429770645320.185, "dur": 4.489,
    "args": {
      "External id": 1168,      "device": 2, "stream": 0,      "correlation": 373, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 373, "pid": 2, "tid": 0, "ts": 7429770645320.185,
    "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": 7429770645324.675, "dur": 4.609,
    "args": {
      "External id": 1169,      "device": 2, "stream": 0,      "correlation": 374, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 374, "pid": 2, "tid": 0, "ts": 7429770645324.675,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770645329.285, "dur": 4.971,
    "args": {
      "External id": 1170,      "device": 2, "stream": 0,      "correlation": 375, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 375, "pid": 2, "tid": 0, "ts": 7429770645329.285,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770645334.257, "dur": 5.532,
    "args": {
      "External id": 1172,      "device": 2, "stream": 0,      "correlation": 376, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 376, "pid": 2, "tid": 0, "ts": 7429770645334.257,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770645339.790, "dur": 63.905,
    "args": {
      "External id": 1180,      "device": 2, "stream": 0,      "correlation": 380, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 380, "pid": 2, "tid": 0, "ts": 7429770645339.790,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770645403.696, "dur": 50.796,
    "args": {
      "External id": 1193,      "device": 2, "stream": 0,      "correlation": 384, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 384, "pid": 2, "tid": 0, "ts": 7429770645403.696,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770645454.493, "dur": 49.513,
    "args": {
      "External id": 1206,      "device": 2, "stream": 0,      "correlation": 388, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 388, "pid": 2, "tid": 0, "ts": 7429770645454.493,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770645504.007, "dur": 5.251,
    "args": {
      "External id": 1216,      "device": 2, "stream": 0,      "correlation": 389, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 389, "pid": 2, "tid": 0, "ts": 7429770645504.007,
    "cat": "ac2g", "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": 7429770645509.259, "dur": 5.572,
    "args": {
      "External id": 1221,      "device": 2, "stream": 0,      "correlation": 390, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 390, "pid": 2, "tid": 0, "ts": 7429770645509.259,
    "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": 7429770645514.832, "dur": 6.213,
    "args": {
      "External id": 1222,      "device": 2, "stream": 0,      "correlation": 391, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 391, "pid": 2, "tid": 0, "ts": 7429770645514.832,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770645521.046, "dur": 4.730,
    "args": {
      "External id": 1223,      "device": 2, "stream": 0,      "correlation": 392, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 392, "pid": 2, "tid": 0, "ts": 7429770645521.046,
    "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::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": 7429770645525.777, "dur": 5.050,
    "args": {
      "External id": 1224,      "device": 2, "stream": 0,      "correlation": 393, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 393, "pid": 2, "tid": 0, "ts": 7429770645525.777,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770645530.828, "dur": 5.171,
    "args": {
      "External id": 1225,      "device": 2, "stream": 0,      "correlation": 394, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 394, "pid": 2, "tid": 0, "ts": 7429770645530.828,
    "cat": "ac2g", "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": 7429770645536.000, "dur": 5.211,
    "args": {
      "External id": 1230,      "device": 2, "stream": 0,      "correlation": 395, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 395, "pid": 2, "tid": 0, "ts": 7429770645536.000,
    "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": 7429770645541.212, "dur": 6.133,
    "args": {
      "External id": 1231,      "device": 2, "stream": 0,      "correlation": 396, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 396, "pid": 2, "tid": 0, "ts": 7429770645541.212,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770645547.346, "dur": 5.131,
    "args": {
      "External id": 1232,      "device": 2, "stream": 0,      "correlation": 397, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 397, "pid": 2, "tid": 0, "ts": 7429770645547.346,
    "cat": "ac2g", "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": 7429770645552.478, "dur": 5.011,
    "args": {
      "External id": 1233,      "device": 2, "stream": 0,      "correlation": 398, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 398, "pid": 2, "tid": 0, "ts": 7429770645552.478,
    "cat": "ac2g", "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": 7429770645557.490, "dur": 6.935,
    "args": {
      "External id": 1250,      "device": 2, "stream": 0,      "correlation": 399, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 399, "pid": 2, "tid": 0, "ts": 7429770645557.490,
    "cat": "ac2g", "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": 7429770645564.426, "dur": 5.451,
    "args": {
      "External id": 1268,      "device": 2, "stream": 0,      "correlation": 400, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 400, "pid": 2, "tid": 0, "ts": 7429770645564.426,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770645569.878, "dur": 9.862,
    "args": {
      "External id": 1281,      "device": 2, "stream": 0,      "correlation": 404, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 404, "pid": 2, "tid": 0, "ts": 7429770645569.878,
    "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": 7429770645579.741, "dur": 4.609,
    "args": {
      "External id": 1284,      "device": 2, "stream": 0,      "correlation": 405, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 405, "pid": 2, "tid": 0, "ts": 7429770645579.741,
    "cat": "ac2g", "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": 7429770645584.351, "dur": 4.770,
    "args": {
      "External id": 1291,      "device": 2, "stream": 0,      "correlation": 406, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 406, "pid": 2, "tid": 0, "ts": 7429770645584.351,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770645589.122, "dur": 5.612,
    "args": {
      "External id": 1294,      "device": 2, "stream": 0,      "correlation": 407, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 407, "pid": 2, "tid": 0, "ts": 7429770645589.122,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770645594.735, "dur": 11.105,
    "args": {
      "External id": 1306,      "device": 2, "stream": 0,      "correlation": 411, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 411, "pid": 2, "tid": 0, "ts": 7429770645594.735,
    "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::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::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": 7429770645605.841, "dur": 5.812,
    "args": {
      "External id": 1315,      "device": 2, "stream": 0,      "correlation": 412, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 412, "pid": 2, "tid": 0, "ts": 7429770645605.841,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770645611.654, "dur": 63.946,
    "args": {
      "External id": 1325,      "device": 2, "stream": 0,      "correlation": 416, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 416, "pid": 2, "tid": 0, "ts": 7429770645611.654,
    "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": 7429770645675.601, "dur": 5.251,
    "args": {
      "External id": 1328,      "device": 2, "stream": 0,      "correlation": 417, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 417, "pid": 2, "tid": 0, "ts": 7429770645675.601,
    "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": 7429770645680.853, "dur": 6.454,
    "args": {
      "External id": 1330,      "device": 2, "stream": 0,      "correlation": 418, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 418, "pid": 2, "tid": 0, "ts": 7429770645680.853,
    "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": 7429770645687.308, "dur": 7.977,
    "args": {
      "External id": 1333,      "device": 2, "stream": 0,      "correlation": 419, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 419, "pid": 2, "tid": 0, "ts": 7429770645687.308,
    "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": 7429770645695.286, "dur": 4.570,
    "args": {
      "External id": 1334,      "device": 2, "stream": 0,      "correlation": 420, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 420, "pid": 2, "tid": 0, "ts": 7429770645695.286,
    "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": 7429770645699.857, "dur": 5.371,
    "args": {
      "External id": 1335,      "device": 2, "stream": 0,      "correlation": 421, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 421, "pid": 2, "tid": 0, "ts": 7429770645699.857,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770645705.229, "dur": 5.371,
    "args": {
      "External id": 1336,      "device": 2, "stream": 0,      "correlation": 422, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 422, "pid": 2, "tid": 0, "ts": 7429770645705.229,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770645710.601, "dur": 5.051,
    "args": {
      "External id": 1338,      "device": 2, "stream": 0,      "correlation": 423, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 423, "pid": 2, "tid": 0, "ts": 7429770645710.601,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770645715.653, "dur": 252.659,
    "args": {
      "External id": 1346,      "device": 2, "stream": 0,      "correlation": 427, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 427, "pid": 2, "tid": 0, "ts": 7429770645715.653,
    "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": 7429770645968.313, "dur": 5.893,
    "args": {
      "External id": 1349,      "device": 2, "stream": 0,      "correlation": 428, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 428, "pid": 2, "tid": 0, "ts": 7429770645968.313,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770645974.207, "dur": 252.218,
    "args": {
      "External id": 1357,      "device": 2, "stream": 0,      "correlation": 432, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 432, "pid": 2, "tid": 0, "ts": 7429770645974.207,
    "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": 7429770646226.426, "dur": 6.093,
    "args": {
      "External id": 1360,      "device": 2, "stream": 0,      "correlation": 433, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 433, "pid": 2, "tid": 0, "ts": 7429770646226.426,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770646232.520, "dur": 165.780,
    "args": {
      "External id": 1368,      "device": 2, "stream": 0,      "correlation": 437, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 437, "pid": 2, "tid": 0, "ts": 7429770646232.520,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770646398.301, "dur": 7.175,
    "args": {
      "External id": 1368,      "device": 2, "stream": 0,      "correlation": 438, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 438, "pid": 2, "tid": 0, "ts": 7429770646398.301,
    "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": 7429770646405.477, "dur": 4.770,
    "args": {
      "External id": 1371,      "device": 2, "stream": 0,      "correlation": 439, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 439, "pid": 2, "tid": 0, "ts": 7429770646405.477,
    "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": 7429770646410.248, "dur": 6.093,
    "args": {
      "External id": 1373,      "device": 2, "stream": 0,      "correlation": 440, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 440, "pid": 2, "tid": 0, "ts": 7429770646410.248,
    "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": 7429770646416.342, "dur": 7.978,
    "args": {
      "External id": 1376,      "device": 2, "stream": 0,      "correlation": 441, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 441, "pid": 2, "tid": 0, "ts": 7429770646416.342,
    "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": 7429770646424.321, "dur": 5.170,
    "args": {
      "External id": 1377,      "device": 2, "stream": 0,      "correlation": 442, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 442, "pid": 2, "tid": 0, "ts": 7429770646424.321,
    "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": 7429770646429.492, "dur": 4.931,
    "args": {
      "External id": 1378,      "device": 2, "stream": 0,      "correlation": 443, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 443, "pid": 2, "tid": 0, "ts": 7429770646429.492,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770646434.424, "dur": 4.850,
    "args": {
      "External id": 1379,      "device": 2, "stream": 0,      "correlation": 444, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 444, "pid": 2, "tid": 0, "ts": 7429770646434.424,
    "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": 7429770646439.275, "dur": 5.331,
    "args": {
      "External id": 1381,      "device": 2, "stream": 0,      "correlation": 445, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 445, "pid": 2, "tid": 0, "ts": 7429770646439.275,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770646444.607, "dur": 64.307,
    "args": {
      "External id": 1389,      "device": 2, "stream": 0,      "correlation": 449, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 449, "pid": 2, "tid": 0, "ts": 7429770646444.607,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770646508.915, "dur": 50.354,
    "args": {
      "External id": 1402,      "device": 2, "stream": 0,      "correlation": 453, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 453, "pid": 2, "tid": 0, "ts": 7429770646508.915,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770646559.270, "dur": 49.674,
    "args": {
      "External id": 1415,      "device": 2, "stream": 0,      "correlation": 457, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 457, "pid": 2, "tid": 0, "ts": 7429770646559.270,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770646608.945, "dur": 4.689,
    "args": {
      "External id": 1425,      "device": 2, "stream": 0,      "correlation": 458, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 458, "pid": 2, "tid": 0, "ts": 7429770646608.945,
    "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::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": 7429770646613.635, "dur": 5.452,
    "args": {
      "External id": 1430,      "device": 2, "stream": 0,      "correlation": 459, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 459, "pid": 2, "tid": 0, "ts": 7429770646613.635,
    "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": 7429770646619.088, "dur": 6.494,
    "args": {
      "External id": 1431,      "device": 2, "stream": 0,      "correlation": 460, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 460, "pid": 2, "tid": 0, "ts": 7429770646619.088,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770646625.583, "dur": 4.649,
    "args": {
      "External id": 1432,      "device": 2, "stream": 0,      "correlation": 461, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 461, "pid": 2, "tid": 0, "ts": 7429770646625.583,
    "cat": "ac2g", "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": 7429770646630.233, "dur": 5.011,
    "args": {
      "External id": 1433,      "device": 2, "stream": 0,      "correlation": 462, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 462, "pid": 2, "tid": 0, "ts": 7429770646630.233,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770646635.245, "dur": 4.850,
    "args": {
      "External id": 1434,      "device": 2, "stream": 0,      "correlation": 463, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 463, "pid": 2, "tid": 0, "ts": 7429770646635.245,
    "cat": "ac2g", "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": 7429770646640.096, "dur": 5.532,
    "args": {
      "External id": 1439,      "device": 2, "stream": 0,      "correlation": 464, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 464, "pid": 2, "tid": 0, "ts": 7429770646640.096,
    "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": 7429770646645.629, "dur": 6.213,
    "args": {
      "External id": 1440,      "device": 2, "stream": 0,      "correlation": 465, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 465, "pid": 2, "tid": 0, "ts": 7429770646645.629,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770646651.843, "dur": 5.251,
    "args": {
      "External id": 1441,      "device": 2, "stream": 0,      "correlation": 466, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 466, "pid": 2, "tid": 0, "ts": 7429770646651.843,
    "cat": "ac2g", "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": 7429770646657.095, "dur": 5.291,
    "args": {
      "External id": 1442,      "device": 2, "stream": 0,      "correlation": 467, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 467, "pid": 2, "tid": 0, "ts": 7429770646657.095,
    "cat": "ac2g", "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": 7429770646662.387, "dur": 5.131,
    "args": {
      "External id": 1459,      "device": 2, "stream": 0,      "correlation": 468, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 468, "pid": 2, "tid": 0, "ts": 7429770646662.387,
    "cat": "ac2g", "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": 7429770646667.519, "dur": 5.893,
    "args": {
      "External id": 1477,      "device": 2, "stream": 0,      "correlation": 469, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 469, "pid": 2, "tid": 0, "ts": 7429770646667.519,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770646673.413, "dur": 9.941,
    "args": {
      "External id": 1490,      "device": 2, "stream": 0,      "correlation": 473, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 473, "pid": 2, "tid": 0, "ts": 7429770646673.413,
    "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": 7429770646683.355, "dur": 4.530,
    "args": {
      "External id": 1493,      "device": 2, "stream": 0,      "correlation": 474, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 474, "pid": 2, "tid": 0, "ts": 7429770646683.355,
    "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::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": 7429770646687.886, "dur": 4.690,
    "args": {
      "External id": 1500,      "device": 2, "stream": 0,      "correlation": 475, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 475, "pid": 2, "tid": 0, "ts": 7429770646687.886,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770646692.577, "dur": 5.612,
    "args": {
      "External id": 1503,      "device": 2, "stream": 0,      "correlation": 476, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 476, "pid": 2, "tid": 0, "ts": 7429770646692.577,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770646698.190, "dur": 10.904,
    "args": {
      "External id": 1515,      "device": 2, "stream": 0,      "correlation": 480, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 480, "pid": 2, "tid": 0, "ts": 7429770646698.190,
    "cat": "ac2g", "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": 7429770646709.095, "dur": 5.010,
    "args": {
      "External id": 1524,      "device": 2, "stream": 0,      "correlation": 481, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 481, "pid": 2, "tid": 0, "ts": 7429770646709.095,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770646714.106, "dur": 64.587,
    "args": {
      "External id": 1534,      "device": 2, "stream": 0,      "correlation": 485, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 485, "pid": 2, "tid": 0, "ts": 7429770646714.106,
    "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": 7429770646778.694, "dur": 4.610,
    "args": {
      "External id": 1537,      "device": 2, "stream": 0,      "correlation": 486, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 486, "pid": 2, "tid": 0, "ts": 7429770646778.694,
    "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": 7429770646783.305, "dur": 5.131,
    "args": {
      "External id": 1539,      "device": 2, "stream": 0,      "correlation": 487, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 487, "pid": 2, "tid": 0, "ts": 7429770646783.305,
    "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": 7429770646788.437, "dur": 8.097,
    "args": {
      "External id": 1542,      "device": 2, "stream": 0,      "correlation": 488, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 488, "pid": 2, "tid": 0, "ts": 7429770646788.437,
    "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": 7429770646796.535, "dur": 4.490,
    "args": {
      "External id": 1543,      "device": 2, "stream": 0,      "correlation": 489, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 489, "pid": 2, "tid": 0, "ts": 7429770646796.535,
    "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": 7429770646801.026, "dur": 5.211,
    "args": {
      "External id": 1544,      "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": 7429770646801.026,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770646806.238, "dur": 5.291,
    "args": {
      "External id": 1545,      "device": 2, "stream": 0,      "correlation": 491, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 491, "pid": 2, "tid": 0, "ts": 7429770646806.238,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770646811.530, "dur": 4.810,
    "args": {
      "External id": 1547,      "device": 2, "stream": 0,      "correlation": 492, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 492, "pid": 2, "tid": 0, "ts": 7429770646811.530,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770646816.341, "dur": 253.020,
    "args": {
      "External id": 1555,      "device": 2, "stream": 0,      "correlation": 498, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 498, "pid": 2, "tid": 0, "ts": 7429770646816.341,
    "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": 7429770647069.362, "dur": 5.652,
    "args": {
      "External id": 1558,      "device": 2, "stream": 0,      "correlation": 499, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 499, "pid": 2, "tid": 0, "ts": 7429770647069.362,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770647075.015, "dur": 252.699,
    "args": {
      "External id": 1566,      "device": 2, "stream": 0,      "correlation": 503, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 503, "pid": 2, "tid": 0, "ts": 7429770647075.015,
    "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": 7429770647327.715, "dur": 5.733,
    "args": {
      "External id": 1569,      "device": 2, "stream": 0,      "correlation": 504, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 504, "pid": 2, "tid": 0, "ts": 7429770647327.715,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770647333.449, "dur": 166.100,
    "args": {
      "External id": 1577,      "device": 2, "stream": 0,      "correlation": 508, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 508, "pid": 2, "tid": 0, "ts": 7429770647333.449,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770647499.550, "dur": 7.336,
    "args": {
      "External id": 1577,      "device": 2, "stream": 0,      "correlation": 509, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 509, "pid": 2, "tid": 0, "ts": 7429770647499.550,
    "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": 7429770647506.887, "dur": 4.650,
    "args": {
      "External id": 1580,      "device": 2, "stream": 0,      "correlation": 510, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 510, "pid": 2, "tid": 0, "ts": 7429770647506.887,
    "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": 7429770647511.538, "dur": 5.692,
    "args": {
      "External id": 1582,      "device": 2, "stream": 0,      "correlation": 511, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 511, "pid": 2, "tid": 0, "ts": 7429770647511.538,
    "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": 7429770647517.231, "dur": 8.057,
    "args": {
      "External id": 1585,      "device": 2, "stream": 0,      "correlation": 512, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 512, "pid": 2, "tid": 0, "ts": 7429770647517.231,
    "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": 7429770647525.289, "dur": 4.409,
    "args": {
      "External id": 1586,      "device": 2, "stream": 0,      "correlation": 513, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 513, "pid": 2, "tid": 0, "ts": 7429770647525.289,
    "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": 7429770647529.699, "dur": 3.928,
    "args": {
      "External id": 1587,      "device": 2, "stream": 0,      "correlation": 514, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 514, "pid": 2, "tid": 0, "ts": 7429770647529.699,
    "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": 7429770647539.201, "dur": 4.209,
    "args": {
      "External id": 1588,      "device": 2, "stream": 0,      "correlation": 515, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 515, "pid": 2, "tid": 0, "ts": 7429770647539.201,
    "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": 7429770647543.411, "dur": 5.010,
    "args": {
      "External id": 1590,      "device": 2, "stream": 0,      "correlation": 516, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 516, "pid": 2, "tid": 0, "ts": 7429770647543.411,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770647548.422, "dur": 63.986,
    "args": {
      "External id": 1598,      "device": 2, "stream": 0,      "correlation": 520, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 520, "pid": 2, "tid": 0, "ts": 7429770647548.422,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770647612.409, "dur": 49.994,
    "args": {
      "External id": 1611,      "device": 2, "stream": 0,      "correlation": 524, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 524, "pid": 2, "tid": 0, "ts": 7429770647612.409,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770647662.404, "dur": 49.713,
    "args": {
      "External id": 1624,      "device": 2, "stream": 0,      "correlation": 528, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 528, "pid": 2, "tid": 0, "ts": 7429770647662.404,
    "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": 7429770647712.118, "dur": 5.452,
    "args": {
      "External id": 1634,      "device": 2, "stream": 0,      "correlation": 529, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 529, "pid": 2, "tid": 0, "ts": 7429770647712.118,
    "cat": "ac2g", "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": 7429770647717.571, "dur": 4.650,
    "args": {
      "External id": 1639,      "device": 2, "stream": 0,      "correlation": 530, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 530, "pid": 2, "tid": 0, "ts": 7429770647717.571,
    "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": 7429770647722.222, "dur": 6.373,
    "args": {
      "External id": 1640,      "device": 2, "stream": 0,      "correlation": 531, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 531, "pid": 2, "tid": 0, "ts": 7429770647722.222,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770647728.596, "dur": 5.131,
    "args": {
      "External id": 1641,      "device": 2, "stream": 0,      "correlation": 532, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 532, "pid": 2, "tid": 0, "ts": 7429770647728.596,
    "cat": "ac2g", "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": 7429770647733.728, "dur": 5.492,
    "args": {
      "External id": 1642,      "device": 2, "stream": 0,      "correlation": 533, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 533, "pid": 2, "tid": 0, "ts": 7429770647733.728,
    "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": 7429770647739.221, "dur": 5.130,
    "args": {
      "External id": 1643,      "device": 2, "stream": 0,      "correlation": 534, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 534, "pid": 2, "tid": 0, "ts": 7429770647739.221,
    "cat": "ac2g", "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": 7429770647744.352, "dur": 4.570,
    "args": {
      "External id": 1648,      "device": 2, "stream": 0,      "correlation": 535, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 535, "pid": 2, "tid": 0, "ts": 7429770647744.352,
    "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": 7429770647748.923, "dur": 6.133,
    "args": {
      "External id": 1649,      "device": 2, "stream": 0,      "correlation": 536, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 536, "pid": 2, "tid": 0, "ts": 7429770647748.923,
    "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": 7429770647755.057, "dur": 4.409,
    "args": {
      "External id": 1650,      "device": 2, "stream": 0,      "correlation": 537, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 537, "pid": 2, "tid": 0, "ts": 7429770647755.057,
    "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": 7429770647759.467, "dur": 4.730,
    "args": {
      "External id": 1651,      "device": 2, "stream": 0,      "correlation": 538, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 538, "pid": 2, "tid": 0, "ts": 7429770647759.467,
    "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::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::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": 7429770647764.198, "dur": 6.133,
    "args": {
      "External id": 1668,      "device": 2, "stream": 0,      "correlation": 539, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 539, "pid": 2, "tid": 0, "ts": 7429770647764.198,
    "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": 7429770647770.332, "dur": 5.492,
    "args": {
      "External id": 1686,      "device": 2, "stream": 0,      "correlation": 540, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 540, "pid": 2, "tid": 0, "ts": 7429770647770.332,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770647775.825, "dur": 9.621,
    "args": {
      "External id": 1699,      "device": 2, "stream": 0,      "correlation": 544, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 544, "pid": 2, "tid": 0, "ts": 7429770647775.825,
    "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": 7429770647785.447, "dur": 5.211,
    "args": {
      "External id": 1702,      "device": 2, "stream": 0,      "correlation": 545, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 545, "pid": 2, "tid": 0, "ts": 7429770647785.447,
    "cat": "ac2g", "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": 7429770647790.659, "dur": 5.211,
    "args": {
      "External id": 1709,      "device": 2, "stream": 0,      "correlation": 546, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 546, "pid": 2, "tid": 0, "ts": 7429770647790.659,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770647795.871, "dur": 5.171,
    "args": {
      "External id": 1712,      "device": 2, "stream": 0,      "correlation": 547, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 547, "pid": 2, "tid": 0, "ts": 7429770647795.871,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770647801.043, "dur": 10.944,
    "args": {
      "External id": 1724,      "device": 2, "stream": 0,      "correlation": 551, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 551, "pid": 2, "tid": 0, "ts": 7429770647801.043,
    "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::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::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": 7429770647811.988, "dur": 4.970,
    "args": {
      "External id": 1733,      "device": 2, "stream": 0,      "correlation": 552, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 552, "pid": 2, "tid": 0, "ts": 7429770647811.988,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770647816.959, "dur": 62.703,
    "args": {
      "External id": 1743,      "device": 2, "stream": 0,      "correlation": 556, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 556, "pid": 2, "tid": 0, "ts": 7429770647816.959,
    "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": 7429770647879.663, "dur": 5.452,
    "args": {
      "External id": 1746,      "device": 2, "stream": 0,      "correlation": 557, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 557, "pid": 2, "tid": 0, "ts": 7429770647879.663,
    "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": 7429770647885.116, "dur": 5.972,
    "args": {
      "External id": 1748,      "device": 2, "stream": 0,      "correlation": 558, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 558, "pid": 2, "tid": 0, "ts": 7429770647885.116,
    "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": 7429770647891.089, "dur": 7.857,
    "args": {
      "External id": 1751,      "device": 2, "stream": 0,      "correlation": 559, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 559, "pid": 2, "tid": 0, "ts": 7429770647891.089,
    "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": 7429770647898.947, "dur": 5.051,
    "args": {
      "External id": 1752,      "device": 2, "stream": 0,      "correlation": 560, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 560, "pid": 2, "tid": 0, "ts": 7429770647898.947,
    "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": 7429770647903.999, "dur": 4.970,
    "args": {
      "External id": 1753,      "device": 2, "stream": 0,      "correlation": 561, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 561, "pid": 2, "tid": 0, "ts": 7429770647903.999,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770647908.970, "dur": 5.011,
    "args": {
      "External id": 1754,      "device": 2, "stream": 0,      "correlation": 562, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 562, "pid": 2, "tid": 0, "ts": 7429770647908.970,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770647913.982, "dur": 5.251,
    "args": {
      "External id": 1756,      "device": 2, "stream": 0,      "correlation": 563, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 563, "pid": 2, "tid": 0, "ts": 7429770647913.982,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770647919.234, "dur": 253.140,
    "args": {
      "External id": 1764,      "device": 2, "stream": 0,      "correlation": 567, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 567, "pid": 2, "tid": 0, "ts": 7429770647919.234,
    "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": 7429770648172.375, "dur": 5.492,
    "args": {
      "External id": 1767,      "device": 2, "stream": 0,      "correlation": 568, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 568, "pid": 2, "tid": 0, "ts": 7429770648172.375,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770648177.868, "dur": 252.299,
    "args": {
      "External id": 1775,      "device": 2, "stream": 0,      "correlation": 572, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 572, "pid": 2, "tid": 0, "ts": 7429770648177.868,
    "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": 7429770648430.168, "dur": 5.932,
    "args": {
      "External id": 1778,      "device": 2, "stream": 0,      "correlation": 573, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 573, "pid": 2, "tid": 0, "ts": 7429770648430.168,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770648436.101, "dur": 164.738,
    "args": {
      "External id": 1786,      "device": 2, "stream": 0,      "correlation": 577, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 577, "pid": 2, "tid": 0, "ts": 7429770648436.101,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770648600.840, "dur": 7.496,
    "args": {
      "External id": 1786,      "device": 2, "stream": 0,      "correlation": 578, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 578, "pid": 2, "tid": 0, "ts": 7429770648600.840,
    "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": 7429770648608.337, "dur": 4.930,
    "args": {
      "External id": 1789,      "device": 2, "stream": 0,      "correlation": 579, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 579, "pid": 2, "tid": 0, "ts": 7429770648608.337,
    "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": 7429770648613.268, "dur": 6.213,
    "args": {
      "External id": 1791,      "device": 2, "stream": 0,      "correlation": 580, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 580, "pid": 2, "tid": 0, "ts": 7429770648613.268,
    "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": 7429770648619.482, "dur": 7.978,
    "args": {
      "External id": 1794,      "device": 2, "stream": 0,      "correlation": 581, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 581, "pid": 2, "tid": 0, "ts": 7429770648619.482,
    "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": 7429770648627.461, "dur": 4.850,
    "args": {
      "External id": 1795,      "device": 2, "stream": 0,      "correlation": 582, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 582, "pid": 2, "tid": 0, "ts": 7429770648627.461,
    "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": 7429770648632.312, "dur": 4.650,
    "args": {
      "External id": 1796,      "device": 2, "stream": 0,      "correlation": 583, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 583, "pid": 2, "tid": 0, "ts": 7429770648632.312,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770648636.963, "dur": 4.769,
    "args": {
      "External id": 1797,      "device": 2, "stream": 0,      "correlation": 584, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 584, "pid": 2, "tid": 0, "ts": 7429770648636.963,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770648641.733, "dur": 4.971,
    "args": {
      "External id": 1799,      "device": 2, "stream": 0,      "correlation": 585, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 585, "pid": 2, "tid": 0, "ts": 7429770648641.733,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770648646.705, "dur": 64.026,
    "args": {
      "External id": 1807,      "device": 2, "stream": 0,      "correlation": 589, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 589, "pid": 2, "tid": 0, "ts": 7429770648646.705,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770648710.732, "dur": 50.435,
    "args": {
      "External id": 1820,      "device": 2, "stream": 0,      "correlation": 593, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 593, "pid": 2, "tid": 0, "ts": 7429770648710.732,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770648761.168, "dur": 49.793,
    "args": {
      "External id": 1833,      "device": 2, "stream": 0,      "correlation": 597, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 597, "pid": 2, "tid": 0, "ts": 7429770648761.168,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770648810.962, "dur": 4.930,
    "args": {
      "External id": 1843,      "device": 2, "stream": 0,      "correlation": 598, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 598, "pid": 2, "tid": 0, "ts": 7429770648810.962,
    "cat": "ac2g", "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": 7429770648815.893, "dur": 5.332,
    "args": {
      "External id": 1848,      "device": 2, "stream": 0,      "correlation": 599, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 599, "pid": 2, "tid": 0, "ts": 7429770648815.893,
    "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": 7429770648821.226, "dur": 6.293,
    "args": {
      "External id": 1849,      "device": 2, "stream": 0,      "correlation": 600, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 600, "pid": 2, "tid": 0, "ts": 7429770648821.226,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770648827.520, "dur": 4.971,
    "args": {
      "External id": 1850,      "device": 2, "stream": 0,      "correlation": 601, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 601, "pid": 2, "tid": 0, "ts": 7429770648827.520,
    "cat": "ac2g", "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": 7429770648832.492, "dur": 5.010,
    "args": {
      "External id": 1851,      "device": 2, "stream": 0,      "correlation": 602, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 602, "pid": 2, "tid": 0, "ts": 7429770648832.492,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770648837.503, "dur": 5.452,
    "args": {
      "External id": 1852,      "device": 2, "stream": 0,      "correlation": 603, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 603, "pid": 2, "tid": 0, "ts": 7429770648837.503,
    "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": 7429770648842.956, "dur": 5.050,
    "args": {
      "External id": 1857,      "device": 2, "stream": 0,      "correlation": 604, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 604, "pid": 2, "tid": 0, "ts": 7429770648842.956,
    "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": 7429770648848.007, "dur": 6.254,
    "args": {
      "External id": 1858,      "device": 2, "stream": 0,      "correlation": 605, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 605, "pid": 2, "tid": 0, "ts": 7429770648848.007,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770648854.262, "dur": 5.130,
    "args": {
      "External id": 1859,      "device": 2, "stream": 0,      "correlation": 606, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 606, "pid": 2, "tid": 0, "ts": 7429770648854.262,
    "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": 7429770648859.393, "dur": 4.971,
    "args": {
      "External id": 1860,      "device": 2, "stream": 0,      "correlation": 607, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 607, "pid": 2, "tid": 0, "ts": 7429770648859.393,
    "cat": "ac2g", "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": 7429770648864.365, "dur": 5.211,
    "args": {
      "External id": 1877,      "device": 2, "stream": 0,      "correlation": 608, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 608, "pid": 2, "tid": 0, "ts": 7429770648864.365,
    "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::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::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": 7429770648869.577, "dur": 5.852,
    "args": {
      "External id": 1895,      "device": 2, "stream": 0,      "correlation": 609, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 609, "pid": 2, "tid": 0, "ts": 7429770648869.577,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770648875.430, "dur": 9.862,
    "args": {
      "External id": 1908,      "device": 2, "stream": 0,      "correlation": 613, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 613, "pid": 2, "tid": 0, "ts": 7429770648875.430,
    "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": 7429770648885.293, "dur": 4.529,
    "args": {
      "External id": 1911,      "device": 2, "stream": 0,      "correlation": 614, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 614, "pid": 2, "tid": 0, "ts": 7429770648885.293,
    "cat": "ac2g", "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": 7429770648889.823, "dur": 4.891,
    "args": {
      "External id": 1918,      "device": 2, "stream": 0,      "correlation": 615, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 615, "pid": 2, "tid": 0, "ts": 7429770648889.823,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770648894.715, "dur": 5.932,
    "args": {
      "External id": 1921,      "device": 2, "stream": 0,      "correlation": 616, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 616, "pid": 2, "tid": 0, "ts": 7429770648894.715,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770648900.648, "dur": 10.864,
    "args": {
      "External id": 1933,      "device": 2, "stream": 0,      "correlation": 620, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 620, "pid": 2, "tid": 0, "ts": 7429770648900.648,
    "cat": "ac2g", "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": 7429770648911.513, "dur": 5.412,
    "args": {
      "External id": 1942,      "device": 2, "stream": 0,      "correlation": 621, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 621, "pid": 2, "tid": 0, "ts": 7429770648911.513,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770648916.926, "dur": 63.945,
    "args": {
      "External id": 1952,      "device": 2, "stream": 0,      "correlation": 625, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 625, "pid": 2, "tid": 0, "ts": 7429770648916.926,
    "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": 7429770648980.872, "dur": 4.730,
    "args": {
      "External id": 1955,      "device": 2, "stream": 0,      "correlation": 626, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 626, "pid": 2, "tid": 0, "ts": 7429770648980.872,
    "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": 7429770648985.603, "dur": 5.131,
    "args": {
      "External id": 1957,      "device": 2, "stream": 0,      "correlation": 627, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 627, "pid": 2, "tid": 0, "ts": 7429770648985.603,
    "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": 7429770648990.735, "dur": 7.977,
    "args": {
      "External id": 1960,      "device": 2, "stream": 0,      "correlation": 628, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 628, "pid": 2, "tid": 0, "ts": 7429770648990.735,
    "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": 7429770648998.713, "dur": 4.770,
    "args": {
      "External id": 1961,      "device": 2, "stream": 0,      "correlation": 629, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 629, "pid": 2, "tid": 0, "ts": 7429770648998.713,
    "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": 7429770649003.484, "dur": 4.690,
    "args": {
      "External id": 1962,      "device": 2, "stream": 0,      "correlation": 630, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 630, "pid": 2, "tid": 0, "ts": 7429770649003.484,
    "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": 7429770649008.175, "dur": 4.850,
    "args": {
      "External id": 1963,      "device": 2, "stream": 0,      "correlation": 631, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 631, "pid": 2, "tid": 0, "ts": 7429770649008.175,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770649013.026, "dur": 5.372,
    "args": {
      "External id": 1965,      "device": 2, "stream": 0,      "correlation": 632, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 632, "pid": 2, "tid": 0, "ts": 7429770649013.026,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770649018.399, "dur": 253.381,
    "args": {
      "External id": 1973,      "device": 2, "stream": 0,      "correlation": 636, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 636, "pid": 2, "tid": 0, "ts": 7429770649018.399,
    "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": 7429770649271.781, "dur": 5.411,
    "args": {
      "External id": 1976,      "device": 2, "stream": 0,      "correlation": 637, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 637, "pid": 2, "tid": 0, "ts": 7429770649271.781,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770649277.193, "dur": 252.419,
    "args": {
      "External id": 1984,      "device": 2, "stream": 0,      "correlation": 641, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 641, "pid": 2, "tid": 0, "ts": 7429770649277.193,
    "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": 7429770649529.613, "dur": 5.692,
    "args": {
      "External id": 1987,      "device": 2, "stream": 0,      "correlation": 642, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 642, "pid": 2, "tid": 0, "ts": 7429770649529.613,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770649535.306, "dur": 164.376,
    "args": {
      "External id": 1995,      "device": 2, "stream": 0,      "correlation": 648, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 648, "pid": 2, "tid": 0, "ts": 7429770649535.306,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770649699.683, "dur": 7.416,
    "args": {
      "External id": 1995,      "device": 2, "stream": 0,      "correlation": 649, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 649, "pid": 2, "tid": 0, "ts": 7429770649699.683,
    "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": 7429770649707.100, "dur": 4.610,
    "args": {
      "External id": 1998,      "device": 2, "stream": 0,      "correlation": 650, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 650, "pid": 2, "tid": 0, "ts": 7429770649707.100,
    "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": 7429770649711.711, "dur": 5.051,
    "args": {
      "External id": 2000,      "device": 2, "stream": 0,      "correlation": 651, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 651, "pid": 2, "tid": 0, "ts": 7429770649711.711,
    "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": 7429770649716.763, "dur": 8.057,
    "args": {
      "External id": 2003,      "device": 2, "stream": 0,      "correlation": 652, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 652, "pid": 2, "tid": 0, "ts": 7429770649716.763,
    "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": 7429770649724.821, "dur": 5.051,
    "args": {
      "External id": 2004,      "device": 2, "stream": 0,      "correlation": 653, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 653, "pid": 2, "tid": 0, "ts": 7429770649724.821,
    "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": 7429770649729.873, "dur": 5.171,
    "args": {
      "External id": 2005,      "device": 2, "stream": 0,      "correlation": 654, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 654, "pid": 2, "tid": 0, "ts": 7429770649729.873,
    "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": 7429770649735.045, "dur": 5.050,
    "args": {
      "External id": 2006,      "device": 2, "stream": 0,      "correlation": 655, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 655, "pid": 2, "tid": 0, "ts": 7429770649735.045,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770649740.096, "dur": 5.291,
    "args": {
      "External id": 2008,      "device": 2, "stream": 0,      "correlation": 656, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 656, "pid": 2, "tid": 0, "ts": 7429770649740.096,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770649745.388, "dur": 63.786,
    "args": {
      "External id": 2016,      "device": 2, "stream": 0,      "correlation": 660, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 660, "pid": 2, "tid": 0, "ts": 7429770649745.388,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770649809.175, "dur": 49.954,
    "args": {
      "External id": 2029,      "device": 2, "stream": 0,      "correlation": 664, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 664, "pid": 2, "tid": 0, "ts": 7429770649809.175,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770649859.130, "dur": 50.234,
    "args": {
      "External id": 2042,      "device": 2, "stream": 0,      "correlation": 668, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 668, "pid": 2, "tid": 0, "ts": 7429770649859.130,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770649909.365, "dur": 5.291,
    "args": {
      "External id": 2052,      "device": 2, "stream": 0,      "correlation": 669, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 669, "pid": 2, "tid": 0, "ts": 7429770649909.365,
    "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::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": 7429770649914.657, "dur": 4.810,
    "args": {
      "External id": 2057,      "device": 2, "stream": 0,      "correlation": 670, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 670, "pid": 2, "tid": 0, "ts": 7429770649914.657,
    "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": 7429770649919.468, "dur": 6.213,
    "args": {
      "External id": 2058,      "device": 2, "stream": 0,      "correlation": 671, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 671, "pid": 2, "tid": 0, "ts": 7429770649919.468,
    "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": 7429770649925.682, "dur": 4.770,
    "args": {
      "External id": 2059,      "device": 2, "stream": 0,      "correlation": 672, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 672, "pid": 2, "tid": 0, "ts": 7429770649925.682,
    "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::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": 7429770649930.453, "dur": 5.091,
    "args": {
      "External id": 2060,      "device": 2, "stream": 0,      "correlation": 673, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 673, "pid": 2, "tid": 0, "ts": 7429770649930.453,
    "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": 7429770649935.545, "dur": 5.091,
    "args": {
      "External id": 2061,      "device": 2, "stream": 0,      "correlation": 674, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 674, "pid": 2, "tid": 0, "ts": 7429770649935.545,
    "cat": "ac2g", "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": 7429770649940.637, "dur": 4.890,
    "args": {
      "External id": 2066,      "device": 2, "stream": 0,      "correlation": 675, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 675, "pid": 2, "tid": 0, "ts": 7429770649940.637,
    "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": 7429770649945.528, "dur": 6.293,
    "args": {
      "External id": 2067,      "device": 2, "stream": 0,      "correlation": 676, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 676, "pid": 2, "tid": 0, "ts": 7429770649945.528,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770649951.822, "dur": 5.091,
    "args": {
      "External id": 2068,      "device": 2, "stream": 0,      "correlation": 677, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 677, "pid": 2, "tid": 0, "ts": 7429770649951.822,
    "cat": "ac2g", "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": 7429770649956.914, "dur": 5.692,
    "args": {
      "External id": 2069,      "device": 2, "stream": 0,      "correlation": 678, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 678, "pid": 2, "tid": 0, "ts": 7429770649956.914,
    "cat": "ac2g", "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": 7429770649962.607, "dur": 5.692,
    "args": {
      "External id": 2086,      "device": 2, "stream": 0,      "correlation": 679, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 679, "pid": 2, "tid": 0, "ts": 7429770649962.607,
    "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::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::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": 7429770649968.300, "dur": 5.532,
    "args": {
      "External id": 2104,      "device": 2, "stream": 0,      "correlation": 680, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 680, "pid": 2, "tid": 0, "ts": 7429770649968.300,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770649973.833, "dur": 9.781,
    "args": {
      "External id": 2117,      "device": 2, "stream": 0,      "correlation": 684, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 684, "pid": 2, "tid": 0, "ts": 7429770649973.833,
    "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": 7429770649983.615, "dur": 5.372,
    "args": {
      "External id": 2120,      "device": 2, "stream": 0,      "correlation": 685, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 685, "pid": 2, "tid": 0, "ts": 7429770649983.615,
    "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::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": 7429770649988.988, "dur": 5.572,
    "args": {
      "External id": 2127,      "device": 2, "stream": 0,      "correlation": 686, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 686, "pid": 2, "tid": 0, "ts": 7429770649988.988,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770649994.561, "dur": 4.970,
    "args": {
      "External id": 2130,      "device": 2, "stream": 0,      "correlation": 687, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 687, "pid": 2, "tid": 0, "ts": 7429770649994.561,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770649999.532, "dur": 10.944,
    "args": {
      "External id": 2142,      "device": 2, "stream": 0,      "correlation": 691, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 691, "pid": 2, "tid": 0, "ts": 7429770649999.532,
    "cat": "ac2g", "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": 7429770650010.477, "dur": 5.091,
    "args": {
      "External id": 2151,      "device": 2, "stream": 0,      "correlation": 692, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 692, "pid": 2, "tid": 0, "ts": 7429770650010.477,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770650015.569, "dur": 63.705,
    "args": {
      "External id": 2161,      "device": 2, "stream": 0,      "correlation": 696, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 696, "pid": 2, "tid": 0, "ts": 7429770650015.569,
    "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": 7429770650079.275, "dur": 4.730,
    "args": {
      "External id": 2164,      "device": 2, "stream": 0,      "correlation": 697, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 697, "pid": 2, "tid": 0, "ts": 7429770650079.275,
    "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": 7429770650084.006, "dur": 5.051,
    "args": {
      "External id": 2166,      "device": 2, "stream": 0,      "correlation": 698, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 698, "pid": 2, "tid": 0, "ts": 7429770650084.006,
    "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": 7429770650089.058, "dur": 7.937,
    "args": {
      "External id": 2169,      "device": 2, "stream": 0,      "correlation": 699, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 699, "pid": 2, "tid": 0, "ts": 7429770650089.058,
    "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": 7429770650096.996, "dur": 4.489,
    "args": {
      "External id": 2170,      "device": 2, "stream": 0,      "correlation": 700, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 700, "pid": 2, "tid": 0, "ts": 7429770650096.996,
    "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": 7429770650101.486, "dur": 4.570,
    "args": {
      "External id": 2171,      "device": 2, "stream": 0,      "correlation": 701, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 701, "pid": 2, "tid": 0, "ts": 7429770650101.486,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770650106.057, "dur": 4.730,
    "args": {
      "External id": 2172,      "device": 2, "stream": 0,      "correlation": 702, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 702, "pid": 2, "tid": 0, "ts": 7429770650106.057,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770650110.788, "dur": 4.930,
    "args": {
      "External id": 2174,      "device": 2, "stream": 0,      "correlation": 703, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 703, "pid": 2, "tid": 0, "ts": 7429770650110.788,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770650115.719, "dur": 252.860,
    "args": {
      "External id": 2182,      "device": 2, "stream": 0,      "correlation": 707, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 707, "pid": 2, "tid": 0, "ts": 7429770650115.719,
    "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": 7429770650368.580, "dur": 5.612,
    "args": {
      "External id": 2185,      "device": 2, "stream": 0,      "correlation": 708, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 708, "pid": 2, "tid": 0, "ts": 7429770650368.580,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770650374.193, "dur": 252.940,
    "args": {
      "External id": 2193,      "device": 2, "stream": 0,      "correlation": 712, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 712, "pid": 2, "tid": 0, "ts": 7429770650374.193,
    "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": 7429770650627.134, "dur": 5.411,
    "args": {
      "External id": 2196,      "device": 2, "stream": 0,      "correlation": 713, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 713, "pid": 2, "tid": 0, "ts": 7429770650627.134,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770650632.546, "dur": 165.058,
    "args": {
      "External id": 2204,      "device": 2, "stream": 0,      "correlation": 717, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 717, "pid": 2, "tid": 0, "ts": 7429770650632.546,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770650797.605, "dur": 7.095,
    "args": {
      "External id": 2204,      "device": 2, "stream": 0,      "correlation": 718, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 718, "pid": 2, "tid": 0, "ts": 7429770650797.605,
    "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": 7429770650804.701, "dur": 4.850,
    "args": {
      "External id": 2207,      "device": 2, "stream": 0,      "correlation": 719, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 719, "pid": 2, "tid": 0, "ts": 7429770650804.701,
    "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": 7429770650809.552, "dur": 6.174,
    "args": {
      "External id": 2209,      "device": 2, "stream": 0,      "correlation": 720, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 720, "pid": 2, "tid": 0, "ts": 7429770650809.552,
    "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": 7429770650815.727, "dur": 8.137,
    "args": {
      "External id": 2212,      "device": 2, "stream": 0,      "correlation": 721, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 721, "pid": 2, "tid": 0, "ts": 7429770650815.727,
    "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": 7429770650823.865, "dur": 4.490,
    "args": {
      "External id": 2213,      "device": 2, "stream": 0,      "correlation": 722, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 722, "pid": 2, "tid": 0, "ts": 7429770650823.865,
    "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": 7429770650828.356, "dur": 4.489,
    "args": {
      "External id": 2214,      "device": 2, "stream": 0,      "correlation": 723, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 723, "pid": 2, "tid": 0, "ts": 7429770650828.356,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770650832.846, "dur": 4.770,
    "args": {
      "External id": 2215,      "device": 2, "stream": 0,      "correlation": 724, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 724, "pid": 2, "tid": 0, "ts": 7429770650832.846,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770650837.617, "dur": 5.010,
    "args": {
      "External id": 2217,      "device": 2, "stream": 0,      "correlation": 725, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 725, "pid": 2, "tid": 0, "ts": 7429770650837.617,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770650842.628, "dur": 65.590,
    "args": {
      "External id": 2225,      "device": 2, "stream": 0,      "correlation": 729, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 729, "pid": 2, "tid": 0, "ts": 7429770650842.628,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770650908.219, "dur": 50.355,
    "args": {
      "External id": 2238,      "device": 2, "stream": 0,      "correlation": 733, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 733, "pid": 2, "tid": 0, "ts": 7429770650908.219,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770650958.575, "dur": 50.234,
    "args": {
      "External id": 2251,      "device": 2, "stream": 0,      "correlation": 737, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 737, "pid": 2, "tid": 0, "ts": 7429770650958.575,
    "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": 7429770651008.810, "dur": 4.770,
    "args": {
      "External id": 2261,      "device": 2, "stream": 0,      "correlation": 738, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 738, "pid": 2, "tid": 0, "ts": 7429770651008.810,
    "cat": "ac2g", "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": 7429770651013.581, "dur": 4.770,
    "args": {
      "External id": 2266,      "device": 2, "stream": 0,      "correlation": 739, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 739, "pid": 2, "tid": 0, "ts": 7429770651013.581,
    "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": 7429770651018.352, "dur": 6.213,
    "args": {
      "External id": 2267,      "device": 2, "stream": 0,      "correlation": 740, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 740, "pid": 2, "tid": 0, "ts": 7429770651018.352,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770651024.566, "dur": 5.051,
    "args": {
      "External id": 2268,      "device": 2, "stream": 0,      "correlation": 741, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 741, "pid": 2, "tid": 0, "ts": 7429770651024.566,
    "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::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": 7429770651029.618, "dur": 5.492,
    "args": {
      "External id": 2269,      "device": 2, "stream": 0,      "correlation": 742, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 742, "pid": 2, "tid": 0, "ts": 7429770651029.618,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770651035.111, "dur": 5.170,
    "args": {
      "External id": 2270,      "device": 2, "stream": 0,      "correlation": 743, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 743, "pid": 2, "tid": 0, "ts": 7429770651035.111,
    "cat": "ac2g", "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": 7429770651040.282, "dur": 4.730,
    "args": {
      "External id": 2275,      "device": 2, "stream": 0,      "correlation": 744, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 744, "pid": 2, "tid": 0, "ts": 7429770651040.282,
    "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": 7429770651045.013, "dur": 6.294,
    "args": {
      "External id": 2276,      "device": 2, "stream": 0,      "correlation": 745, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 745, "pid": 2, "tid": 0, "ts": 7429770651045.013,
    "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": 7429770651051.308, "dur": 4.489,
    "args": {
      "External id": 2277,      "device": 2, "stream": 0,      "correlation": 746, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 746, "pid": 2, "tid": 0, "ts": 7429770651051.308,
    "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::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": 7429770651055.798, "dur": 4.690,
    "args": {
      "External id": 2278,      "device": 2, "stream": 0,      "correlation": 747, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 747, "pid": 2, "tid": 0, "ts": 7429770651055.798,
    "cat": "ac2g", "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": 7429770651060.489, "dur": 4.970,
    "args": {
      "External id": 2295,      "device": 2, "stream": 0,      "correlation": 748, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 748, "pid": 2, "tid": 0, "ts": 7429770651060.489,
    "cat": "ac2g", "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": 7429770651065.460, "dur": 5.412,
    "args": {
      "External id": 2313,      "device": 2, "stream": 0,      "correlation": 749, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 749, "pid": 2, "tid": 0, "ts": 7429770651065.460,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770651070.873, "dur": 9.380,
    "args": {
      "External id": 2326,      "device": 2, "stream": 0,      "correlation": 753, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 753, "pid": 2, "tid": 0, "ts": 7429770651070.873,
    "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": 7429770651080.254, "dur": 4.610,
    "args": {
      "External id": 2329,      "device": 2, "stream": 0,      "correlation": 754, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 754, "pid": 2, "tid": 0, "ts": 7429770651080.254,
    "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::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": 7429770651084.865, "dur": 4.770,
    "args": {
      "External id": 2336,      "device": 2, "stream": 0,      "correlation": 755, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 755, "pid": 2, "tid": 0, "ts": 7429770651084.865,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770651089.636, "dur": 5.411,
    "args": {
      "External id": 2339,      "device": 2, "stream": 0,      "correlation": 756, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 756, "pid": 2, "tid": 0, "ts": 7429770651089.636,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770651095.048, "dur": 10.944,
    "args": {
      "External id": 2351,      "device": 2, "stream": 0,      "correlation": 760, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 760, "pid": 2, "tid": 0, "ts": 7429770651095.048,
    "cat": "ac2g", "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": 7429770651105.993, "dur": 5.131,
    "args": {
      "External id": 2360,      "device": 2, "stream": 0,      "correlation": 761, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 761, "pid": 2, "tid": 0, "ts": 7429770651105.993,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770651111.125, "dur": 63.465,
    "args": {
      "External id": 2370,      "device": 2, "stream": 0,      "correlation": 765, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 765, "pid": 2, "tid": 0, "ts": 7429770651111.125,
    "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": 7429770651174.591, "dur": 4.650,
    "args": {
      "External id": 2373,      "device": 2, "stream": 0,      "correlation": 766, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 766, "pid": 2, "tid": 0, "ts": 7429770651174.591,
    "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": 7429770651179.242, "dur": 5.411,
    "args": {
      "External id": 2375,      "device": 2, "stream": 0,      "correlation": 767, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 767, "pid": 2, "tid": 0, "ts": 7429770651179.242,
    "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": 7429770651184.654, "dur": 7.857,
    "args": {
      "External id": 2378,      "device": 2, "stream": 0,      "correlation": 768, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 768, "pid": 2, "tid": 0, "ts": 7429770651184.654,
    "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": 7429770651192.512, "dur": 5.091,
    "args": {
      "External id": 2379,      "device": 2, "stream": 0,      "correlation": 769, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 769, "pid": 2, "tid": 0, "ts": 7429770651192.512,
    "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": 7429770651197.604, "dur": 5.211,
    "args": {
      "External id": 2380,      "device": 2, "stream": 0,      "correlation": 770, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 770, "pid": 2, "tid": 0, "ts": 7429770651197.604,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770651202.816, "dur": 4.810,
    "args": {
      "External id": 2381,      "device": 2, "stream": 0,      "correlation": 771, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 771, "pid": 2, "tid": 0, "ts": 7429770651202.816,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770651207.627, "dur": 5.131,
    "args": {
      "External id": 2383,      "device": 2, "stream": 0,      "correlation": 772, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 772, "pid": 2, "tid": 0, "ts": 7429770651207.627,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770651212.759, "dur": 252.499,
    "args": {
      "External id": 2391,      "device": 2, "stream": 0,      "correlation": 776, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 776, "pid": 2, "tid": 0, "ts": 7429770651212.759,
    "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": 7429770651465.259, "dur": 5.210,
    "args": {
      "External id": 2394,      "device": 2, "stream": 0,      "correlation": 777, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 777, "pid": 2, "tid": 0, "ts": 7429770651465.259,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770651470.470, "dur": 252.700,
    "args": {
      "External id": 2402,      "device": 2, "stream": 0,      "correlation": 781, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 781, "pid": 2, "tid": 0, "ts": 7429770651470.470,
    "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": 7429770651723.171, "dur": 5.652,
    "args": {
      "External id": 2405,      "device": 2, "stream": 0,      "correlation": 782, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 782, "pid": 2, "tid": 0, "ts": 7429770651723.171,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770651728.824, "dur": 164.898,
    "args": {
      "External id": 2413,      "device": 2, "stream": 0,      "correlation": 786, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 786, "pid": 2, "tid": 0, "ts": 7429770651728.824,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770651893.723, "dur": 7.496,
    "args": {
      "External id": 2413,      "device": 2, "stream": 0,      "correlation": 787, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 787, "pid": 2, "tid": 0, "ts": 7429770651893.723,
    "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": 7429770651901.220, "dur": 5.010,
    "args": {
      "External id": 2416,      "device": 2, "stream": 0,      "correlation": 788, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 788, "pid": 2, "tid": 0, "ts": 7429770651901.220,
    "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": 7429770651906.231, "dur": 4.730,
    "args": {
      "External id": 2418,      "device": 2, "stream": 0,      "correlation": 789, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 789, "pid": 2, "tid": 0, "ts": 7429770651906.231,
    "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": 7429770651910.962, "dur": 7.937,
    "args": {
      "External id": 2421,      "device": 2, "stream": 0,      "correlation": 790, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 790, "pid": 2, "tid": 0, "ts": 7429770651910.962,
    "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": 7429770651918.900, "dur": 5.091,
    "args": {
      "External id": 2422,      "device": 2, "stream": 0,      "correlation": 791, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 791, "pid": 2, "tid": 0, "ts": 7429770651918.900,
    "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": 7429770651923.992, "dur": 4.930,
    "args": {
      "External id": 2423,      "device": 2, "stream": 0,      "correlation": 792, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 792, "pid": 2, "tid": 0, "ts": 7429770651923.992,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770651928.923, "dur": 4.851,
    "args": {
      "External id": 2424,      "device": 2, "stream": 0,      "correlation": 793, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 793, "pid": 2, "tid": 0, "ts": 7429770651928.923,
    "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": 7429770651933.775, "dur": 5.451,
    "args": {
      "External id": 2426,      "device": 2, "stream": 0,      "correlation": 794, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 794, "pid": 2, "tid": 0, "ts": 7429770651933.775,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770651939.227, "dur": 64.507,
    "args": {
      "External id": 2434,      "device": 2, "stream": 0,      "correlation": 798, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 798, "pid": 2, "tid": 0, "ts": 7429770651939.227,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770652003.735, "dur": 50.956,
    "args": {
      "External id": 2447,      "device": 2, "stream": 0,      "correlation": 802, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 802, "pid": 2, "tid": 0, "ts": 7429770652003.735,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770652054.692, "dur": 50.716,
    "args": {
      "External id": 2460,      "device": 2, "stream": 0,      "correlation": 806, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 806, "pid": 2, "tid": 0, "ts": 7429770652054.692,
    "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": 7429770652105.409, "dur": 4.770,
    "args": {
      "External id": 2470,      "device": 2, "stream": 0,      "correlation": 807, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 807, "pid": 2, "tid": 0, "ts": 7429770652105.409,
    "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": 7429770652110.180, "dur": 4.690,
    "args": {
      "External id": 2475,      "device": 2, "stream": 0,      "correlation": 808, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 808, "pid": 2, "tid": 0, "ts": 7429770652110.180,
    "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": 7429770652114.871, "dur": 6.293,
    "args": {
      "External id": 2476,      "device": 2, "stream": 0,      "correlation": 809, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 809, "pid": 2, "tid": 0, "ts": 7429770652114.871,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770652121.165, "dur": 4.970,
    "args": {
      "External id": 2477,      "device": 2, "stream": 0,      "correlation": 810, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 810, "pid": 2, "tid": 0, "ts": 7429770652121.165,
    "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::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": 7429770652126.136, "dur": 4.931,
    "args": {
      "External id": 2478,      "device": 2, "stream": 0,      "correlation": 811, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 811, "pid": 2, "tid": 0, "ts": 7429770652126.136,
    "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": 7429770652131.068, "dur": 4.569,
    "args": {
      "External id": 2479,      "device": 2, "stream": 0,      "correlation": 812, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 812, "pid": 2, "tid": 0, "ts": 7429770652131.068,
    "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::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": 7429770652135.638, "dur": 4.409,
    "args": {
      "External id": 2484,      "device": 2, "stream": 0,      "correlation": 813, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 813, "pid": 2, "tid": 0, "ts": 7429770652135.638,
    "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": 7429770652140.048, "dur": 5.933,
    "args": {
      "External id": 2485,      "device": 2, "stream": 0,      "correlation": 814, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 814, "pid": 2, "tid": 0, "ts": 7429770652140.048,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770652145.982, "dur": 4.850,
    "args": {
      "External id": 2486,      "device": 2, "stream": 0,      "correlation": 815, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 815, "pid": 2, "tid": 0, "ts": 7429770652145.982,
    "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": 7429770652150.833, "dur": 4.850,
    "args": {
      "External id": 2487,      "device": 2, "stream": 0,      "correlation": 816, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 816, "pid": 2, "tid": 0, "ts": 7429770652150.833,
    "cat": "ac2g", "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": 7429770652155.684, "dur": 5.773,
    "args": {
      "External id": 2504,      "device": 2, "stream": 0,      "correlation": 817, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 817, "pid": 2, "tid": 0, "ts": 7429770652155.684,
    "cat": "ac2g", "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": 7429770652161.458, "dur": 5.972,
    "args": {
      "External id": 2522,      "device": 2, "stream": 0,      "correlation": 818, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 818, "pid": 2, "tid": 0, "ts": 7429770652161.458,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770652167.431, "dur": 9.501,
    "args": {
      "External id": 2535,      "device": 2, "stream": 0,      "correlation": 822, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 822, "pid": 2, "tid": 0, "ts": 7429770652167.431,
    "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": 7429770652176.933, "dur": 4.890,
    "args": {
      "External id": 2538,      "device": 2, "stream": 0,      "correlation": 823, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 823, "pid": 2, "tid": 0, "ts": 7429770652176.933,
    "cat": "ac2g", "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": 7429770652181.824, "dur": 4.850,
    "args": {
      "External id": 2545,      "device": 2, "stream": 0,      "correlation": 824, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 824, "pid": 2, "tid": 0, "ts": 7429770652181.824,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770652186.675, "dur": 5.091,
    "args": {
      "External id": 2548,      "device": 2, "stream": 0,      "correlation": 825, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 825, "pid": 2, "tid": 0, "ts": 7429770652186.675,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770652191.767, "dur": 10.784,
    "args": {
      "External id": 2560,      "device": 2, "stream": 0,      "correlation": 831, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 831, "pid": 2, "tid": 0, "ts": 7429770652191.767,
    "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::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::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": 7429770652202.552, "dur": 5.051,
    "args": {
      "External id": 2569,      "device": 2, "stream": 0,      "correlation": 832, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 832, "pid": 2, "tid": 0, "ts": 7429770652202.552,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770652207.604, "dur": 63.985,
    "args": {
      "External id": 2579,      "device": 2, "stream": 0,      "correlation": 836, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 836, "pid": 2, "tid": 0, "ts": 7429770652207.604,
    "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": 7429770652271.590, "dur": 4.811,
    "args": {
      "External id": 2582,      "device": 2, "stream": 0,      "correlation": 837, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 837, "pid": 2, "tid": 0, "ts": 7429770652271.590,
    "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": 7429770652276.402, "dur": 5.451,
    "args": {
      "External id": 2584,      "device": 2, "stream": 0,      "correlation": 838, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 838, "pid": 2, "tid": 0, "ts": 7429770652276.402,
    "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": 7429770652281.854, "dur": 8.017,
    "args": {
      "External id": 2587,      "device": 2, "stream": 0,      "correlation": 839, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 839, "pid": 2, "tid": 0, "ts": 7429770652281.854,
    "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": 7429770652289.872, "dur": 4.730,
    "args": {
      "External id": 2588,      "device": 2, "stream": 0,      "correlation": 840, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 840, "pid": 2, "tid": 0, "ts": 7429770652289.872,
    "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": 7429770652294.603, "dur": 4.650,
    "args": {
      "External id": 2589,      "device": 2, "stream": 0,      "correlation": 841, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 841, "pid": 2, "tid": 0, "ts": 7429770652294.603,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770652299.254, "dur": 4.770,
    "args": {
      "External id": 2590,      "device": 2, "stream": 0,      "correlation": 842, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 842, "pid": 2, "tid": 0, "ts": 7429770652299.254,
    "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": 7429770652304.025, "dur": 5.010,
    "args": {
      "External id": 2592,      "device": 2, "stream": 0,      "correlation": 843, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 843, "pid": 2, "tid": 0, "ts": 7429770652304.025,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770652309.036, "dur": 252.740,
    "args": {
      "External id": 2600,      "device": 2, "stream": 0,      "correlation": 847, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 847, "pid": 2, "tid": 0, "ts": 7429770652309.036,
    "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": 7429770652561.777, "dur": 5.492,
    "args": {
      "External id": 2603,      "device": 2, "stream": 0,      "correlation": 848, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 848, "pid": 2, "tid": 0, "ts": 7429770652561.777,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770652567.270, "dur": 252.539,
    "args": {
      "External id": 2611,      "device": 2, "stream": 0,      "correlation": 852, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 852, "pid": 2, "tid": 0, "ts": 7429770652567.270,
    "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": 7429770652819.810, "dur": 5.652,
    "args": {
      "External id": 2614,      "device": 2, "stream": 0,      "correlation": 853, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 853, "pid": 2, "tid": 0, "ts": 7429770652819.810,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770652825.463, "dur": 165.419,
    "args": {
      "External id": 2622,      "device": 2, "stream": 0,      "correlation": 857, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 857, "pid": 2, "tid": 0, "ts": 7429770652825.463,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770652990.883, "dur": 7.095,
    "args": {
      "External id": 2622,      "device": 2, "stream": 0,      "correlation": 858, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 858, "pid": 2, "tid": 0, "ts": 7429770652990.883,
    "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": 7429770652997.979, "dur": 4.609,
    "args": {
      "External id": 2625,      "device": 2, "stream": 0,      "correlation": 859, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 859, "pid": 2, "tid": 0, "ts": 7429770652997.979,
    "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": 7429770653002.589, "dur": 5.652,
    "args": {
      "External id": 2627,      "device": 2, "stream": 0,      "correlation": 860, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 860, "pid": 2, "tid": 0, "ts": 7429770653002.589,
    "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": 7429770653008.242, "dur": 8.178,
    "args": {
      "External id": 2630,      "device": 2, "stream": 0,      "correlation": 861, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 861, "pid": 2, "tid": 0, "ts": 7429770653008.242,
    "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": 7429770653016.421, "dur": 4.530,
    "args": {
      "External id": 2631,      "device": 2, "stream": 0,      "correlation": 862, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 862, "pid": 2, "tid": 0, "ts": 7429770653016.421,
    "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": 7429770653020.952, "dur": 4.649,
    "args": {
      "External id": 2632,      "device": 2, "stream": 0,      "correlation": 863, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 863, "pid": 2, "tid": 0, "ts": 7429770653020.952,
    "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": 7429770653025.602, "dur": 4.770,
    "args": {
      "External id": 2633,      "device": 2, "stream": 0,      "correlation": 864, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 864, "pid": 2, "tid": 0, "ts": 7429770653025.602,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770653030.373, "dur": 5.011,
    "args": {
      "External id": 2635,      "device": 2, "stream": 0,      "correlation": 865, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 865, "pid": 2, "tid": 0, "ts": 7429770653030.373,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770653035.385, "dur": 64.868,
    "args": {
      "External id": 2643,      "device": 2, "stream": 0,      "correlation": 869, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 869, "pid": 2, "tid": 0, "ts": 7429770653035.385,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770653100.254, "dur": 50.715,
    "args": {
      "External id": 2656,      "device": 2, "stream": 0,      "correlation": 873, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 873, "pid": 2, "tid": 0, "ts": 7429770653100.254,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770653150.970, "dur": 50.756,
    "args": {
      "External id": 2669,      "device": 2, "stream": 0,      "correlation": 877, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 877, "pid": 2, "tid": 0, "ts": 7429770653150.970,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770653201.727, "dur": 4.810,
    "args": {
      "External id": 2679,      "device": 2, "stream": 0,      "correlation": 878, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 878, "pid": 2, "tid": 0, "ts": 7429770653201.727,
    "cat": "ac2g", "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": 7429770653206.538, "dur": 4.810,
    "args": {
      "External id": 2684,      "device": 2, "stream": 0,      "correlation": 879, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 879, "pid": 2, "tid": 0, "ts": 7429770653206.538,
    "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": 7429770653211.349, "dur": 6.253,
    "args": {
      "External id": 2685,      "device": 2, "stream": 0,      "correlation": 880, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 880, "pid": 2, "tid": 0, "ts": 7429770653211.349,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770653217.603, "dur": 4.850,
    "args": {
      "External id": 2686,      "device": 2, "stream": 0,      "correlation": 881, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 881, "pid": 2, "tid": 0, "ts": 7429770653217.603,
    "cat": "ac2g", "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": 7429770653222.454, "dur": 5.211,
    "args": {
      "External id": 2687,      "device": 2, "stream": 0,      "correlation": 882, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 882, "pid": 2, "tid": 0, "ts": 7429770653222.454,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770653227.666, "dur": 4.971,
    "args": {
      "External id": 2688,      "device": 2, "stream": 0,      "correlation": 883, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 883, "pid": 2, "tid": 0, "ts": 7429770653227.666,
    "cat": "ac2g", "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": 7429770653232.638, "dur": 4.609,
    "args": {
      "External id": 2693,      "device": 2, "stream": 0,      "correlation": 884, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 884, "pid": 2, "tid": 0, "ts": 7429770653232.638,
    "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": 7429770653237.248, "dur": 6.133,
    "args": {
      "External id": 2694,      "device": 2, "stream": 0,      "correlation": 885, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 885, "pid": 2, "tid": 0, "ts": 7429770653237.248,
    "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": 7429770653243.382, "dur": 4.690,
    "args": {
      "External id": 2695,      "device": 2, "stream": 0,      "correlation": 886, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 886, "pid": 2, "tid": 0, "ts": 7429770653243.382,
    "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::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": 7429770653248.073, "dur": 4.810,
    "args": {
      "External id": 2696,      "device": 2, "stream": 0,      "correlation": 887, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 887, "pid": 2, "tid": 0, "ts": 7429770653248.073,
    "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::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::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": 7429770653252.884, "dur": 7.497,
    "args": {
      "External id": 2713,      "device": 2, "stream": 0,      "correlation": 888, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 888, "pid": 2, "tid": 0, "ts": 7429770653252.884,
    "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::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::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": 7429770653260.382, "dur": 5.531,
    "args": {
      "External id": 2731,      "device": 2, "stream": 0,      "correlation": 889, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 889, "pid": 2, "tid": 0, "ts": 7429770653260.382,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770653265.914, "dur": 9.541,
    "args": {
      "External id": 2744,      "device": 2, "stream": 0,      "correlation": 893, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 893, "pid": 2, "tid": 0, "ts": 7429770653265.914,
    "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": 7429770653275.456, "dur": 4.690,
    "args": {
      "External id": 2747,      "device": 2, "stream": 0,      "correlation": 894, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 894, "pid": 2, "tid": 0, "ts": 7429770653275.456,
    "cat": "ac2g", "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": 7429770653280.147, "dur": 4.810,
    "args": {
      "External id": 2754,      "device": 2, "stream": 0,      "correlation": 895, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 895, "pid": 2, "tid": 0, "ts": 7429770653280.147,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770653284.958, "dur": 5.131,
    "args": {
      "External id": 2757,      "device": 2, "stream": 0,      "correlation": 896, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 896, "pid": 2, "tid": 0, "ts": 7429770653284.958,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770653290.090, "dur": 10.784,
    "args": {
      "External id": 2769,      "device": 2, "stream": 0,      "correlation": 900, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 900, "pid": 2, "tid": 0, "ts": 7429770653290.090,
    "cat": "ac2g", "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": 7429770653300.875, "dur": 5.050,
    "args": {
      "External id": 2778,      "device": 2, "stream": 0,      "correlation": 901, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 901, "pid": 2, "tid": 0, "ts": 7429770653300.875,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770653305.926, "dur": 64.187,
    "args": {
      "External id": 2788,      "device": 2, "stream": 0,      "correlation": 905, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 905, "pid": 2, "tid": 0, "ts": 7429770653305.926,
    "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": 7429770653370.114, "dur": 4.850,
    "args": {
      "External id": 2791,      "device": 2, "stream": 0,      "correlation": 906, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 906, "pid": 2, "tid": 0, "ts": 7429770653370.114,
    "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": 7429770653374.965, "dur": 5.571,
    "args": {
      "External id": 2793,      "device": 2, "stream": 0,      "correlation": 907, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 907, "pid": 2, "tid": 0, "ts": 7429770653374.965,
    "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": 7429770653380.537, "dur": 8.018,
    "args": {
      "External id": 2796,      "device": 2, "stream": 0,      "correlation": 908, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 908, "pid": 2, "tid": 0, "ts": 7429770653380.537,
    "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": 7429770653388.556, "dur": 4.449,
    "args": {
      "External id": 2797,      "device": 2, "stream": 0,      "correlation": 909, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 909, "pid": 2, "tid": 0, "ts": 7429770653388.556,
    "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": 7429770653393.006, "dur": 4.570,
    "args": {
      "External id": 2798,      "device": 2, "stream": 0,      "correlation": 910, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 910, "pid": 2, "tid": 0, "ts": 7429770653393.006,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770653397.577, "dur": 4.810,
    "args": {
      "External id": 2799,      "device": 2, "stream": 0,      "correlation": 911, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 911, "pid": 2, "tid": 0, "ts": 7429770653397.577,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770653402.388, "dur": 4.930,
    "args": {
      "External id": 2801,      "device": 2, "stream": 0,      "correlation": 912, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 912, "pid": 2, "tid": 0, "ts": 7429770653402.388,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770653407.319, "dur": 253.100,
    "args": {
      "External id": 2809,      "device": 2, "stream": 0,      "correlation": 916, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 916, "pid": 2, "tid": 0, "ts": 7429770653407.319,
    "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": 7429770653660.420, "dur": 5.853,
    "args": {
      "External id": 2812,      "device": 2, "stream": 0,      "correlation": 917, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 917, "pid": 2, "tid": 0, "ts": 7429770653660.420,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770653666.274, "dur": 252.819,
    "args": {
      "External id": 2820,      "device": 2, "stream": 0,      "correlation": 921, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 921, "pid": 2, "tid": 0, "ts": 7429770653666.274,
    "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": 7429770653919.094, "dur": 5.733,
    "args": {
      "External id": 2823,      "device": 2, "stream": 0,      "correlation": 922, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 922, "pid": 2, "tid": 0, "ts": 7429770653919.094,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770653924.828, "dur": 165.499,
    "args": {
      "External id": 2831,      "device": 2, "stream": 0,      "correlation": 926, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 926, "pid": 2, "tid": 0, "ts": 7429770653924.828,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770654090.328, "dur": 7.456,
    "args": {
      "External id": 2831,      "device": 2, "stream": 0,      "correlation": 927, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 927, "pid": 2, "tid": 0, "ts": 7429770654090.328,
    "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": 7429770654097.785, "dur": 5.010,
    "args": {
      "External id": 2834,      "device": 2, "stream": 0,      "correlation": 928, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 928, "pid": 2, "tid": 0, "ts": 7429770654097.785,
    "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": 7429770654102.796, "dur": 6.454,
    "args": {
      "External id": 2836,      "device": 2, "stream": 0,      "correlation": 929, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 929, "pid": 2, "tid": 0, "ts": 7429770654102.796,
    "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": 7429770654109.251, "dur": 7.817,
    "args": {
      "External id": 2839,      "device": 2, "stream": 0,      "correlation": 930, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 930, "pid": 2, "tid": 0, "ts": 7429770654109.251,
    "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": 7429770654117.069, "dur": 5.051,
    "args": {
      "External id": 2840,      "device": 2, "stream": 0,      "correlation": 931, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 931, "pid": 2, "tid": 0, "ts": 7429770654117.069,
    "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": 7429770654122.121, "dur": 5.291,
    "args": {
      "External id": 2841,      "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": 7429770654122.121,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770654127.413, "dur": 5.171,
    "args": {
      "External id": 2842,      "device": 2, "stream": 0,      "correlation": 933, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 933, "pid": 2, "tid": 0, "ts": 7429770654127.413,
    "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": 7429770654132.585, "dur": 4.930,
    "args": {
      "External id": 2844,      "device": 2, "stream": 0,      "correlation": 934, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 934, "pid": 2, "tid": 0, "ts": 7429770654132.585,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770654137.516, "dur": 63.746,
    "args": {
      "External id": 2852,      "device": 2, "stream": 0,      "correlation": 938, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 938, "pid": 2, "tid": 0, "ts": 7429770654137.516,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770654201.263, "dur": 50.114,
    "args": {
      "External id": 2865,      "device": 2, "stream": 0,      "correlation": 942, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 942, "pid": 2, "tid": 0, "ts": 7429770654201.263,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770654251.378, "dur": 49.633,
    "args": {
      "External id": 2878,      "device": 2, "stream": 0,      "correlation": 946, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 946, "pid": 2, "tid": 0, "ts": 7429770654251.378,
    "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": 7429770654301.012, "dur": 5.531,
    "args": {
      "External id": 2888,      "device": 2, "stream": 0,      "correlation": 947, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 947, "pid": 2, "tid": 0, "ts": 7429770654301.012,
    "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::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": 7429770654306.544, "dur": 4.770,
    "args": {
      "External id": 2893,      "device": 2, "stream": 0,      "correlation": 948, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 948, "pid": 2, "tid": 0, "ts": 7429770654306.544,
    "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": 7429770654311.315, "dur": 6.254,
    "args": {
      "External id": 2894,      "device": 2, "stream": 0,      "correlation": 949, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 949, "pid": 2, "tid": 0, "ts": 7429770654311.315,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770654317.570, "dur": 4.970,
    "args": {
      "External id": 2895,      "device": 2, "stream": 0,      "correlation": 950, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 950, "pid": 2, "tid": 0, "ts": 7429770654317.570,
    "cat": "ac2g", "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": 7429770654322.541, "dur": 5.532,
    "args": {
      "External id": 2896,      "device": 2, "stream": 0,      "correlation": 951, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 951, "pid": 2, "tid": 0, "ts": 7429770654322.541,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770654328.074, "dur": 5.491,
    "args": {
      "External id": 2897,      "device": 2, "stream": 0,      "correlation": 952, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 952, "pid": 2, "tid": 0, "ts": 7429770654328.074,
    "cat": "ac2g", "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": 7429770654333.566, "dur": 5.131,
    "args": {
      "External id": 2902,      "device": 2, "stream": 0,      "correlation": 953, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 953, "pid": 2, "tid": 0, "ts": 7429770654333.566,
    "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": 7429770654338.698, "dur": 6.013,
    "args": {
      "External id": 2903,      "device": 2, "stream": 0,      "correlation": 954, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 954, "pid": 2, "tid": 0, "ts": 7429770654338.698,
    "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": 7429770654344.712, "dur": 4.930,
    "args": {
      "External id": 2904,      "device": 2, "stream": 0,      "correlation": 955, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 955, "pid": 2, "tid": 0, "ts": 7429770654344.712,
    "cat": "ac2g", "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": 7429770654349.643, "dur": 4.850,
    "args": {
      "External id": 2905,      "device": 2, "stream": 0,      "correlation": 956, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 956, "pid": 2, "tid": 0, "ts": 7429770654349.643,
    "cat": "ac2g", "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": 7429770654354.494, "dur": 6.815,
    "args": {
      "External id": 2922,      "device": 2, "stream": 0,      "correlation": 957, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 957, "pid": 2, "tid": 0, "ts": 7429770654354.494,
    "cat": "ac2g", "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": 7429770654361.310, "dur": 5.371,
    "args": {
      "External id": 2940,      "device": 2, "stream": 0,      "correlation": 958, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 958, "pid": 2, "tid": 0, "ts": 7429770654361.310,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770654366.682, "dur": 9.742,
    "args": {
      "External id": 2953,      "device": 2, "stream": 0,      "correlation": 962, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 962, "pid": 2, "tid": 0, "ts": 7429770654366.682,
    "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": 7429770654376.425, "dur": 5.211,
    "args": {
      "External id": 2956,      "device": 2, "stream": 0,      "correlation": 963, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 963, "pid": 2, "tid": 0, "ts": 7429770654376.425,
    "cat": "ac2g", "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": 7429770654381.637, "dur": 4.770,
    "args": {
      "External id": 2963,      "device": 2, "stream": 0,      "correlation": 964, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 964, "pid": 2, "tid": 0, "ts": 7429770654381.637,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770654386.408, "dur": 5.090,
    "args": {
      "External id": 2966,      "device": 2, "stream": 0,      "correlation": 965, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 965, "pid": 2, "tid": 0, "ts": 7429770654386.408,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770654391.499, "dur": 10.905,
    "args": {
      "External id": 2978,      "device": 2, "stream": 0,      "correlation": 969, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 969, "pid": 2, "tid": 0, "ts": 7429770654391.499,
    "cat": "ac2g", "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": 7429770654402.405, "dur": 4.970,
    "args": {
      "External id": 2987,      "device": 2, "stream": 0,      "correlation": 970, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 970, "pid": 2, "tid": 0, "ts": 7429770654402.405,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770654407.376, "dur": 63.826,
    "args": {
      "External id": 2997,      "device": 2, "stream": 0,      "correlation": 974, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 974, "pid": 2, "tid": 0, "ts": 7429770654407.376,
    "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": 7429770654471.203, "dur": 5.611,
    "args": {
      "External id": 3000,      "device": 2, "stream": 0,      "correlation": 975, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 975, "pid": 2, "tid": 0, "ts": 7429770654471.203,
    "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": 7429770654476.815, "dur": 6.053,
    "args": {
      "External id": 3002,      "device": 2, "stream": 0,      "correlation": 976, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 976, "pid": 2, "tid": 0, "ts": 7429770654476.815,
    "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": 7429770654482.869, "dur": 7.897,
    "args": {
      "External id": 3005,      "device": 2, "stream": 0,      "correlation": 977, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 977, "pid": 2, "tid": 0, "ts": 7429770654482.869,
    "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": 7429770654490.767, "dur": 4.530,
    "args": {
      "External id": 3006,      "device": 2, "stream": 0,      "correlation": 978, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 978, "pid": 2, "tid": 0, "ts": 7429770654490.767,
    "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": 7429770654495.298, "dur": 4.569,
    "args": {
      "External id": 3007,      "device": 2, "stream": 0,      "correlation": 979, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 979, "pid": 2, "tid": 0, "ts": 7429770654495.298,
    "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": 7429770654499.868, "dur": 4.851,
    "args": {
      "External id": 3008,      "device": 2, "stream": 0,      "correlation": 980, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 980, "pid": 2, "tid": 0, "ts": 7429770654499.868,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770654504.720, "dur": 4.970,
    "args": {
      "External id": 3010,      "device": 2, "stream": 0,      "correlation": 981, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 981, "pid": 2, "tid": 0, "ts": 7429770654504.720,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770654509.691, "dur": 252.980,
    "args": {
      "External id": 3018,      "device": 2, "stream": 0,      "correlation": 985, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 985, "pid": 2, "tid": 0, "ts": 7429770654509.691,
    "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": 7429770654762.672, "dur": 5.411,
    "args": {
      "External id": 3021,      "device": 2, "stream": 0,      "correlation": 986, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 986, "pid": 2, "tid": 0, "ts": 7429770654762.672,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770654768.084, "dur": 253.101,
    "args": {
      "External id": 3029,      "device": 2, "stream": 0,      "correlation": 992, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 992, "pid": 2, "tid": 0, "ts": 7429770654768.084,
    "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": 7429770655021.186, "dur": 5.491,
    "args": {
      "External id": 3032,      "device": 2, "stream": 0,      "correlation": 993, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 993, "pid": 2, "tid": 0, "ts": 7429770655021.186,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770655026.678, "dur": 164.738,
    "args": {
      "External id": 3040,      "device": 2, "stream": 0,      "correlation": 997, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 997, "pid": 2, "tid": 0, "ts": 7429770655026.678,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770655191.417, "dur": 7.336,
    "args": {
      "External id": 3040,      "device": 2, "stream": 0,      "correlation": 998, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 998, "pid": 2, "tid": 0, "ts": 7429770655191.417,
    "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": 7429770655198.754, "dur": 5.371,
    "args": {
      "External id": 3043,      "device": 2, "stream": 0,      "correlation": 999, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 999, "pid": 2, "tid": 0, "ts": 7429770655198.754,
    "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": 7429770655204.126, "dur": 5.692,
    "args": {
      "External id": 3045,      "device": 2, "stream": 0,      "correlation": 1000, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1000, "pid": 2, "tid": 0, "ts": 7429770655204.126,
    "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": 7429770655209.819, "dur": 7.977,
    "args": {
      "External id": 3048,      "device": 2, "stream": 0,      "correlation": 1001, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1001, "pid": 2, "tid": 0, "ts": 7429770655209.819,
    "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": 7429770655217.797, "dur": 4.490,
    "args": {
      "External id": 3049,      "device": 2, "stream": 0,      "correlation": 1002, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1002, "pid": 2, "tid": 0, "ts": 7429770655217.797,
    "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": 7429770655222.288, "dur": 4.529,
    "args": {
      "External id": 3050,      "device": 2, "stream": 0,      "correlation": 1003, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1003, "pid": 2, "tid": 0, "ts": 7429770655222.288,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770655226.818, "dur": 4.730,
    "args": {
      "External id": 3051,      "device": 2, "stream": 0,      "correlation": 1004, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1004, "pid": 2, "tid": 0, "ts": 7429770655226.818,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770655231.549, "dur": 5.010,
    "args": {
      "External id": 3053,      "device": 2, "stream": 0,      "correlation": 1005, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1005, "pid": 2, "tid": 0, "ts": 7429770655231.549,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770655236.560, "dur": 64.588,
    "args": {
      "External id": 3061,      "device": 2, "stream": 0,      "correlation": 1009, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1009, "pid": 2, "tid": 0, "ts": 7429770655236.560,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770655301.149, "dur": 50.795,
    "args": {
      "External id": 3074,      "device": 2, "stream": 0,      "correlation": 1013, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1013, "pid": 2, "tid": 0, "ts": 7429770655301.149,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770655351.945, "dur": 50.916,
    "args": {
      "External id": 3087,      "device": 2, "stream": 0,      "correlation": 1017, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1017, "pid": 2, "tid": 0, "ts": 7429770655351.945,
    "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": 7429770655402.862, "dur": 4.730,
    "args": {
      "External id": 3097,      "device": 2, "stream": 0,      "correlation": 1018, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1018, "pid": 2, "tid": 0, "ts": 7429770655402.862,
    "cat": "ac2g", "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": 7429770655407.593, "dur": 4.730,
    "args": {
      "External id": 3102,      "device": 2, "stream": 0,      "correlation": 1019, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1019, "pid": 2, "tid": 0, "ts": 7429770655407.593,
    "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": 7429770655412.324, "dur": 6.374,
    "args": {
      "External id": 3103,      "device": 2, "stream": 0,      "correlation": 1020, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1020, "pid": 2, "tid": 0, "ts": 7429770655412.324,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770655418.699, "dur": 4.770,
    "args": {
      "External id": 3104,      "device": 2, "stream": 0,      "correlation": 1021, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1021, "pid": 2, "tid": 0, "ts": 7429770655418.699,
    "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::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": 7429770655423.470, "dur": 5.491,
    "args": {
      "External id": 3105,      "device": 2, "stream": 0,      "correlation": 1022, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1022, "pid": 2, "tid": 0, "ts": 7429770655423.470,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770655428.962, "dur": 4.650,
    "args": {
      "External id": 3106,      "device": 2, "stream": 0,      "correlation": 1023, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1023, "pid": 2, "tid": 0, "ts": 7429770655428.962,
    "cat": "ac2g", "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": 7429770655433.613, "dur": 4.529,
    "args": {
      "External id": 3111,      "device": 2, "stream": 0,      "correlation": 1024, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1024, "pid": 2, "tid": 0, "ts": 7429770655433.613,
    "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": 7429770655438.143, "dur": 6.053,
    "args": {
      "External id": 3112,      "device": 2, "stream": 0,      "correlation": 1025, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1025, "pid": 2, "tid": 0, "ts": 7429770655438.143,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770655444.197, "dur": 4.530,
    "args": {
      "External id": 3113,      "device": 2, "stream": 0,      "correlation": 1026, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1026, "pid": 2, "tid": 0, "ts": 7429770655444.197,
    "cat": "ac2g", "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": 7429770655448.728, "dur": 5.571,
    "args": {
      "External id": 3114,      "device": 2, "stream": 0,      "correlation": 1027, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1027, "pid": 2, "tid": 0, "ts": 7429770655448.728,
    "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::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::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": 7429770655454.300, "dur": 6.013,
    "args": {
      "External id": 3131,      "device": 2, "stream": 0,      "correlation": 1028, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1028, "pid": 2, "tid": 0, "ts": 7429770655454.300,
    "cat": "ac2g", "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": 7429770655460.314, "dur": 5.612,
    "args": {
      "External id": 3149,      "device": 2, "stream": 0,      "correlation": 1029, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1029, "pid": 2, "tid": 0, "ts": 7429770655460.314,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770655465.927, "dur": 9.782,
    "args": {
      "External id": 3162,      "device": 2, "stream": 0,      "correlation": 1035, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1035, "pid": 2, "tid": 0, "ts": 7429770655465.927,
    "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": 7429770655475.710, "dur": 5.050,
    "args": {
      "External id": 3165,      "device": 2, "stream": 0,      "correlation": 1036, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1036, "pid": 2, "tid": 0, "ts": 7429770655475.710,
    "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::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": 7429770655480.761, "dur": 4.891,
    "args": {
      "External id": 3172,      "device": 2, "stream": 0,      "correlation": 1037, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1037, "pid": 2, "tid": 0, "ts": 7429770655480.761,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770655485.653, "dur": 5.491,
    "args": {
      "External id": 3175,      "device": 2, "stream": 0,      "correlation": 1038, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 1038, "pid": 2, "tid": 0, "ts": 7429770655485.653,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770655491.145, "dur": 10.904,
    "args": {
      "External id": 3187,      "device": 2, "stream": 0,      "correlation": 1042, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1042, "pid": 2, "tid": 0, "ts": 7429770655491.145,
    "cat": "ac2g", "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": 7429770655502.050, "dur": 5.171,
    "args": {
      "External id": 3196,      "device": 2, "stream": 0,      "correlation": 1043, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1043, "pid": 2, "tid": 0, "ts": 7429770655502.050,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770655507.222, "dur": 64.627,
    "args": {
      "External id": 3206,      "device": 2, "stream": 0,      "correlation": 1047, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1047, "pid": 2, "tid": 0, "ts": 7429770655507.222,
    "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": 7429770655571.850, "dur": 5.211,
    "args": {
      "External id": 3209,      "device": 2, "stream": 0,      "correlation": 1048, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1048, "pid": 2, "tid": 0, "ts": 7429770655571.850,
    "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": 7429770655577.062, "dur": 5.733,
    "args": {
      "External id": 3211,      "device": 2, "stream": 0,      "correlation": 1049, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1049, "pid": 2, "tid": 0, "ts": 7429770655577.062,
    "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": 7429770655582.796, "dur": 8.017,
    "args": {
      "External id": 3214,      "device": 2, "stream": 0,      "correlation": 1050, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1050, "pid": 2, "tid": 0, "ts": 7429770655582.796,
    "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": 7429770655590.814, "dur": 4.489,
    "args": {
      "External id": 3215,      "device": 2, "stream": 0,      "correlation": 1051, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1051, "pid": 2, "tid": 0, "ts": 7429770655590.814,
    "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": 7429770655595.304, "dur": 5.251,
    "args": {
      "External id": 3216,      "device": 2, "stream": 0,      "correlation": 1052, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1052, "pid": 2, "tid": 0, "ts": 7429770655595.304,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770655600.556, "dur": 5.091,
    "args": {
      "External id": 3217,      "device": 2, "stream": 0,      "correlation": 1053, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1053, "pid": 2, "tid": 0, "ts": 7429770655600.556,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770655605.648, "dur": 4.971,
    "args": {
      "External id": 3219,      "device": 2, "stream": 0,      "correlation": 1054, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1054, "pid": 2, "tid": 0, "ts": 7429770655605.648,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770655610.620, "dur": 252.859,
    "args": {
      "External id": 3227,      "device": 2, "stream": 0,      "correlation": 1058, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1058, "pid": 2, "tid": 0, "ts": 7429770655610.620,
    "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": 7429770655863.480, "dur": 5.612,
    "args": {
      "External id": 3230,      "device": 2, "stream": 0,      "correlation": 1059, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1059, "pid": 2, "tid": 0, "ts": 7429770655863.480,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770655869.093, "dur": 252.579,
    "args": {
      "External id": 3238,      "device": 2, "stream": 0,      "correlation": 1063, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1063, "pid": 2, "tid": 0, "ts": 7429770655869.093,
    "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": 7429770656121.673, "dur": 5.813,
    "args": {
      "External id": 3241,      "device": 2, "stream": 0,      "correlation": 1064, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1064, "pid": 2, "tid": 0, "ts": 7429770656121.673,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770656127.487, "dur": 165.699,
    "args": {
      "External id": 3249,      "device": 2, "stream": 0,      "correlation": 1068, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1068, "pid": 2, "tid": 0, "ts": 7429770656127.487,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770656293.187, "dur": 7.336,
    "args": {
      "External id": 3249,      "device": 2, "stream": 0,      "correlation": 1069, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1069, "pid": 2, "tid": 0, "ts": 7429770656293.187,
    "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": 7429770656300.524, "dur": 4.650,
    "args": {
      "External id": 3252,      "device": 2, "stream": 0,      "correlation": 1070, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1070, "pid": 2, "tid": 0, "ts": 7429770656300.524,
    "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": 7429770656305.175, "dur": 5.371,
    "args": {
      "External id": 3254,      "device": 2, "stream": 0,      "correlation": 1071, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1071, "pid": 2, "tid": 0, "ts": 7429770656305.175,
    "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": 7429770656310.547, "dur": 7.977,
    "args": {
      "External id": 3257,      "device": 2, "stream": 0,      "correlation": 1072, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1072, "pid": 2, "tid": 0, "ts": 7429770656310.547,
    "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": 7429770656318.525, "dur": 4.530,
    "args": {
      "External id": 3258,      "device": 2, "stream": 0,      "correlation": 1073, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1073, "pid": 2, "tid": 0, "ts": 7429770656318.525,
    "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": 7429770656323.056, "dur": 4.650,
    "args": {
      "External id": 3259,      "device": 2, "stream": 0,      "correlation": 1074, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1074, "pid": 2, "tid": 0, "ts": 7429770656323.056,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770656327.707, "dur": 4.769,
    "args": {
      "External id": 3260,      "device": 2, "stream": 0,      "correlation": 1075, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1075, "pid": 2, "tid": 0, "ts": 7429770656327.707,
    "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": 7429770656332.477, "dur": 4.971,
    "args": {
      "External id": 3262,      "device": 2, "stream": 0,      "correlation": 1076, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1076, "pid": 2, "tid": 0, "ts": 7429770656332.477,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770656337.449, "dur": 64.026,
    "args": {
      "External id": 3270,      "device": 2, "stream": 0,      "correlation": 1080, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1080, "pid": 2, "tid": 0, "ts": 7429770656337.449,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770656401.476, "dur": 49.994,
    "args": {
      "External id": 3283,      "device": 2, "stream": 0,      "correlation": 1084, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1084, "pid": 2, "tid": 0, "ts": 7429770656401.476,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770656451.471, "dur": 50.475,
    "args": {
      "External id": 3296,      "device": 2, "stream": 0,      "correlation": 1088, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1088, "pid": 2, "tid": 0, "ts": 7429770656451.471,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770656501.947, "dur": 4.730,
    "args": {
      "External id": 3306,      "device": 2, "stream": 0,      "correlation": 1089, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1089, "pid": 2, "tid": 0, "ts": 7429770656501.947,
    "cat": "ac2g", "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": 7429770656506.678, "dur": 5.411,
    "args": {
      "External id": 3311,      "device": 2, "stream": 0,      "correlation": 1090, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1090, "pid": 2, "tid": 0, "ts": 7429770656506.678,
    "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": 7429770656512.090, "dur": 6.293,
    "args": {
      "External id": 3312,      "device": 2, "stream": 0,      "correlation": 1091, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1091, "pid": 2, "tid": 0, "ts": 7429770656512.090,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770656518.384, "dur": 4.931,
    "args": {
      "External id": 3313,      "device": 2, "stream": 0,      "correlation": 1092, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1092, "pid": 2, "tid": 0, "ts": 7429770656518.384,
    "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::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": 7429770656523.316, "dur": 5.010,
    "args": {
      "External id": 3314,      "device": 2, "stream": 0,      "correlation": 1093, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1093, "pid": 2, "tid": 0, "ts": 7429770656523.316,
    "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": 7429770656528.327, "dur": 4.770,
    "args": {
      "External id": 3315,      "device": 2, "stream": 0,      "correlation": 1094, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1094, "pid": 2, "tid": 0, "ts": 7429770656528.327,
    "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::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": 7429770656533.098, "dur": 5.372,
    "args": {
      "External id": 3320,      "device": 2, "stream": 0,      "correlation": 1095, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1095, "pid": 2, "tid": 0, "ts": 7429770656533.098,
    "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": 7429770656538.471, "dur": 6.413,
    "args": {
      "External id": 3321,      "device": 2, "stream": 0,      "correlation": 1096, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1096, "pid": 2, "tid": 0, "ts": 7429770656538.471,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770656544.885, "dur": 4.730,
    "args": {
      "External id": 3322,      "device": 2, "stream": 0,      "correlation": 1097, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1097, "pid": 2, "tid": 0, "ts": 7429770656544.885,
    "cat": "ac2g", "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": 7429770656549.616, "dur": 4.850,
    "args": {
      "External id": 3323,      "device": 2, "stream": 0,      "correlation": 1098, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1098, "pid": 2, "tid": 0, "ts": 7429770656549.616,
    "cat": "ac2g", "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": 7429770656554.467, "dur": 4.971,
    "args": {
      "External id": 3340,      "device": 2, "stream": 0,      "correlation": 1099, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1099, "pid": 2, "tid": 0, "ts": 7429770656554.467,
    "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::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::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": 7429770656559.439, "dur": 5.491,
    "args": {
      "External id": 3358,      "device": 2, "stream": 0,      "correlation": 1100, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1100, "pid": 2, "tid": 0, "ts": 7429770656559.439,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770656564.931, "dur": 9.461,
    "args": {
      "External id": 3371,      "device": 2, "stream": 0,      "correlation": 1106, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1106, "pid": 2, "tid": 0, "ts": 7429770656564.931,
    "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": 7429770656574.393, "dur": 5.452,
    "args": {
      "External id": 3374,      "device": 2, "stream": 0,      "correlation": 1107, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1107, "pid": 2, "tid": 0, "ts": 7429770656574.393,
    "cat": "ac2g", "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": 7429770656579.846, "dur": 5.852,
    "args": {
      "External id": 3381,      "device": 2, "stream": 0,      "correlation": 1108, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1108, "pid": 2, "tid": 0, "ts": 7429770656579.846,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770656585.699, "dur": 5.211,
    "args": {
      "External id": 3384,      "device": 2, "stream": 0,      "correlation": 1109, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 1109, "pid": 2, "tid": 0, "ts": 7429770656585.699,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770656590.911, "dur": 10.944,
    "args": {
      "External id": 3396,      "device": 2, "stream": 0,      "correlation": 1113, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1113, "pid": 2, "tid": 0, "ts": 7429770656590.911,
    "cat": "ac2g", "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": 7429770656601.856, "dur": 5.211,
    "args": {
      "External id": 3405,      "device": 2, "stream": 0,      "correlation": 1114, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1114, "pid": 2, "tid": 0, "ts": 7429770656601.856,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770656607.068, "dur": 63.585,
    "args": {
      "External id": 3415,      "device": 2, "stream": 0,      "correlation": 1118, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1118, "pid": 2, "tid": 0, "ts": 7429770656607.068,
    "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": 7429770656670.654, "dur": 4.570,
    "args": {
      "External id": 3418,      "device": 2, "stream": 0,      "correlation": 1119, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1119, "pid": 2, "tid": 0, "ts": 7429770656670.654,
    "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": 7429770656675.225, "dur": 5.291,
    "args": {
      "External id": 3420,      "device": 2, "stream": 0,      "correlation": 1120, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1120, "pid": 2, "tid": 0, "ts": 7429770656675.225,
    "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": 7429770656680.517, "dur": 8.138,
    "args": {
      "External id": 3423,      "device": 2, "stream": 0,      "correlation": 1121, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1121, "pid": 2, "tid": 0, "ts": 7429770656680.517,
    "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": 7429770656688.656, "dur": 4.409,
    "args": {
      "External id": 3424,      "device": 2, "stream": 0,      "correlation": 1122, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1122, "pid": 2, "tid": 0, "ts": 7429770656688.656,
    "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": 7429770656693.066, "dur": 5.411,
    "args": {
      "External id": 3425,      "device": 2, "stream": 0,      "correlation": 1123, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1123, "pid": 2, "tid": 0, "ts": 7429770656693.066,
    "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": 7429770656698.478, "dur": 5.091,
    "args": {
      "External id": 3426,      "device": 2, "stream": 0,      "correlation": 1124, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1124, "pid": 2, "tid": 0, "ts": 7429770656698.478,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770656703.570, "dur": 4.970,
    "args": {
      "External id": 3428,      "device": 2, "stream": 0,      "correlation": 1125, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1125, "pid": 2, "tid": 0, "ts": 7429770656703.570,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770656708.541, "dur": 253.181,
    "args": {
      "External id": 3436,      "device": 2, "stream": 0,      "correlation": 1129, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1129, "pid": 2, "tid": 0, "ts": 7429770656708.541,
    "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": 7429770656961.723, "dur": 5.612,
    "args": {
      "External id": 3439,      "device": 2, "stream": 0,      "correlation": 1130, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1130, "pid": 2, "tid": 0, "ts": 7429770656961.723,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770656967.336, "dur": 252.619,
    "args": {
      "External id": 3447,      "device": 2, "stream": 0,      "correlation": 1134, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1134, "pid": 2, "tid": 0, "ts": 7429770656967.336,
    "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": 7429770657219.956, "dur": 5.572,
    "args": {
      "External id": 3450,      "device": 2, "stream": 0,      "correlation": 1135, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1135, "pid": 2, "tid": 0, "ts": 7429770657219.956,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770657225.529, "dur": 165.940,
    "args": {
      "External id": 3458,      "device": 2, "stream": 0,      "correlation": 1139, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1139, "pid": 2, "tid": 0, "ts": 7429770657225.529,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770657391.470, "dur": 8.137,
    "args": {
      "External id": 3458,      "device": 2, "stream": 0,      "correlation": 1140, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1140, "pid": 2, "tid": 0, "ts": 7429770657391.470,
    "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": 7429770657399.608, "dur": 5.252,
    "args": {
      "External id": 3461,      "device": 2, "stream": 0,      "correlation": 1141, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1141, "pid": 2, "tid": 0, "ts": 7429770657399.608,
    "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": 7429770657404.861, "dur": 5.571,
    "args": {
      "External id": 3463,      "device": 2, "stream": 0,      "correlation": 1142, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1142, "pid": 2, "tid": 0, "ts": 7429770657404.861,
    "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": 7429770657410.433, "dur": 8.218,
    "args": {
      "External id": 3466,      "device": 2, "stream": 0,      "correlation": 1143, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1143, "pid": 2, "tid": 0, "ts": 7429770657410.433,
    "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": 7429770657418.652, "dur": 5.372,
    "args": {
      "External id": 3467,      "device": 2, "stream": 0,      "correlation": 1144, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1144, "pid": 2, "tid": 0, "ts": 7429770657418.652,
    "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": 7429770657424.025, "dur": 5.211,
    "args": {
      "External id": 3468,      "device": 2, "stream": 0,      "correlation": 1145, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1145, "pid": 2, "tid": 0, "ts": 7429770657424.025,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770657429.237, "dur": 5.090,
    "args": {
      "External id": 3469,      "device": 2, "stream": 0,      "correlation": 1146, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1146, "pid": 2, "tid": 0, "ts": 7429770657429.237,
    "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": 7429770657434.328, "dur": 5.171,
    "args": {
      "External id": 3471,      "device": 2, "stream": 0,      "correlation": 1147, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1147, "pid": 2, "tid": 0, "ts": 7429770657434.328,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770657439.500, "dur": 62.984,
    "args": {
      "External id": 3479,      "device": 2, "stream": 0,      "correlation": 1153, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1153, "pid": 2, "tid": 0, "ts": 7429770657439.500,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770657502.485, "dur": 51.116,
    "args": {
      "External id": 3492,      "device": 2, "stream": 0,      "correlation": 1157, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1157, "pid": 2, "tid": 0, "ts": 7429770657502.485,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770657553.602, "dur": 50.956,
    "args": {
      "External id": 3505,      "device": 2, "stream": 0,      "correlation": 1161, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1161, "pid": 2, "tid": 0, "ts": 7429770657553.602,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770657604.559, "dur": 4.810,
    "args": {
      "External id": 3515,      "device": 2, "stream": 0,      "correlation": 1162, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1162, "pid": 2, "tid": 0, "ts": 7429770657604.559,
    "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::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": 7429770657609.370, "dur": 4.810,
    "args": {
      "External id": 3520,      "device": 2, "stream": 0,      "correlation": 1163, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1163, "pid": 2, "tid": 0, "ts": 7429770657609.370,
    "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": 7429770657614.181, "dur": 6.614,
    "args": {
      "External id": 3521,      "device": 2, "stream": 0,      "correlation": 1164, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1164, "pid": 2, "tid": 0, "ts": 7429770657614.181,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770657620.796, "dur": 4.730,
    "args": {
      "External id": 3522,      "device": 2, "stream": 0,      "correlation": 1165, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1165, "pid": 2, "tid": 0, "ts": 7429770657620.796,
    "cat": "ac2g", "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": 7429770657625.527, "dur": 5.091,
    "args": {
      "External id": 3523,      "device": 2, "stream": 0,      "correlation": 1166, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1166, "pid": 2, "tid": 0, "ts": 7429770657625.527,
    "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": 7429770657630.619, "dur": 4.810,
    "args": {
      "External id": 3524,      "device": 2, "stream": 0,      "correlation": 1167, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1167, "pid": 2, "tid": 0, "ts": 7429770657630.619,
    "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::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": 7429770657635.430, "dur": 4.610,
    "args": {
      "External id": 3529,      "device": 2, "stream": 0,      "correlation": 1168, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1168, "pid": 2, "tid": 0, "ts": 7429770657635.430,
    "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": 7429770657640.041, "dur": 6.454,
    "args": {
      "External id": 3530,      "device": 2, "stream": 0,      "correlation": 1169, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1169, "pid": 2, "tid": 0, "ts": 7429770657640.041,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770657646.496, "dur": 5.170,
    "args": {
      "External id": 3531,      "device": 2, "stream": 0,      "correlation": 1170, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1170, "pid": 2, "tid": 0, "ts": 7429770657646.496,
    "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": 7429770657651.667, "dur": 4.891,
    "args": {
      "External id": 3532,      "device": 2, "stream": 0,      "correlation": 1171, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1171, "pid": 2, "tid": 0, "ts": 7429770657651.667,
    "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": 7429770657656.559, "dur": 6.012,
    "args": {
      "External id": 3549,      "device": 2, "stream": 0,      "correlation": 1172, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1172, "pid": 2, "tid": 0, "ts": 7429770657656.559,
    "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::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::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": 7429770657662.572, "dur": 5.412,
    "args": {
      "External id": 3567,      "device": 2, "stream": 0,      "correlation": 1173, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1173, "pid": 2, "tid": 0, "ts": 7429770657662.572,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770657667.985, "dur": 9.581,
    "args": {
      "External id": 3580,      "device": 2, "stream": 0,      "correlation": 1179, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1179, "pid": 2, "tid": 0, "ts": 7429770657667.985,
    "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": 7429770657677.567, "dur": 5.131,
    "args": {
      "External id": 3583,      "device": 2, "stream": 0,      "correlation": 1180, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1180, "pid": 2, "tid": 0, "ts": 7429770657677.567,
    "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": 7429770657682.699, "dur": 5.451,
    "args": {
      "External id": 3590,      "device": 2, "stream": 0,      "correlation": 1181, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1181, "pid": 2, "tid": 0, "ts": 7429770657682.699,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770657688.151, "dur": 5.171,
    "args": {
      "External id": 3593,      "device": 2, "stream": 0,      "correlation": 1182, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 1182, "pid": 2, "tid": 0, "ts": 7429770657688.151,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770657693.323, "dur": 10.944,
    "args": {
      "External id": 3605,      "device": 2, "stream": 0,      "correlation": 1186, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1186, "pid": 2, "tid": 0, "ts": 7429770657693.323,
    "cat": "ac2g", "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": 7429770657704.268, "dur": 5.091,
    "args": {
      "External id": 3614,      "device": 2, "stream": 0,      "correlation": 1187, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1187, "pid": 2, "tid": 0, "ts": 7429770657704.268,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770657709.360, "dur": 63.465,
    "args": {
      "External id": 3624,      "device": 2, "stream": 0,      "correlation": 1191, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1191, "pid": 2, "tid": 0, "ts": 7429770657709.360,
    "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": 7429770657772.826, "dur": 4.850,
    "args": {
      "External id": 3627,      "device": 2, "stream": 0,      "correlation": 1192, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1192, "pid": 2, "tid": 0, "ts": 7429770657772.826,
    "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": 7429770657777.677, "dur": 5.612,
    "args": {
      "External id": 3629,      "device": 2, "stream": 0,      "correlation": 1193, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1193, "pid": 2, "tid": 0, "ts": 7429770657777.677,
    "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": 7429770657783.290, "dur": 7.937,
    "args": {
      "External id": 3632,      "device": 2, "stream": 0,      "correlation": 1194, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1194, "pid": 2, "tid": 0, "ts": 7429770657783.290,
    "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": 7429770657791.228, "dur": 4.730,
    "args": {
      "External id": 3633,      "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": 7429770657791.228,
    "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": 7429770657795.959, "dur": 5.291,
    "args": {
      "External id": 3634,      "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": 7429770657795.959,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770657801.251, "dur": 5.211,
    "args": {
      "External id": 3635,      "device": 2, "stream": 0,      "correlation": 1197, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1197, "pid": 2, "tid": 0, "ts": 7429770657801.251,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770657806.463, "dur": 5.131,
    "args": {
      "External id": 3637,      "device": 2, "stream": 0,      "correlation": 1198, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1198, "pid": 2, "tid": 0, "ts": 7429770657806.463,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770657811.595, "dur": 257.671,
    "args": {
      "External id": 3645,      "device": 2, "stream": 0,      "correlation": 1202, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1202, "pid": 2, "tid": 0, "ts": 7429770657811.595,
    "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": 7429770658069.267, "dur": 5.571,
    "args": {
      "External id": 3648,      "device": 2, "stream": 0,      "correlation": 1203, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1203, "pid": 2, "tid": 0, "ts": 7429770658069.267,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770658074.839, "dur": 256.228,
    "args": {
      "External id": 3656,      "device": 2, "stream": 0,      "correlation": 1207, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1207, "pid": 2, "tid": 0, "ts": 7429770658074.839,
    "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": 7429770658331.068, "dur": 5.572,
    "args": {
      "External id": 3659,      "device": 2, "stream": 0,      "correlation": 1208, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1208, "pid": 2, "tid": 0, "ts": 7429770658331.068,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770658336.641, "dur": 165.419,
    "args": {
      "External id": 3667,      "device": 2, "stream": 0,      "correlation": 1212, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1212, "pid": 2, "tid": 0, "ts": 7429770658336.641,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770658502.061, "dur": 7.977,
    "args": {
      "External id": 3667,      "device": 2, "stream": 0,      "correlation": 1213, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1213, "pid": 2, "tid": 0, "ts": 7429770658502.061,
    "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": 7429770658510.039, "dur": 5.131,
    "args": {
      "External id": 3670,      "device": 2, "stream": 0,      "correlation": 1214, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1214, "pid": 2, "tid": 0, "ts": 7429770658510.039,
    "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": 7429770658515.171, "dur": 5.291,
    "args": {
      "External id": 3672,      "device": 2, "stream": 0,      "correlation": 1215, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1215, "pid": 2, "tid": 0, "ts": 7429770658515.171,
    "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": 7429770658520.463, "dur": 8.218,
    "args": {
      "External id": 3675,      "device": 2, "stream": 0,      "correlation": 1216, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1216, "pid": 2, "tid": 0, "ts": 7429770658520.463,
    "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": 7429770658528.682, "dur": 5.050,
    "args": {
      "External id": 3676,      "device": 2, "stream": 0,      "correlation": 1217, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1217, "pid": 2, "tid": 0, "ts": 7429770658528.682,
    "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": 7429770658533.733, "dur": 5.011,
    "args": {
      "External id": 3677,      "device": 2, "stream": 0,      "correlation": 1218, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1218, "pid": 2, "tid": 0, "ts": 7429770658533.733,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770658538.745, "dur": 4.770,
    "args": {
      "External id": 3678,      "device": 2, "stream": 0,      "correlation": 1219, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1219, "pid": 2, "tid": 0, "ts": 7429770658538.745,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770658543.516, "dur": 4.970,
    "args": {
      "External id": 3680,      "device": 2, "stream": 0,      "correlation": 1220, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1220, "pid": 2, "tid": 0, "ts": 7429770658543.516,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770658548.487, "dur": 64.668,
    "args": {
      "External id": 3688,      "device": 2, "stream": 0,      "correlation": 1224, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1224, "pid": 2, "tid": 0, "ts": 7429770658548.487,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770658613.156, "dur": 50.034,
    "args": {
      "External id": 3701,      "device": 2, "stream": 0,      "correlation": 1228, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1228, "pid": 2, "tid": 0, "ts": 7429770658613.156,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770658663.191, "dur": 50.434,
    "args": {
      "External id": 3714,      "device": 2, "stream": 0,      "correlation": 1232, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1232, "pid": 2, "tid": 0, "ts": 7429770658663.191,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770658713.626, "dur": 4.690,
    "args": {
      "External id": 3724,      "device": 2, "stream": 0,      "correlation": 1233, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1233, "pid": 2, "tid": 0, "ts": 7429770658713.626,
    "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::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": 7429770658718.317, "dur": 4.770,
    "args": {
      "External id": 3729,      "device": 2, "stream": 0,      "correlation": 1234, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1234, "pid": 2, "tid": 0, "ts": 7429770658718.317,
    "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": 7429770658723.088, "dur": 6.614,
    "args": {
      "External id": 3730,      "device": 2, "stream": 0,      "correlation": 1235, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1235, "pid": 2, "tid": 0, "ts": 7429770658723.088,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770658729.703, "dur": 4.770,
    "args": {
      "External id": 3731,      "device": 2, "stream": 0,      "correlation": 1236, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1236, "pid": 2, "tid": 0, "ts": 7429770658729.703,
    "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::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": 7429770658734.474, "dur": 4.971,
    "args": {
      "External id": 3732,      "device": 2, "stream": 0,      "correlation": 1237, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1237, "pid": 2, "tid": 0, "ts": 7429770658734.474,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770658739.446, "dur": 4.569,
    "args": {
      "External id": 3733,      "device": 2, "stream": 0,      "correlation": 1238, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1238, "pid": 2, "tid": 0, "ts": 7429770658739.446,
    "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::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": 7429770658744.016, "dur": 4.490,
    "args": {
      "External id": 3738,      "device": 2, "stream": 0,      "correlation": 1239, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1239, "pid": 2, "tid": 0, "ts": 7429770658744.016,
    "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": 7429770658748.507, "dur": 6.253,
    "args": {
      "External id": 3739,      "device": 2, "stream": 0,      "correlation": 1240, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1240, "pid": 2, "tid": 0, "ts": 7429770658748.507,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770658754.761, "dur": 4.770,
    "args": {
      "External id": 3740,      "device": 2, "stream": 0,      "correlation": 1241, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1241, "pid": 2, "tid": 0, "ts": 7429770658754.761,
    "cat": "ac2g", "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": 7429770658759.532, "dur": 5.451,
    "args": {
      "External id": 3741,      "device": 2, "stream": 0,      "correlation": 1242, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1242, "pid": 2, "tid": 0, "ts": 7429770658759.532,
    "cat": "ac2g", "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": 7429770658764.984, "dur": 5.612,
    "args": {
      "External id": 3758,      "device": 2, "stream": 0,      "correlation": 1243, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1243, "pid": 2, "tid": 0, "ts": 7429770658764.984,
    "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::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::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": 7429770658770.597, "dur": 5.412,
    "args": {
      "External id": 3776,      "device": 2, "stream": 0,      "correlation": 1244, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1244, "pid": 2, "tid": 0, "ts": 7429770658770.597,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770658776.010, "dur": 9.581,
    "args": {
      "External id": 3789,      "device": 2, "stream": 0,      "correlation": 1250, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1250, "pid": 2, "tid": 0, "ts": 7429770658776.010,
    "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": 7429770658785.592, "dur": 4.529,
    "args": {
      "External id": 3792,      "device": 2, "stream": 0,      "correlation": 1251, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1251, "pid": 2, "tid": 0, "ts": 7429770658785.592,
    "cat": "ac2g", "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": 7429770658790.122, "dur": 4.850,
    "args": {
      "External id": 3799,      "device": 2, "stream": 0,      "correlation": 1252, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1252, "pid": 2, "tid": 0, "ts": 7429770658790.122,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770658794.973, "dur": 5.171,
    "args": {
      "External id": 3802,      "device": 2, "stream": 0,      "correlation": 1253, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 1253, "pid": 2, "tid": 0, "ts": 7429770658794.973,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770658800.145, "dur": 10.904,
    "args": {
      "External id": 3814,      "device": 2, "stream": 0,      "correlation": 1257, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1257, "pid": 2, "tid": 0, "ts": 7429770658800.145,
    "cat": "ac2g", "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": 7429770658811.050, "dur": 5.091,
    "args": {
      "External id": 3823,      "device": 2, "stream": 0,      "correlation": 1258, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1258, "pid": 2, "tid": 0, "ts": 7429770658811.050,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770658816.142, "dur": 63.785,
    "args": {
      "External id": 3833,      "device": 2, "stream": 0,      "correlation": 1262, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1262, "pid": 2, "tid": 0, "ts": 7429770658816.142,
    "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": 7429770658879.928, "dur": 5.131,
    "args": {
      "External id": 3836,      "device": 2, "stream": 0,      "correlation": 1263, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1263, "pid": 2, "tid": 0, "ts": 7429770658879.928,
    "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": 7429770658885.060, "dur": 5.452,
    "args": {
      "External id": 3838,      "device": 2, "stream": 0,      "correlation": 1264, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1264, "pid": 2, "tid": 0, "ts": 7429770658885.060,
    "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": 7429770658890.513, "dur": 7.817,
    "args": {
      "External id": 3841,      "device": 2, "stream": 0,      "correlation": 1265, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1265, "pid": 2, "tid": 0, "ts": 7429770658890.513,
    "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": 7429770658898.331, "dur": 4.449,
    "args": {
      "External id": 3842,      "device": 2, "stream": 0,      "correlation": 1266, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1266, "pid": 2, "tid": 0, "ts": 7429770658898.331,
    "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": 7429770658902.781, "dur": 4.650,
    "args": {
      "External id": 3843,      "device": 2, "stream": 0,      "correlation": 1267, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1267, "pid": 2, "tid": 0, "ts": 7429770658902.781,
    "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": 7429770658907.432, "dur": 4.770,
    "args": {
      "External id": 3844,      "device": 2, "stream": 0,      "correlation": 1268, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1268, "pid": 2, "tid": 0, "ts": 7429770658907.432,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770658912.203, "dur": 5.010,
    "args": {
      "External id": 3846,      "device": 2, "stream": 0,      "correlation": 1269, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1269, "pid": 2, "tid": 0, "ts": 7429770658912.203,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770658917.214, "dur": 252.820,
    "args": {
      "External id": 3854,      "device": 2, "stream": 0,      "correlation": 1273, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1273, "pid": 2, "tid": 0, "ts": 7429770658917.214,
    "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": 7429770659170.035, "dur": 5.612,
    "args": {
      "External id": 3857,      "device": 2, "stream": 0,      "correlation": 1274, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1274, "pid": 2, "tid": 0, "ts": 7429770659170.035,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770659175.648, "dur": 253.060,
    "args": {
      "External id": 3865,      "device": 2, "stream": 0,      "correlation": 1278, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1278, "pid": 2, "tid": 0, "ts": 7429770659175.648,
    "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": 7429770659428.709, "dur": 5.772,
    "args": {
      "External id": 3868,      "device": 2, "stream": 0,      "correlation": 1279, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1279, "pid": 2, "tid": 0, "ts": 7429770659428.709,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770659434.482, "dur": 165.419,
    "args": {
      "External id": 3876,      "device": 2, "stream": 0,      "correlation": 1283, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1283, "pid": 2, "tid": 0, "ts": 7429770659434.482,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770659599.902, "dur": 7.416,
    "args": {
      "External id": 3876,      "device": 2, "stream": 0,      "correlation": 1284, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1284, "pid": 2, "tid": 0, "ts": 7429770659599.902,
    "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": 7429770659607.319, "dur": 4.690,
    "args": {
      "External id": 3879,      "device": 2, "stream": 0,      "correlation": 1285, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1285, "pid": 2, "tid": 0, "ts": 7429770659607.319,
    "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": 7429770659612.010, "dur": 6.374,
    "args": {
      "External id": 3881,      "device": 2, "stream": 0,      "correlation": 1286, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1286, "pid": 2, "tid": 0, "ts": 7429770659612.010,
    "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": 7429770659618.385, "dur": 8.017,
    "args": {
      "External id": 3884,      "device": 2, "stream": 0,      "correlation": 1287, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1287, "pid": 2, "tid": 0, "ts": 7429770659618.385,
    "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": 7429770659626.403, "dur": 4.369,
    "args": {
      "External id": 3885,      "device": 2, "stream": 0,      "correlation": 1288, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1288, "pid": 2, "tid": 0, "ts": 7429770659626.403,
    "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": 7429770659630.773, "dur": 4.610,
    "args": {
      "External id": 3886,      "device": 2, "stream": 0,      "correlation": 1289, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1289, "pid": 2, "tid": 0, "ts": 7429770659630.773,
    "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": 7429770659635.384, "dur": 5.090,
    "args": {
      "External id": 3887,      "device": 2, "stream": 0,      "correlation": 1290, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1290, "pid": 2, "tid": 0, "ts": 7429770659635.384,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770659640.475, "dur": 5.492,
    "args": {
      "External id": 3889,      "device": 2, "stream": 0,      "correlation": 1291, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1291, "pid": 2, "tid": 0, "ts": 7429770659640.475,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770659645.968, "dur": 63.825,
    "args": {
      "External id": 3897,      "device": 2, "stream": 0,      "correlation": 1295, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1295, "pid": 2, "tid": 0, "ts": 7429770659645.968,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770659709.794, "dur": 50.997,
    "args": {
      "External id": 3910,      "device": 2, "stream": 0,      "correlation": 1299, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1299, "pid": 2, "tid": 0, "ts": 7429770659709.794,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770659760.792, "dur": 49.913,
    "args": {
      "External id": 3923,      "device": 2, "stream": 0,      "correlation": 1303, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1303, "pid": 2, "tid": 0, "ts": 7429770659760.792,
    "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": 7429770659810.706, "dur": 5.051,
    "args": {
      "External id": 3933,      "device": 2, "stream": 0,      "correlation": 1304, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1304, "pid": 2, "tid": 0, "ts": 7429770659810.706,
    "cat": "ac2g", "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": 7429770659815.758, "dur": 4.850,
    "args": {
      "External id": 3938,      "device": 2, "stream": 0,      "correlation": 1305, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1305, "pid": 2, "tid": 0, "ts": 7429770659815.758,
    "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": 7429770659820.609, "dur": 6.213,
    "args": {
      "External id": 3939,      "device": 2, "stream": 0,      "correlation": 1306, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1306, "pid": 2, "tid": 0, "ts": 7429770659820.609,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770659826.823, "dur": 4.730,
    "args": {
      "External id": 3940,      "device": 2, "stream": 0,      "correlation": 1307, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1307, "pid": 2, "tid": 0, "ts": 7429770659826.823,
    "cat": "ac2g", "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": 7429770659831.554, "dur": 5.011,
    "args": {
      "External id": 3941,      "device": 2, "stream": 0,      "correlation": 1308, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1308, "pid": 2, "tid": 0, "ts": 7429770659831.554,
    "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": 7429770659836.566, "dur": 4.529,
    "args": {
      "External id": 3942,      "device": 2, "stream": 0,      "correlation": 1309, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1309, "pid": 2, "tid": 0, "ts": 7429770659836.566,
    "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": 7429770659841.096, "dur": 4.529,
    "args": {
      "External id": 3947,      "device": 2, "stream": 0,      "correlation": 1310, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1310, "pid": 2, "tid": 0, "ts": 7429770659841.096,
    "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": 7429770659845.626, "dur": 6.093,
    "args": {
      "External id": 3948,      "device": 2, "stream": 0,      "correlation": 1311, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1311, "pid": 2, "tid": 0, "ts": 7429770659845.626,
    "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": 7429770659851.720, "dur": 4.690,
    "args": {
      "External id": 3949,      "device": 2, "stream": 0,      "correlation": 1312, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1312, "pid": 2, "tid": 0, "ts": 7429770659851.720,
    "cat": "ac2g", "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": 7429770659856.411, "dur": 5.532,
    "args": {
      "External id": 3950,      "device": 2, "stream": 0,      "correlation": 1313, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1313, "pid": 2, "tid": 0, "ts": 7429770659856.411,
    "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::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::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": 7429770659861.944, "dur": 5.652,
    "args": {
      "External id": 3967,      "device": 2, "stream": 0,      "correlation": 1314, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1314, "pid": 2, "tid": 0, "ts": 7429770659861.944,
    "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::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::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": 7429770659867.597, "dur": 5.492,
    "args": {
      "External id": 3985,      "device": 2, "stream": 0,      "correlation": 1315, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1315, "pid": 2, "tid": 0, "ts": 7429770659867.597,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770659873.090, "dur": 9.781,
    "args": {
      "External id": 3998,      "device": 2, "stream": 0,      "correlation": 1321, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1321, "pid": 2, "tid": 0, "ts": 7429770659873.090,
    "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": 7429770659882.872, "dur": 4.569,
    "args": {
      "External id": 4001,      "device": 2, "stream": 0,      "correlation": 1322, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1322, "pid": 2, "tid": 0, "ts": 7429770659882.872,
    "cat": "ac2g", "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": 7429770659887.442, "dur": 4.730,
    "args": {
      "External id": 4008,      "device": 2, "stream": 0,      "correlation": 1323, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1323, "pid": 2, "tid": 0, "ts": 7429770659887.442,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770659892.173, "dur": 5.412,
    "args": {
      "External id": 4011,      "device": 2, "stream": 0,      "correlation": 1324, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 1324, "pid": 2, "tid": 0, "ts": 7429770659892.173,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770659897.586, "dur": 11.064,
    "args": {
      "External id": 4023,      "device": 2, "stream": 0,      "correlation": 1328, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1328, "pid": 2, "tid": 0, "ts": 7429770659897.586,
    "cat": "ac2g", "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": 7429770659908.651, "dur": 5.372,
    "args": {
      "External id": 4032,      "device": 2, "stream": 0,      "correlation": 1329, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1329, "pid": 2, "tid": 0, "ts": 7429770659908.651,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770659914.024, "dur": 64.346,
    "args": {
      "External id": 4042,      "device": 2, "stream": 0,      "correlation": 1333, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1333, "pid": 2, "tid": 0, "ts": 7429770659914.024,
    "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": 7429770659978.371, "dur": 5.492,
    "args": {
      "External id": 4045,      "device": 2, "stream": 0,      "correlation": 1334, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1334, "pid": 2, "tid": 0, "ts": 7429770659978.371,
    "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": 7429770659983.864, "dur": 6.494,
    "args": {
      "External id": 4047,      "device": 2, "stream": 0,      "correlation": 1335, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1335, "pid": 2, "tid": 0, "ts": 7429770659983.864,
    "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": 7429770659990.359, "dur": 8.619,
    "args": {
      "External id": 4050,      "device": 2, "stream": 0,      "correlation": 1336, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1336, "pid": 2, "tid": 0, "ts": 7429770659990.359,
    "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": 7429770659998.979, "dur": 4.369,
    "args": {
      "External id": 4051,      "device": 2, "stream": 0,      "correlation": 1337, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1337, "pid": 2, "tid": 0, "ts": 7429770659998.979,
    "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": 7429770660003.349, "dur": 5.331,
    "args": {
      "External id": 4052,      "device": 2, "stream": 0,      "correlation": 1338, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1338, "pid": 2, "tid": 0, "ts": 7429770660003.349,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770660008.681, "dur": 5.251,
    "args": {
      "External id": 4053,      "device": 2, "stream": 0,      "correlation": 1339, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1339, "pid": 2, "tid": 0, "ts": 7429770660008.681,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770660013.933, "dur": 5.371,
    "args": {
      "External id": 4055,      "device": 2, "stream": 0,      "correlation": 1340, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1340, "pid": 2, "tid": 0, "ts": 7429770660013.933,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770660019.305, "dur": 253.181,
    "args": {
      "External id": 4063,      "device": 2, "stream": 0,      "correlation": 1346, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1346, "pid": 2, "tid": 0, "ts": 7429770660019.305,
    "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": 7429770660272.487, "dur": 5.451,
    "args": {
      "External id": 4066,      "device": 2, "stream": 0,      "correlation": 1347, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1347, "pid": 2, "tid": 0, "ts": 7429770660272.487,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770660277.939, "dur": 252.620,
    "args": {
      "External id": 4074,      "device": 2, "stream": 0,      "correlation": 1351, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1351, "pid": 2, "tid": 0, "ts": 7429770660277.939,
    "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": 7429770660530.560, "dur": 5.692,
    "args": {
      "External id": 4077,      "device": 2, "stream": 0,      "correlation": 1352, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1352, "pid": 2, "tid": 0, "ts": 7429770660530.560,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770660536.253, "dur": 166.301,
    "args": {
      "External id": 4085,      "device": 2, "stream": 0,      "correlation": 1356, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1356, "pid": 2, "tid": 0, "ts": 7429770660536.253,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770660702.555, "dur": 7.095,
    "args": {
      "External id": 4085,      "device": 2, "stream": 0,      "correlation": 1357, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1357, "pid": 2, "tid": 0, "ts": 7429770660702.555,
    "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": 7429770660709.651, "dur": 5.171,
    "args": {
      "External id": 4088,      "device": 2, "stream": 0,      "correlation": 1358, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1358, "pid": 2, "tid": 0, "ts": 7429770660709.651,
    "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": 7429770660714.823, "dur": 5.812,
    "args": {
      "External id": 4090,      "device": 2, "stream": 0,      "correlation": 1359, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1359, "pid": 2, "tid": 0, "ts": 7429770660714.823,
    "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": 7429770660720.636, "dur": 8.218,
    "args": {
      "External id": 4093,      "device": 2, "stream": 0,      "correlation": 1360, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1360, "pid": 2, "tid": 0, "ts": 7429770660720.636,
    "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": 7429770660728.855, "dur": 4.529,
    "args": {
      "External id": 4094,      "device": 2, "stream": 0,      "correlation": 1361, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1361, "pid": 2, "tid": 0, "ts": 7429770660728.855,
    "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": 7429770660733.385, "dur": 4.530,
    "args": {
      "External id": 4095,      "device": 2, "stream": 0,      "correlation": 1362, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1362, "pid": 2, "tid": 0, "ts": 7429770660733.385,
    "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": 7429770660737.916, "dur": 4.810,
    "args": {
      "External id": 4096,      "device": 2, "stream": 0,      "correlation": 1363, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1363, "pid": 2, "tid": 0, "ts": 7429770660737.916,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770660742.727, "dur": 4.970,
    "args": {
      "External id": 4098,      "device": 2, "stream": 0,      "correlation": 1364, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1364, "pid": 2, "tid": 0, "ts": 7429770660742.727,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770660747.698, "dur": 63.425,
    "args": {
      "External id": 4106,      "device": 2, "stream": 0,      "correlation": 1368, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1368, "pid": 2, "tid": 0, "ts": 7429770660747.698,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770660811.124, "dur": 50.796,
    "args": {
      "External id": 4119,      "device": 2, "stream": 0,      "correlation": 1372, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1372, "pid": 2, "tid": 0, "ts": 7429770660811.124,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770660861.921, "dur": 50.314,
    "args": {
      "External id": 4132,      "device": 2, "stream": 0,      "correlation": 1376, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1376, "pid": 2, "tid": 0, "ts": 7429770660861.921,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770660912.236, "dur": 4.730,
    "args": {
      "External id": 4142,      "device": 2, "stream": 0,      "correlation": 1377, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1377, "pid": 2, "tid": 0, "ts": 7429770660912.236,
    "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": 7429770660916.967, "dur": 4.650,
    "args": {
      "External id": 4147,      "device": 2, "stream": 0,      "correlation": 1378, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1378, "pid": 2, "tid": 0, "ts": 7429770660916.967,
    "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": 7429770660927.110, "dur": 5.652,
    "args": {
      "External id": 4148,      "device": 2, "stream": 0,      "correlation": 1379, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1379, "pid": 2, "tid": 0, "ts": 7429770660927.110,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770660932.763, "dur": 5.211,
    "args": {
      "External id": 4149,      "device": 2, "stream": 0,      "correlation": 1380, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1380, "pid": 2, "tid": 0, "ts": 7429770660932.763,
    "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::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": 7429770660937.975, "dur": 5.171,
    "args": {
      "External id": 4150,      "device": 2, "stream": 0,      "correlation": 1381, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1381, "pid": 2, "tid": 0, "ts": 7429770660937.975,
    "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": 7429770660943.147, "dur": 4.850,
    "args": {
      "External id": 4151,      "device": 2, "stream": 0,      "correlation": 1382, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1382, "pid": 2, "tid": 0, "ts": 7429770660943.147,
    "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": 7429770660947.998, "dur": 4.369,
    "args": {
      "External id": 4156,      "device": 2, "stream": 0,      "correlation": 1383, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1383, "pid": 2, "tid": 0, "ts": 7429770660947.998,
    "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": 7429770660952.368, "dur": 6.133,
    "args": {
      "External id": 4157,      "device": 2, "stream": 0,      "correlation": 1384, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1384, "pid": 2, "tid": 0, "ts": 7429770660952.368,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770660958.502, "dur": 5.332,
    "args": {
      "External id": 4158,      "device": 2, "stream": 0,      "correlation": 1385, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1385, "pid": 2, "tid": 0, "ts": 7429770660958.502,
    "cat": "ac2g", "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": 7429770660963.835, "dur": 5.451,
    "args": {
      "External id": 4159,      "device": 2, "stream": 0,      "correlation": 1386, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1386, "pid": 2, "tid": 0, "ts": 7429770660963.835,
    "cat": "ac2g", "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": 7429770660969.287, "dur": 5.532,
    "args": {
      "External id": 4176,      "device": 2, "stream": 0,      "correlation": 1387, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1387, "pid": 2, "tid": 0, "ts": 7429770660969.287,
    "cat": "ac2g", "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": 7429770660974.820, "dur": 5.572,
    "args": {
      "External id": 4194,      "device": 2, "stream": 0,      "correlation": 1388, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1388, "pid": 2, "tid": 0, "ts": 7429770660974.820,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770660980.393, "dur": 9.821,
    "args": {
      "External id": 4207,      "device": 2, "stream": 0,      "correlation": 1394, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1394, "pid": 2, "tid": 0, "ts": 7429770660980.393,
    "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": 7429770660990.215, "dur": 4.450,
    "args": {
      "External id": 4210,      "device": 2, "stream": 0,      "correlation": 1395, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1395, "pid": 2, "tid": 0, "ts": 7429770660990.215,
    "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::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": 7429770660994.666, "dur": 4.810,
    "args": {
      "External id": 4217,      "device": 2, "stream": 0,      "correlation": 1396, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1396, "pid": 2, "tid": 0, "ts": 7429770660994.666,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770660999.477, "dur": 4.930,
    "args": {
      "External id": 4220,      "device": 2, "stream": 0,      "correlation": 1397, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 1397, "pid": 2, "tid": 0, "ts": 7429770660999.477,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770661004.408, "dur": 10.984,
    "args": {
      "External id": 4232,      "device": 2, "stream": 0,      "correlation": 1401, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1401, "pid": 2, "tid": 0, "ts": 7429770661004.408,
    "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::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::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": 7429770661015.393, "dur": 5.251,
    "args": {
      "External id": 4241,      "device": 2, "stream": 0,      "correlation": 1402, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1402, "pid": 2, "tid": 0, "ts": 7429770661015.393,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770661020.645, "dur": 64.628,
    "args": {
      "External id": 4251,      "device": 2, "stream": 0,      "correlation": 1406, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1406, "pid": 2, "tid": 0, "ts": 7429770661020.645,
    "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": 7429770661085.274, "dur": 4.649,
    "args": {
      "External id": 4254,      "device": 2, "stream": 0,      "correlation": 1407, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1407, "pid": 2, "tid": 0, "ts": 7429770661085.274,
    "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": 7429770661089.924, "dur": 5.933,
    "args": {
      "External id": 4256,      "device": 2, "stream": 0,      "correlation": 1408, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1408, "pid": 2, "tid": 0, "ts": 7429770661089.924,
    "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": 7429770661095.858, "dur": 7.897,
    "args": {
      "External id": 4259,      "device": 2, "stream": 0,      "correlation": 1409, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1409, "pid": 2, "tid": 0, "ts": 7429770661095.858,
    "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": 7429770661103.756, "dur": 4.930,
    "args": {
      "External id": 4260,      "device": 2, "stream": 0,      "correlation": 1410, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1410, "pid": 2, "tid": 0, "ts": 7429770661103.756,
    "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": 7429770661108.687, "dur": 5.251,
    "args": {
      "External id": 4261,      "device": 2, "stream": 0,      "correlation": 1411, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1411, "pid": 2, "tid": 0, "ts": 7429770661108.687,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770661113.939, "dur": 5.372,
    "args": {
      "External id": 4262,      "device": 2, "stream": 0,      "correlation": 1412, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1412, "pid": 2, "tid": 0, "ts": 7429770661113.939,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770661119.312, "dur": 4.890,
    "args": {
      "External id": 4264,      "device": 2, "stream": 0,      "correlation": 1413, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1413, "pid": 2, "tid": 0, "ts": 7429770661119.312,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770661124.203, "dur": 253.140,
    "args": {
      "External id": 4272,      "device": 2, "stream": 0,      "correlation": 1417, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1417, "pid": 2, "tid": 0, "ts": 7429770661124.203,
    "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": 7429770661377.344, "dur": 5.252,
    "args": {
      "External id": 4275,      "device": 2, "stream": 0,      "correlation": 1418, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1418, "pid": 2, "tid": 0, "ts": 7429770661377.344,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770661382.597, "dur": 252.699,
    "args": {
      "External id": 4283,      "device": 2, "stream": 0,      "correlation": 1422, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1422, "pid": 2, "tid": 0, "ts": 7429770661382.597,
    "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": 7429770661635.297, "dur": 5.492,
    "args": {
      "External id": 4286,      "device": 2, "stream": 0,      "correlation": 1423, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1423, "pid": 2, "tid": 0, "ts": 7429770661635.297,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770661640.790, "dur": 164.937,
    "args": {
      "External id": 4294,      "device": 2, "stream": 0,      "correlation": 1427, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1427, "pid": 2, "tid": 0, "ts": 7429770661640.790,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770661805.728, "dur": 7.336,
    "args": {
      "External id": 4294,      "device": 2, "stream": 0,      "correlation": 1428, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1428, "pid": 2, "tid": 0, "ts": 7429770661805.728,
    "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": 7429770661813.065, "dur": 5.372,
    "args": {
      "External id": 4297,      "device": 2, "stream": 0,      "correlation": 1429, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1429, "pid": 2, "tid": 0, "ts": 7429770661813.065,
    "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": 7429770661818.438, "dur": 6.173,
    "args": {
      "External id": 4299,      "device": 2, "stream": 0,      "correlation": 1430, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1430, "pid": 2, "tid": 0, "ts": 7429770661818.438,
    "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": 7429770661824.612, "dur": 7.897,
    "args": {
      "External id": 4302,      "device": 2, "stream": 0,      "correlation": 1431, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1431, "pid": 2, "tid": 0, "ts": 7429770661824.612,
    "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": 7429770661832.510, "dur": 4.409,
    "args": {
      "External id": 4303,      "device": 2, "stream": 0,      "correlation": 1432, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1432, "pid": 2, "tid": 0, "ts": 7429770661832.510,
    "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": 7429770661836.920, "dur": 4.770,
    "args": {
      "External id": 4304,      "device": 2, "stream": 0,      "correlation": 1433, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1433, "pid": 2, "tid": 0, "ts": 7429770661836.920,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770661841.691, "dur": 4.730,
    "args": {
      "External id": 4305,      "device": 2, "stream": 0,      "correlation": 1434, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1434, "pid": 2, "tid": 0, "ts": 7429770661841.691,
    "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": 7429770661846.422, "dur": 4.970,
    "args": {
      "External id": 4307,      "device": 2, "stream": 0,      "correlation": 1435, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1435, "pid": 2, "tid": 0, "ts": 7429770661846.422,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770661851.393, "dur": 63.665,
    "args": {
      "External id": 4315,      "device": 2, "stream": 0,      "correlation": 1439, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1439, "pid": 2, "tid": 0, "ts": 7429770661851.393,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770661915.059, "dur": 50.315,
    "args": {
      "External id": 4328,      "device": 2, "stream": 0,      "correlation": 1443, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1443, "pid": 2, "tid": 0, "ts": 7429770661915.059,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770661965.375, "dur": 49.874,
    "args": {
      "External id": 4341,      "device": 2, "stream": 0,      "correlation": 1447, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1447, "pid": 2, "tid": 0, "ts": 7429770661965.375,
    "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": 7429770662015.250, "dur": 4.689,
    "args": {
      "External id": 4351,      "device": 2, "stream": 0,      "correlation": 1448, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1448, "pid": 2, "tid": 0, "ts": 7429770662015.250,
    "cat": "ac2g", "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": 7429770662019.940, "dur": 4.610,
    "args": {
      "External id": 4356,      "device": 2, "stream": 0,      "correlation": 1449, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1449, "pid": 2, "tid": 0, "ts": 7429770662019.940,
    "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": 7429770662024.551, "dur": 6.374,
    "args": {
      "External id": 4357,      "device": 2, "stream": 0,      "correlation": 1450, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1450, "pid": 2, "tid": 0, "ts": 7429770662024.551,
    "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": 7429770662030.926, "dur": 5.050,
    "args": {
      "External id": 4358,      "device": 2, "stream": 0,      "correlation": 1451, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1451, "pid": 2, "tid": 0, "ts": 7429770662030.926,
    "cat": "ac2g", "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": 7429770662035.977, "dur": 5.291,
    "args": {
      "External id": 4359,      "device": 2, "stream": 0,      "correlation": 1452, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1452, "pid": 2, "tid": 0, "ts": 7429770662035.977,
    "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": 7429770662041.269, "dur": 5.091,
    "args": {
      "External id": 4360,      "device": 2, "stream": 0,      "correlation": 1453, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1453, "pid": 2, "tid": 0, "ts": 7429770662041.269,
    "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::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": 7429770662046.361, "dur": 4.329,
    "args": {
      "External id": 4365,      "device": 2, "stream": 0,      "correlation": 1454, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1454, "pid": 2, "tid": 0, "ts": 7429770662046.361,
    "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": 7429770662050.691, "dur": 6.213,
    "args": {
      "External id": 4366,      "device": 2, "stream": 0,      "correlation": 1455, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1455, "pid": 2, "tid": 0, "ts": 7429770662050.691,
    "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": 7429770662056.905, "dur": 4.690,
    "args": {
      "External id": 4367,      "device": 2, "stream": 0,      "correlation": 1456, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1456, "pid": 2, "tid": 0, "ts": 7429770662056.905,
    "cat": "ac2g", "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": 7429770662061.596, "dur": 4.690,
    "args": {
      "External id": 4368,      "device": 2, "stream": 0,      "correlation": 1457, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1457, "pid": 2, "tid": 0, "ts": 7429770662061.596,
    "cat": "ac2g", "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": 7429770662066.287, "dur": 5.251,
    "args": {
      "External id": 4385,      "device": 2, "stream": 0,      "correlation": 1458, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1458, "pid": 2, "tid": 0, "ts": 7429770662066.287,
    "cat": "ac2g", "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": 7429770662071.539, "dur": 5.692,
    "args": {
      "External id": 4403,      "device": 2, "stream": 0,      "correlation": 1459, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1459, "pid": 2, "tid": 0, "ts": 7429770662071.539,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770662077.232, "dur": 9.701,
    "args": {
      "External id": 4416,      "device": 2, "stream": 0,      "correlation": 1465, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1465, "pid": 2, "tid": 0, "ts": 7429770662077.232,
    "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": 7429770662086.934, "dur": 4.890,
    "args": {
      "External id": 4419,      "device": 2, "stream": 0,      "correlation": 1466, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1466, "pid": 2, "tid": 0, "ts": 7429770662086.934,
    "cat": "ac2g", "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": 7429770662091.825, "dur": 4.811,
    "args": {
      "External id": 4426,      "device": 2, "stream": 0,      "correlation": 1467, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1467, "pid": 2, "tid": 0, "ts": 7429770662091.825,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770662096.637, "dur": 5.010,
    "args": {
      "External id": 4429,      "device": 2, "stream": 0,      "correlation": 1468, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 1468, "pid": 2, "tid": 0, "ts": 7429770662096.637,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770662101.648, "dur": 11.145,
    "args": {
      "External id": 4441,      "device": 2, "stream": 0,      "correlation": 1472, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1472, "pid": 2, "tid": 0, "ts": 7429770662101.648,
    "cat": "ac2g", "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": 7429770662112.794, "dur": 5.652,
    "args": {
      "External id": 4450,      "device": 2, "stream": 0,      "correlation": 1473, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1473, "pid": 2, "tid": 0, "ts": 7429770662112.794,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770662118.447, "dur": 64.306,
    "args": {
      "External id": 4460,      "device": 2, "stream": 0,      "correlation": 1477, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1477, "pid": 2, "tid": 0, "ts": 7429770662118.447,
    "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": 7429770662182.754, "dur": 5.211,
    "args": {
      "External id": 4463,      "device": 2, "stream": 0,      "correlation": 1478, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1478, "pid": 2, "tid": 0, "ts": 7429770662182.754,
    "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": 7429770662187.966, "dur": 6.214,
    "args": {
      "External id": 4465,      "device": 2, "stream": 0,      "correlation": 1479, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1479, "pid": 2, "tid": 0, "ts": 7429770662187.966,
    "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": 7429770662194.181, "dur": 7.857,
    "args": {
      "External id": 4468,      "device": 2, "stream": 0,      "correlation": 1480, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1480, "pid": 2, "tid": 0, "ts": 7429770662194.181,
    "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": 7429770662202.039, "dur": 4.449,
    "args": {
      "External id": 4469,      "device": 2, "stream": 0,      "correlation": 1481, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1481, "pid": 2, "tid": 0, "ts": 7429770662202.039,
    "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": 7429770662206.489, "dur": 5.491,
    "args": {
      "External id": 4470,      "device": 2, "stream": 0,      "correlation": 1482, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1482, "pid": 2, "tid": 0, "ts": 7429770662206.489,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770662211.981, "dur": 5.452,
    "args": {
      "External id": 4471,      "device": 2, "stream": 0,      "correlation": 1483, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1483, "pid": 2, "tid": 0, "ts": 7429770662211.981,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770662217.434, "dur": 5.452,
    "args": {
      "External id": 4473,      "device": 2, "stream": 0,      "correlation": 1484, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1484, "pid": 2, "tid": 0, "ts": 7429770662217.434,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770662222.887, "dur": 252.980,
    "args": {
      "External id": 4481,      "device": 2, "stream": 0,      "correlation": 1488, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1488, "pid": 2, "tid": 0, "ts": 7429770662222.887,
    "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": 7429770662475.868, "dur": 5.251,
    "args": {
      "External id": 4484,      "device": 2, "stream": 0,      "correlation": 1489, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1489, "pid": 2, "tid": 0, "ts": 7429770662475.868,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770662481.120, "dur": 252.579,
    "args": {
      "External id": 4492,      "device": 2, "stream": 0,      "correlation": 1495, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1495, "pid": 2, "tid": 0, "ts": 7429770662481.120,
    "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": 7429770662733.700, "dur": 5.932,
    "args": {
      "External id": 4495,      "device": 2, "stream": 0,      "correlation": 1496, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1496, "pid": 2, "tid": 0, "ts": 7429770662733.700,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770662739.633, "dur": 164.818,
    "args": {
      "External id": 4503,      "device": 2, "stream": 0,      "correlation": 1500, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1500, "pid": 2, "tid": 0, "ts": 7429770662739.633,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770662904.452, "dur": 7.657,
    "args": {
      "External id": 4503,      "device": 2, "stream": 0,      "correlation": 1501, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1501, "pid": 2, "tid": 0, "ts": 7429770662904.452,
    "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": 7429770662912.110, "dur": 5.090,
    "args": {
      "External id": 4506,      "device": 2, "stream": 0,      "correlation": 1502, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1502, "pid": 2, "tid": 0, "ts": 7429770662912.110,
    "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": 7429770662917.201, "dur": 4.890,
    "args": {
      "External id": 4508,      "device": 2, "stream": 0,      "correlation": 1503, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1503, "pid": 2, "tid": 0, "ts": 7429770662917.201,
    "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": 7429770662922.092, "dur": 8.178,
    "args": {
      "External id": 4511,      "device": 2, "stream": 0,      "correlation": 1504, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1504, "pid": 2, "tid": 0, "ts": 7429770662922.092,
    "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": 7429770662930.271, "dur": 5.011,
    "args": {
      "External id": 4512,      "device": 2, "stream": 0,      "correlation": 1505, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1505, "pid": 2, "tid": 0, "ts": 7429770662930.271,
    "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": 7429770662935.283, "dur": 5.291,
    "args": {
      "External id": 4513,      "device": 2, "stream": 0,      "correlation": 1506, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1506, "pid": 2, "tid": 0, "ts": 7429770662935.283,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770662940.575, "dur": 5.050,
    "args": {
      "External id": 4514,      "device": 2, "stream": 0,      "correlation": 1507, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1507, "pid": 2, "tid": 0, "ts": 7429770662940.575,
    "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": 7429770662945.626, "dur": 5.372,
    "args": {
      "External id": 4516,      "device": 2, "stream": 0,      "correlation": 1508, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1508, "pid": 2, "tid": 0, "ts": 7429770662945.626,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770662950.999, "dur": 63.866,
    "args": {
      "External id": 4524,      "device": 2, "stream": 0,      "correlation": 1512, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1512, "pid": 2, "tid": 0, "ts": 7429770662950.999,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770663014.866, "dur": 51.156,
    "args": {
      "External id": 4537,      "device": 2, "stream": 0,      "correlation": 1516, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1516, "pid": 2, "tid": 0, "ts": 7429770663014.866,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770663066.023, "dur": 50.876,
    "args": {
      "External id": 4550,      "device": 2, "stream": 0,      "correlation": 1520, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1520, "pid": 2, "tid": 0, "ts": 7429770663066.023,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770663116.900, "dur": 4.730,
    "args": {
      "External id": 4560,      "device": 2, "stream": 0,      "correlation": 1521, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1521, "pid": 2, "tid": 0, "ts": 7429770663116.900,
    "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::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": 7429770663121.631, "dur": 4.770,
    "args": {
      "External id": 4565,      "device": 2, "stream": 0,      "correlation": 1522, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1522, "pid": 2, "tid": 0, "ts": 7429770663121.631,
    "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": 7429770663126.402, "dur": 6.373,
    "args": {
      "External id": 4566,      "device": 2, "stream": 0,      "correlation": 1523, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1523, "pid": 2, "tid": 0, "ts": 7429770663126.402,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770663132.776, "dur": 4.810,
    "args": {
      "External id": 4567,      "device": 2, "stream": 0,      "correlation": 1524, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1524, "pid": 2, "tid": 0, "ts": 7429770663132.776,
    "cat": "ac2g", "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": 7429770663137.587, "dur": 5.131,
    "args": {
      "External id": 4568,      "device": 2, "stream": 0,      "correlation": 1525, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1525, "pid": 2, "tid": 0, "ts": 7429770663137.587,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770663142.719, "dur": 5.091,
    "args": {
      "External id": 4569,      "device": 2, "stream": 0,      "correlation": 1526, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1526, "pid": 2, "tid": 0, "ts": 7429770663142.719,
    "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::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": 7429770663147.811, "dur": 5.131,
    "args": {
      "External id": 4574,      "device": 2, "stream": 0,      "correlation": 1527, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1527, "pid": 2, "tid": 0, "ts": 7429770663147.811,
    "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": 7429770663152.943, "dur": 6.213,
    "args": {
      "External id": 4575,      "device": 2, "stream": 0,      "correlation": 1528, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1528, "pid": 2, "tid": 0, "ts": 7429770663152.943,
    "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": 7429770663159.157, "dur": 4.690,
    "args": {
      "External id": 4576,      "device": 2, "stream": 0,      "correlation": 1529, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1529, "pid": 2, "tid": 0, "ts": 7429770663159.157,
    "cat": "ac2g", "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": 7429770663163.848, "dur": 4.730,
    "args": {
      "External id": 4577,      "device": 2, "stream": 0,      "correlation": 1530, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1530, "pid": 2, "tid": 0, "ts": 7429770663163.848,
    "cat": "ac2g", "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": 7429770663168.579, "dur": 5.411,
    "args": {
      "External id": 4594,      "device": 2, "stream": 0,      "correlation": 1531, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1531, "pid": 2, "tid": 0, "ts": 7429770663168.579,
    "cat": "ac2g", "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": 7429770663173.991, "dur": 5.532,
    "args": {
      "External id": 4612,      "device": 2, "stream": 0,      "correlation": 1532, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1532, "pid": 2, "tid": 0, "ts": 7429770663173.991,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770663179.524, "dur": 9.861,
    "args": {
      "External id": 4625,      "device": 2, "stream": 0,      "correlation": 1538, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1538, "pid": 2, "tid": 0, "ts": 7429770663179.524,
    "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": 7429770663189.386, "dur": 4.450,
    "args": {
      "External id": 4628,      "device": 2, "stream": 0,      "correlation": 1539, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1539, "pid": 2, "tid": 0, "ts": 7429770663189.386,
    "cat": "ac2g", "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": 7429770663193.837, "dur": 4.770,
    "args": {
      "External id": 4635,      "device": 2, "stream": 0,      "correlation": 1540, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1540, "pid": 2, "tid": 0, "ts": 7429770663193.837,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770663198.608, "dur": 5.130,
    "args": {
      "External id": 4638,      "device": 2, "stream": 0,      "correlation": 1541, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 1541, "pid": 2, "tid": 0, "ts": 7429770663198.608,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770663203.739, "dur": 10.985,
    "args": {
      "External id": 4650,      "device": 2, "stream": 0,      "correlation": 1545, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1545, "pid": 2, "tid": 0, "ts": 7429770663203.739,
    "cat": "ac2g", "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": 7429770663214.725, "dur": 5.090,
    "args": {
      "External id": 4659,      "device": 2, "stream": 0,      "correlation": 1546, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1546, "pid": 2, "tid": 0, "ts": 7429770663214.725,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770663219.816, "dur": 63.866,
    "args": {
      "External id": 4669,      "device": 2, "stream": 0,      "correlation": 1550, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1550, "pid": 2, "tid": 0, "ts": 7429770663219.816,
    "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": 7429770663283.683, "dur": 5.291,
    "args": {
      "External id": 4672,      "device": 2, "stream": 0,      "correlation": 1551, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1551, "pid": 2, "tid": 0, "ts": 7429770663283.683,
    "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": 7429770663288.975, "dur": 5.532,
    "args": {
      "External id": 4674,      "device": 2, "stream": 0,      "correlation": 1552, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1552, "pid": 2, "tid": 0, "ts": 7429770663288.975,
    "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": 7429770663294.508, "dur": 8.178,
    "args": {
      "External id": 4677,      "device": 2, "stream": 0,      "correlation": 1553, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1553, "pid": 2, "tid": 0, "ts": 7429770663294.508,
    "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": 7429770663302.687, "dur": 4.409,
    "args": {
      "External id": 4678,      "device": 2, "stream": 0,      "correlation": 1554, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1554, "pid": 2, "tid": 0, "ts": 7429770663302.687,
    "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": 7429770663307.097, "dur": 4.649,
    "args": {
      "External id": 4679,      "device": 2, "stream": 0,      "correlation": 1555, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1555, "pid": 2, "tid": 0, "ts": 7429770663307.097,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770663311.747, "dur": 4.730,
    "args": {
      "External id": 4680,      "device": 2, "stream": 0,      "correlation": 1556, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1556, "pid": 2, "tid": 0, "ts": 7429770663311.747,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770663316.478, "dur": 4.931,
    "args": {
      "External id": 4682,      "device": 2, "stream": 0,      "correlation": 1557, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1557, "pid": 2, "tid": 0, "ts": 7429770663316.478,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770663321.410, "dur": 253.501,
    "args": {
      "External id": 4690,      "device": 2, "stream": 0,      "correlation": 1561, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1561, "pid": 2, "tid": 0, "ts": 7429770663321.410,
    "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": 7429770663574.912, "dur": 5.251,
    "args": {
      "External id": 4693,      "device": 2, "stream": 0,      "correlation": 1562, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1562, "pid": 2, "tid": 0, "ts": 7429770663574.912,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770663580.164, "dur": 252.940,
    "args": {
      "External id": 4701,      "device": 2, "stream": 0,      "correlation": 1566, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1566, "pid": 2, "tid": 0, "ts": 7429770663580.164,
    "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": 7429770663833.105, "dur": 5.772,
    "args": {
      "External id": 4704,      "device": 2, "stream": 0,      "correlation": 1567, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1567, "pid": 2, "tid": 0, "ts": 7429770663833.105,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770663838.878, "dur": 164.096,
    "args": {
      "External id": 4712,      "device": 2, "stream": 0,      "correlation": 1571, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1571, "pid": 2, "tid": 0, "ts": 7429770663838.878,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770664002.975, "dur": 7.296,
    "args": {
      "External id": 4712,      "device": 2, "stream": 0,      "correlation": 1572, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1572, "pid": 2, "tid": 0, "ts": 7429770664002.975,
    "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": 7429770664010.272, "dur": 5.371,
    "args": {
      "External id": 4715,      "device": 2, "stream": 0,      "correlation": 1573, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1573, "pid": 2, "tid": 0, "ts": 7429770664010.272,
    "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": 7429770664015.644, "dur": 6.294,
    "args": {
      "External id": 4717,      "device": 2, "stream": 0,      "correlation": 1574, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1574, "pid": 2, "tid": 0, "ts": 7429770664015.644,
    "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": 7429770664021.939, "dur": 8.177,
    "args": {
      "External id": 4720,      "device": 2, "stream": 0,      "correlation": 1575, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1575, "pid": 2, "tid": 0, "ts": 7429770664021.939,
    "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": 7429770664030.117, "dur": 4.450,
    "args": {
      "External id": 4721,      "device": 2, "stream": 0,      "correlation": 1576, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1576, "pid": 2, "tid": 0, "ts": 7429770664030.117,
    "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": 7429770664034.568, "dur": 4.569,
    "args": {
      "External id": 4722,      "device": 2, "stream": 0,      "correlation": 1577, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1577, "pid": 2, "tid": 0, "ts": 7429770664034.568,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770664039.138, "dur": 4.730,
    "args": {
      "External id": 4723,      "device": 2, "stream": 0,      "correlation": 1578, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1578, "pid": 2, "tid": 0, "ts": 7429770664039.138,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770664043.869, "dur": 5.051,
    "args": {
      "External id": 4725,      "device": 2, "stream": 0,      "correlation": 1579, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1579, "pid": 2, "tid": 0, "ts": 7429770664043.869,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770664048.921, "dur": 64.587,
    "args": {
      "External id": 4733,      "device": 2, "stream": 0,      "correlation": 1583, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1583, "pid": 2, "tid": 0, "ts": 7429770664048.921,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770664113.509, "dur": 50.956,
    "args": {
      "External id": 4746,      "device": 2, "stream": 0,      "correlation": 1587, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1587, "pid": 2, "tid": 0, "ts": 7429770664113.509,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770664164.466, "dur": 50.635,
    "args": {
      "External id": 4759,      "device": 2, "stream": 0,      "correlation": 1591, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1591, "pid": 2, "tid": 0, "ts": 7429770664164.466,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770664215.102, "dur": 5.171,
    "args": {
      "External id": 4769,      "device": 2, "stream": 0,      "correlation": 1592, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1592, "pid": 2, "tid": 0, "ts": 7429770664215.102,
    "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": 7429770664220.274, "dur": 5.412,
    "args": {
      "External id": 4774,      "device": 2, "stream": 0,      "correlation": 1593, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1593, "pid": 2, "tid": 0, "ts": 7429770664220.274,
    "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": 7429770664225.687, "dur": 6.333,
    "args": {
      "External id": 4775,      "device": 2, "stream": 0,      "correlation": 1594, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1594, "pid": 2, "tid": 0, "ts": 7429770664225.687,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770664232.021, "dur": 5.051,
    "args": {
      "External id": 4776,      "device": 2, "stream": 0,      "correlation": 1595, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1595, "pid": 2, "tid": 0, "ts": 7429770664232.021,
    "cat": "ac2g", "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": 7429770664237.073, "dur": 5.090,
    "args": {
      "External id": 4777,      "device": 2, "stream": 0,      "correlation": 1596, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1596, "pid": 2, "tid": 0, "ts": 7429770664237.073,
    "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": 7429770664242.164, "dur": 4.931,
    "args": {
      "External id": 4778,      "device": 2, "stream": 0,      "correlation": 1597, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1597, "pid": 2, "tid": 0, "ts": 7429770664242.164,
    "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::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": 7429770664247.096, "dur": 4.369,
    "args": {
      "External id": 4783,      "device": 2, "stream": 0,      "correlation": 1598, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1598, "pid": 2, "tid": 0, "ts": 7429770664247.096,
    "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": 7429770664251.466, "dur": 6.173,
    "args": {
      "External id": 4784,      "device": 2, "stream": 0,      "correlation": 1599, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1599, "pid": 2, "tid": 0, "ts": 7429770664251.466,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770664257.640, "dur": 4.529,
    "args": {
      "External id": 4785,      "device": 2, "stream": 0,      "correlation": 1600, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1600, "pid": 2, "tid": 0, "ts": 7429770664257.640,
    "cat": "ac2g", "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": 7429770664262.170, "dur": 4.650,
    "args": {
      "External id": 4786,      "device": 2, "stream": 0,      "correlation": 1601, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1601, "pid": 2, "tid": 0, "ts": 7429770664262.170,
    "cat": "ac2g", "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": 7429770664266.821, "dur": 4.970,
    "args": {
      "External id": 4803,      "device": 2, "stream": 0,      "correlation": 1602, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1602, "pid": 2, "tid": 0, "ts": 7429770664266.821,
    "cat": "ac2g", "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": 7429770664271.792, "dur": 5.452,
    "args": {
      "External id": 4821,      "device": 2, "stream": 0,      "correlation": 1603, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1603, "pid": 2, "tid": 0, "ts": 7429770664271.792,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770664277.245, "dur": 9.621,
    "args": {
      "External id": 4834,      "device": 2, "stream": 0,      "correlation": 1609, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1609, "pid": 2, "tid": 0, "ts": 7429770664277.245,
    "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": 7429770664286.867, "dur": 4.610,
    "args": {
      "External id": 4837,      "device": 2, "stream": 0,      "correlation": 1610, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1610, "pid": 2, "tid": 0, "ts": 7429770664286.867,
    "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": 7429770664291.478, "dur": 4.930,
    "args": {
      "External id": 4844,      "device": 2, "stream": 0,      "correlation": 1611, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1611, "pid": 2, "tid": 0, "ts": 7429770664291.478,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770664296.409, "dur": 5.732,
    "args": {
      "External id": 4847,      "device": 2, "stream": 0,      "correlation": 1612, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 1612, "pid": 2, "tid": 0, "ts": 7429770664296.409,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770664302.142, "dur": 10.824,
    "args": {
      "External id": 4859,      "device": 2, "stream": 0,      "correlation": 1616, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1616, "pid": 2, "tid": 0, "ts": 7429770664302.142,
    "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": 7429770664312.967, "dur": 5.211,
    "args": {
      "External id": 4868,      "device": 2, "stream": 0,      "correlation": 1617, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1617, "pid": 2, "tid": 0, "ts": 7429770664312.967,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770664318.179, "dur": 62.944,
    "args": {
      "External id": 4878,      "device": 2, "stream": 0,      "correlation": 1621, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1621, "pid": 2, "tid": 0, "ts": 7429770664318.179,
    "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": 7429770664381.124, "dur": 4.769,
    "args": {
      "External id": 4881,      "device": 2, "stream": 0,      "correlation": 1622, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1622, "pid": 2, "tid": 0, "ts": 7429770664381.124,
    "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": 7429770664385.894, "dur": 5.372,
    "args": {
      "External id": 4883,      "device": 2, "stream": 0,      "correlation": 1623, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1623, "pid": 2, "tid": 0, "ts": 7429770664385.894,
    "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": 7429770664391.267, "dur": 7.937,
    "args": {
      "External id": 4886,      "device": 2, "stream": 0,      "correlation": 1624, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1624, "pid": 2, "tid": 0, "ts": 7429770664391.267,
    "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": 7429770664399.205, "dur": 4.449,
    "args": {
      "External id": 4887,      "device": 2, "stream": 0,      "correlation": 1625, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1625, "pid": 2, "tid": 0, "ts": 7429770664399.205,
    "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": 7429770664403.655, "dur": 4.650,
    "args": {
      "External id": 4888,      "device": 2, "stream": 0,      "correlation": 1626, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1626, "pid": 2, "tid": 0, "ts": 7429770664403.655,
    "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": 7429770664408.306, "dur": 4.770,
    "args": {
      "External id": 4889,      "device": 2, "stream": 0,      "correlation": 1627, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1627, "pid": 2, "tid": 0, "ts": 7429770664408.306,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770664413.077, "dur": 4.890,
    "args": {
      "External id": 4891,      "device": 2, "stream": 0,      "correlation": 1628, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1628, "pid": 2, "tid": 0, "ts": 7429770664413.077,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770664417.968, "dur": 252.900,
    "args": {
      "External id": 4899,      "device": 2, "stream": 0,      "correlation": 1632, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1632, "pid": 2, "tid": 0, "ts": 7429770664417.968,
    "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": 7429770664670.869, "dur": 5.251,
    "args": {
      "External id": 4902,      "device": 2, "stream": 0,      "correlation": 1633, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1633, "pid": 2, "tid": 0, "ts": 7429770664670.869,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770664676.121, "dur": 252.980,
    "args": {
      "External id": 4910,      "device": 2, "stream": 0,      "correlation": 1637, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1637, "pid": 2, "tid": 0, "ts": 7429770664676.121,
    "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": 7429770664929.102, "dur": 5.572,
    "args": {
      "External id": 4913,      "device": 2, "stream": 0,      "correlation": 1638, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1638, "pid": 2, "tid": 0, "ts": 7429770664929.102,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770664934.675, "dur": 165.138,
    "args": {
      "External id": 4921,      "device": 2, "stream": 0,      "correlation": 1642, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1642, "pid": 2, "tid": 0, "ts": 7429770664934.675,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770665099.814, "dur": 7.336,
    "args": {
      "External id": 4921,      "device": 2, "stream": 0,      "correlation": 1643, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1643, "pid": 2, "tid": 0, "ts": 7429770665099.814,
    "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": 7429770665107.151, "dur": 4.970,
    "args": {
      "External id": 4924,      "device": 2, "stream": 0,      "correlation": 1644, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1644, "pid": 2, "tid": 0, "ts": 7429770665107.151,
    "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": 7429770665112.122, "dur": 5.693,
    "args": {
      "External id": 4926,      "device": 2, "stream": 0,      "correlation": 1645, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1645, "pid": 2, "tid": 0, "ts": 7429770665112.122,
    "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": 7429770665117.816, "dur": 7.937,
    "args": {
      "External id": 4929,      "device": 2, "stream": 0,      "correlation": 1646, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1646, "pid": 2, "tid": 0, "ts": 7429770665117.816,
    "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": 7429770665125.754, "dur": 5.211,
    "args": {
      "External id": 4930,      "device": 2, "stream": 0,      "correlation": 1647, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1647, "pid": 2, "tid": 0, "ts": 7429770665125.754,
    "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": 7429770665130.966, "dur": 4.930,
    "args": {
      "External id": 4931,      "device": 2, "stream": 0,      "correlation": 1648, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1648, "pid": 2, "tid": 0, "ts": 7429770665130.966,
    "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": 7429770665135.897, "dur": 4.730,
    "args": {
      "External id": 4932,      "device": 2, "stream": 0,      "correlation": 1649, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1649, "pid": 2, "tid": 0, "ts": 7429770665135.897,
    "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": 7429770665140.628, "dur": 4.890,
    "args": {
      "External id": 4934,      "device": 2, "stream": 0,      "correlation": 1650, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1650, "pid": 2, "tid": 0, "ts": 7429770665140.628,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770665145.519, "dur": 64.588,
    "args": {
      "External id": 4942,      "device": 2, "stream": 0,      "correlation": 1654, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1654, "pid": 2, "tid": 0, "ts": 7429770665145.519,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770665210.108, "dur": 50.795,
    "args": {
      "External id": 4955,      "device": 2, "stream": 0,      "correlation": 1658, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1658, "pid": 2, "tid": 0, "ts": 7429770665210.108,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770665260.904, "dur": 49.834,
    "args": {
      "External id": 4968,      "device": 2, "stream": 0,      "correlation": 1662, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1662, "pid": 2, "tid": 0, "ts": 7429770665260.904,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770665310.739, "dur": 4.689,
    "args": {
      "External id": 4978,      "device": 2, "stream": 0,      "correlation": 1663, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1663, "pid": 2, "tid": 0, "ts": 7429770665310.739,
    "cat": "ac2g", "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": 7429770665315.429, "dur": 5.332,
    "args": {
      "External id": 4983,      "device": 2, "stream": 0,      "correlation": 1664, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1664, "pid": 2, "tid": 0, "ts": 7429770665315.429,
    "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": 7429770665320.762, "dur": 6.253,
    "args": {
      "External id": 4984,      "device": 2, "stream": 0,      "correlation": 1665, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1665, "pid": 2, "tid": 0, "ts": 7429770665320.762,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770665327.016, "dur": 5.131,
    "args": {
      "External id": 4985,      "device": 2, "stream": 0,      "correlation": 1666, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1666, "pid": 2, "tid": 0, "ts": 7429770665327.016,
    "cat": "ac2g", "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": 7429770665332.148, "dur": 5.091,
    "args": {
      "External id": 4986,      "device": 2, "stream": 0,      "correlation": 1667, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1667, "pid": 2, "tid": 0, "ts": 7429770665332.148,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770665337.240, "dur": 5.331,
    "args": {
      "External id": 4987,      "device": 2, "stream": 0,      "correlation": 1668, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1668, "pid": 2, "tid": 0, "ts": 7429770665337.240,
    "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::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": 7429770665342.572, "dur": 5.090,
    "args": {
      "External id": 4992,      "device": 2, "stream": 0,      "correlation": 1669, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1669, "pid": 2, "tid": 0, "ts": 7429770665342.572,
    "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": 7429770665347.663, "dur": 6.093,
    "args": {
      "External id": 4993,      "device": 2, "stream": 0,      "correlation": 1670, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1670, "pid": 2, "tid": 0, "ts": 7429770665347.663,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770665353.757, "dur": 4.650,
    "args": {
      "External id": 4994,      "device": 2, "stream": 0,      "correlation": 1671, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1671, "pid": 2, "tid": 0, "ts": 7429770665353.757,
    "cat": "ac2g", "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": 7429770665358.408, "dur": 4.730,
    "args": {
      "External id": 4995,      "device": 2, "stream": 0,      "correlation": 1672, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1672, "pid": 2, "tid": 0, "ts": 7429770665358.408,
    "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::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::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": 7429770665363.139, "dur": 5.051,
    "args": {
      "External id": 5012,      "device": 2, "stream": 0,      "correlation": 1673, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1673, "pid": 2, "tid": 0, "ts": 7429770665363.139,
    "cat": "ac2g", "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": 7429770665368.191, "dur": 5.451,
    "args": {
      "External id": 5030,      "device": 2, "stream": 0,      "correlation": 1674, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1674, "pid": 2, "tid": 0, "ts": 7429770665368.191,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770665373.643, "dur": 9.822,
    "args": {
      "External id": 5043,      "device": 2, "stream": 0,      "correlation": 1682, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1682, "pid": 2, "tid": 0, "ts": 7429770665373.643,
    "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": 7429770665383.466, "dur": 4.489,
    "args": {
      "External id": 5046,      "device": 2, "stream": 0,      "correlation": 1683, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1683, "pid": 2, "tid": 0, "ts": 7429770665383.466,
    "cat": "ac2g", "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": 7429770665387.956, "dur": 4.890,
    "args": {
      "External id": 5053,      "device": 2, "stream": 0,      "correlation": 1684, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1684, "pid": 2, "tid": 0, "ts": 7429770665387.956,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770665392.847, "dur": 5.131,
    "args": {
      "External id": 5056,      "device": 2, "stream": 0,      "correlation": 1685, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 1685, "pid": 2, "tid": 0, "ts": 7429770665392.847,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770665397.979, "dur": 10.784,
    "args": {
      "External id": 5068,      "device": 2, "stream": 0,      "correlation": 1689, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1689, "pid": 2, "tid": 0, "ts": 7429770665397.979,
    "cat": "ac2g", "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": 7429770665408.764, "dur": 5.612,
    "args": {
      "External id": 5077,      "device": 2, "stream": 0,      "correlation": 1690, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1690, "pid": 2, "tid": 0, "ts": 7429770665408.764,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770665414.377, "dur": 64.346,
    "args": {
      "External id": 5087,      "device": 2, "stream": 0,      "correlation": 1694, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1694, "pid": 2, "tid": 0, "ts": 7429770665414.377,
    "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": 7429770665478.724, "dur": 4.811,
    "args": {
      "External id": 5090,      "device": 2, "stream": 0,      "correlation": 1695, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1695, "pid": 2, "tid": 0, "ts": 7429770665478.724,
    "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": 7429770665483.536, "dur": 4.850,
    "args": {
      "External id": 5092,      "device": 2, "stream": 0,      "correlation": 1696, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1696, "pid": 2, "tid": 0, "ts": 7429770665483.536,
    "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": 7429770665488.387, "dur": 7.937,
    "args": {
      "External id": 5095,      "device": 2, "stream": 0,      "correlation": 1697, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1697, "pid": 2, "tid": 0, "ts": 7429770665488.387,
    "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": 7429770665496.325, "dur": 5.131,
    "args": {
      "External id": 5096,      "device": 2, "stream": 0,      "correlation": 1698, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1698, "pid": 2, "tid": 0, "ts": 7429770665496.325,
    "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": 7429770665501.457, "dur": 5.130,
    "args": {
      "External id": 5097,      "device": 2, "stream": 0,      "correlation": 1699, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1699, "pid": 2, "tid": 0, "ts": 7429770665501.457,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770665506.588, "dur": 4.770,
    "args": {
      "External id": 5098,      "device": 2, "stream": 0,      "correlation": 1700, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1700, "pid": 2, "tid": 0, "ts": 7429770665506.588,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770665511.359, "dur": 4.971,
    "args": {
      "External id": 5100,      "device": 2, "stream": 0,      "correlation": 1701, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1701, "pid": 2, "tid": 0, "ts": 7429770665511.359,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770665516.331, "dur": 253.060,
    "args": {
      "External id": 5108,      "device": 2, "stream": 0,      "correlation": 1705, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1705, "pid": 2, "tid": 0, "ts": 7429770665516.331,
    "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": 7429770665769.392, "dur": 5.652,
    "args": {
      "External id": 5111,      "device": 2, "stream": 0,      "correlation": 1706, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1706, "pid": 2, "tid": 0, "ts": 7429770665769.392,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770665775.045, "dur": 252.940,
    "args": {
      "External id": 5119,      "device": 2, "stream": 0,      "correlation": 1710, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1710, "pid": 2, "tid": 0, "ts": 7429770665775.045,
    "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": 7429770666027.986, "dur": 5.532,
    "args": {
      "External id": 5122,      "device": 2, "stream": 0,      "correlation": 1711, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1711, "pid": 2, "tid": 0, "ts": 7429770666027.986,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770666033.519, "dur": 165.178,
    "args": {
      "External id": 5130,      "device": 2, "stream": 0,      "correlation": 1715, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1715, "pid": 2, "tid": 0, "ts": 7429770666033.519,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770666198.698, "dur": 7.777,
    "args": {
      "External id": 5130,      "device": 2, "stream": 0,      "correlation": 1716, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1716, "pid": 2, "tid": 0, "ts": 7429770666198.698,
    "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": 7429770666206.476, "dur": 4.970,
    "args": {
      "External id": 5133,      "device": 2, "stream": 0,      "correlation": 1717, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1717, "pid": 2, "tid": 0, "ts": 7429770666206.476,
    "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": 7429770666211.447, "dur": 5.532,
    "args": {
      "External id": 5135,      "device": 2, "stream": 0,      "correlation": 1718, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1718, "pid": 2, "tid": 0, "ts": 7429770666211.447,
    "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": 7429770666216.980, "dur": 8.138,
    "args": {
      "External id": 5138,      "device": 2, "stream": 0,      "correlation": 1719, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1719, "pid": 2, "tid": 0, "ts": 7429770666216.980,
    "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": 7429770666225.119, "dur": 5.211,
    "args": {
      "External id": 5139,      "device": 2, "stream": 0,      "correlation": 1720, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1720, "pid": 2, "tid": 0, "ts": 7429770666225.119,
    "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": 7429770666230.331, "dur": 4.850,
    "args": {
      "External id": 5140,      "device": 2, "stream": 0,      "correlation": 1721, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1721, "pid": 2, "tid": 0, "ts": 7429770666230.331,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770666235.182, "dur": 4.730,
    "args": {
      "External id": 5141,      "device": 2, "stream": 0,      "correlation": 1722, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1722, "pid": 2, "tid": 0, "ts": 7429770666235.182,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770666239.913, "dur": 5.090,
    "args": {
      "External id": 5143,      "device": 2, "stream": 0,      "correlation": 1723, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1723, "pid": 2, "tid": 0, "ts": 7429770666239.913,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770666245.004, "dur": 63.706,
    "args": {
      "External id": 5151,      "device": 2, "stream": 0,      "correlation": 1727, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1727, "pid": 2, "tid": 0, "ts": 7429770666245.004,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770666308.711, "dur": 50.515,
    "args": {
      "External id": 5164,      "device": 2, "stream": 0,      "correlation": 1731, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1731, "pid": 2, "tid": 0, "ts": 7429770666308.711,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770666359.227, "dur": 50.435,
    "args": {
      "External id": 5177,      "device": 2, "stream": 0,      "correlation": 1735, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1735, "pid": 2, "tid": 0, "ts": 7429770666359.227,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770666409.663, "dur": 4.770,
    "args": {
      "External id": 5187,      "device": 2, "stream": 0,      "correlation": 1736, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1736, "pid": 2, "tid": 0, "ts": 7429770666409.663,
    "cat": "ac2g", "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": 7429770666414.434, "dur": 4.729,
    "args": {
      "External id": 5192,      "device": 2, "stream": 0,      "correlation": 1737, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1737, "pid": 2, "tid": 0, "ts": 7429770666414.434,
    "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": 7429770666419.164, "dur": 6.214,
    "args": {
      "External id": 5193,      "device": 2, "stream": 0,      "correlation": 1738, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1738, "pid": 2, "tid": 0, "ts": 7429770666419.164,
    "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": 7429770666425.379, "dur": 4.770,
    "args": {
      "External id": 5194,      "device": 2, "stream": 0,      "correlation": 1739, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1739, "pid": 2, "tid": 0, "ts": 7429770666425.379,
    "cat": "ac2g", "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": 7429770666430.150, "dur": 5.090,
    "args": {
      "External id": 5195,      "device": 2, "stream": 0,      "correlation": 1740, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1740, "pid": 2, "tid": 0, "ts": 7429770666430.150,
    "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": 7429770666435.241, "dur": 4.610,
    "args": {
      "External id": 5196,      "device": 2, "stream": 0,      "correlation": 1741, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1741, "pid": 2, "tid": 0, "ts": 7429770666435.241,
    "cat": "ac2g", "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": 7429770666439.852, "dur": 4.529,
    "args": {
      "External id": 5201,      "device": 2, "stream": 0,      "correlation": 1742, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1742, "pid": 2, "tid": 0, "ts": 7429770666439.852,
    "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": 7429770666444.382, "dur": 6.093,
    "args": {
      "External id": 5202,      "device": 2, "stream": 0,      "correlation": 1743, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1743, "pid": 2, "tid": 0, "ts": 7429770666444.382,
    "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": 7429770666450.476, "dur": 4.610,
    "args": {
      "External id": 5203,      "device": 2, "stream": 0,      "correlation": 1744, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1744, "pid": 2, "tid": 0, "ts": 7429770666450.476,
    "cat": "ac2g", "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": 7429770666455.087, "dur": 5.411,
    "args": {
      "External id": 5204,      "device": 2, "stream": 0,      "correlation": 1745, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1745, "pid": 2, "tid": 0, "ts": 7429770666455.087,
    "cat": "ac2g", "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": 7429770666460.499, "dur": 5.372,
    "args": {
      "External id": 5221,      "device": 2, "stream": 0,      "correlation": 1746, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1746, "pid": 2, "tid": 0, "ts": 7429770666460.499,
    "cat": "ac2g", "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": 7429770666465.872, "dur": 5.451,
    "args": {
      "External id": 5239,      "device": 2, "stream": 0,      "correlation": 1747, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1747, "pid": 2, "tid": 0, "ts": 7429770666465.872,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770666471.324, "dur": 9.822,
    "args": {
      "External id": 5252,      "device": 2, "stream": 0,      "correlation": 1753, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1753, "pid": 2, "tid": 0, "ts": 7429770666471.324,
    "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": 7429770666481.147, "dur": 5.291,
    "args": {
      "External id": 5255,      "device": 2, "stream": 0,      "correlation": 1754, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1754, "pid": 2, "tid": 0, "ts": 7429770666481.147,
    "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::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": 7429770666486.439, "dur": 5.652,
    "args": {
      "External id": 5262,      "device": 2, "stream": 0,      "correlation": 1755, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1755, "pid": 2, "tid": 0, "ts": 7429770666486.439,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770666492.092, "dur": 5.812,
    "args": {
      "External id": 5265,      "device": 2, "stream": 0,      "correlation": 1756, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 1756, "pid": 2, "tid": 0, "ts": 7429770666492.092,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770666497.905, "dur": 10.985,
    "args": {
      "External id": 5277,      "device": 2, "stream": 0,      "correlation": 1760, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1760, "pid": 2, "tid": 0, "ts": 7429770666497.905,
    "cat": "ac2g", "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": 7429770666508.891, "dur": 5.331,
    "args": {
      "External id": 5286,      "device": 2, "stream": 0,      "correlation": 1761, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1761, "pid": 2, "tid": 0, "ts": 7429770666508.891,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770666514.223, "dur": 63.785,
    "args": {
      "External id": 5296,      "device": 2, "stream": 0,      "correlation": 1765, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1765, "pid": 2, "tid": 0, "ts": 7429770666514.223,
    "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": 7429770666578.009, "dur": 5.412,
    "args": {
      "External id": 5299,      "device": 2, "stream": 0,      "correlation": 1766, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1766, "pid": 2, "tid": 0, "ts": 7429770666578.009,
    "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": 7429770666583.422, "dur": 5.812,
    "args": {
      "External id": 5301,      "device": 2, "stream": 0,      "correlation": 1767, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1767, "pid": 2, "tid": 0, "ts": 7429770666583.422,
    "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": 7429770666589.235, "dur": 7.897,
    "args": {
      "External id": 5304,      "device": 2, "stream": 0,      "correlation": 1768, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1768, "pid": 2, "tid": 0, "ts": 7429770666589.235,
    "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": 7429770666597.133, "dur": 5.011,
    "args": {
      "External id": 5305,      "device": 2, "stream": 0,      "correlation": 1769, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1769, "pid": 2, "tid": 0, "ts": 7429770666597.133,
    "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": 7429770666602.145, "dur": 4.930,
    "args": {
      "External id": 5306,      "device": 2, "stream": 0,      "correlation": 1770, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1770, "pid": 2, "tid": 0, "ts": 7429770666602.145,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770666607.076, "dur": 4.770,
    "args": {
      "External id": 5307,      "device": 2, "stream": 0,      "correlation": 1771, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1771, "pid": 2, "tid": 0, "ts": 7429770666607.076,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770666611.847, "dur": 4.850,
    "args": {
      "External id": 5309,      "device": 2, "stream": 0,      "correlation": 1772, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1772, "pid": 2, "tid": 0, "ts": 7429770666611.847,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770666616.698, "dur": 252.900,
    "args": {
      "External id": 5317,      "device": 2, "stream": 0,      "correlation": 1776, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1776, "pid": 2, "tid": 0, "ts": 7429770666616.698,
    "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": 7429770666869.599, "dur": 5.371,
    "args": {
      "External id": 5320,      "device": 2, "stream": 0,      "correlation": 1777, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1777, "pid": 2, "tid": 0, "ts": 7429770666869.599,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770666874.971, "dur": 252.339,
    "args": {
      "External id": 5328,      "device": 2, "stream": 0,      "correlation": 1781, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1781, "pid": 2, "tid": 0, "ts": 7429770666874.971,
    "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": 7429770667127.311, "dur": 5.652,
    "args": {
      "External id": 5331,      "device": 2, "stream": 0,      "correlation": 1782, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1782, "pid": 2, "tid": 0, "ts": 7429770667127.311,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770667132.964, "dur": 164.737,
    "args": {
      "External id": 5339,      "device": 2, "stream": 0,      "correlation": 1786, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1786, "pid": 2, "tid": 0, "ts": 7429770667132.964,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770667297.702, "dur": 7.015,
    "args": {
      "External id": 5339,      "device": 2, "stream": 0,      "correlation": 1787, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1787, "pid": 2, "tid": 0, "ts": 7429770667297.702,
    "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": 7429770667304.718, "dur": 4.690,
    "args": {
      "External id": 5342,      "device": 2, "stream": 0,      "correlation": 1788, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1788, "pid": 2, "tid": 0, "ts": 7429770667304.718,
    "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": 7429770667309.409, "dur": 5.692,
    "args": {
      "External id": 5344,      "device": 2, "stream": 0,      "correlation": 1789, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1789, "pid": 2, "tid": 0, "ts": 7429770667309.409,
    "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": 7429770667315.102, "dur": 8.098,
    "args": {
      "External id": 5347,      "device": 2, "stream": 0,      "correlation": 1790, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1790, "pid": 2, "tid": 0, "ts": 7429770667315.102,
    "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": 7429770667323.201, "dur": 4.529,
    "args": {
      "External id": 5348,      "device": 2, "stream": 0,      "correlation": 1791, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1791, "pid": 2, "tid": 0, "ts": 7429770667323.201,
    "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": 7429770667327.731, "dur": 4.650,
    "args": {
      "External id": 5349,      "device": 2, "stream": 0,      "correlation": 1792, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1792, "pid": 2, "tid": 0, "ts": 7429770667327.731,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770667332.382, "dur": 4.770,
    "args": {
      "External id": 5350,      "device": 2, "stream": 0,      "correlation": 1793, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1793, "pid": 2, "tid": 0, "ts": 7429770667332.382,
    "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": 7429770667337.153, "dur": 4.970,
    "args": {
      "External id": 5352,      "device": 2, "stream": 0,      "correlation": 1794, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1794, "pid": 2, "tid": 0, "ts": 7429770667337.153,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770667342.124, "dur": 64.107,
    "args": {
      "External id": 5360,      "device": 2, "stream": 0,      "correlation": 1798, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1798, "pid": 2, "tid": 0, "ts": 7429770667342.124,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770667406.232, "dur": 50.795,
    "args": {
      "External id": 5373,      "device": 2, "stream": 0,      "correlation": 1802, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1802, "pid": 2, "tid": 0, "ts": 7429770667406.232,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770667457.028, "dur": 50.315,
    "args": {
      "External id": 5386,      "device": 2, "stream": 0,      "correlation": 1806, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1806, "pid": 2, "tid": 0, "ts": 7429770667457.028,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770667507.344, "dur": 4.810,
    "args": {
      "External id": 5396,      "device": 2, "stream": 0,      "correlation": 1807, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1807, "pid": 2, "tid": 0, "ts": 7429770667507.344,
    "cat": "ac2g", "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": 7429770667512.155, "dur": 4.730,
    "args": {
      "External id": 5401,      "device": 2, "stream": 0,      "correlation": 1808, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1808, "pid": 2, "tid": 0, "ts": 7429770667512.155,
    "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": 7429770667516.886, "dur": 6.454,
    "args": {
      "External id": 5402,      "device": 2, "stream": 0,      "correlation": 1809, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1809, "pid": 2, "tid": 0, "ts": 7429770667516.886,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770667523.341, "dur": 5.050,
    "args": {
      "External id": 5403,      "device": 2, "stream": 0,      "correlation": 1810, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1810, "pid": 2, "tid": 0, "ts": 7429770667523.341,
    "cat": "ac2g", "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": 7429770667528.392, "dur": 5.051,
    "args": {
      "External id": 5404,      "device": 2, "stream": 0,      "correlation": 1811, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1811, "pid": 2, "tid": 0, "ts": 7429770667528.392,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770667533.444, "dur": 4.489,
    "args": {
      "External id": 5405,      "device": 2, "stream": 0,      "correlation": 1812, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1812, "pid": 2, "tid": 0, "ts": 7429770667533.444,
    "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::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": 7429770667537.934, "dur": 4.489,
    "args": {
      "External id": 5410,      "device": 2, "stream": 0,      "correlation": 1813, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1813, "pid": 2, "tid": 0, "ts": 7429770667537.934,
    "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": 7429770667542.424, "dur": 6.093,
    "args": {
      "External id": 5411,      "device": 2, "stream": 0,      "correlation": 1814, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1814, "pid": 2, "tid": 0, "ts": 7429770667542.424,
    "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": 7429770667548.518, "dur": 5.171,
    "args": {
      "External id": 5412,      "device": 2, "stream": 0,      "correlation": 1815, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1815, "pid": 2, "tid": 0, "ts": 7429770667548.518,
    "cat": "ac2g", "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": 7429770667553.690, "dur": 4.610,
    "args": {
      "External id": 5413,      "device": 2, "stream": 0,      "correlation": 1816, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1816, "pid": 2, "tid": 0, "ts": 7429770667553.690,
    "cat": "ac2g", "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": 7429770667558.301, "dur": 6.654,
    "args": {
      "External id": 5430,      "device": 2, "stream": 0,      "correlation": 1817, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1817, "pid": 2, "tid": 0, "ts": 7429770667558.301,
    "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": 7429770667564.956, "dur": 5.452,
    "args": {
      "External id": 5448,      "device": 2, "stream": 0,      "correlation": 1818, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1818, "pid": 2, "tid": 0, "ts": 7429770667564.956,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770667570.409, "dur": 9.701,
    "args": {
      "External id": 5461,      "device": 2, "stream": 0,      "correlation": 1824, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1824, "pid": 2, "tid": 0, "ts": 7429770667570.409,
    "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": 7429770667580.111, "dur": 4.529,
    "args": {
      "External id": 5464,      "device": 2, "stream": 0,      "correlation": 1825, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1825, "pid": 2, "tid": 0, "ts": 7429770667580.111,
    "cat": "ac2g", "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": 7429770667584.641, "dur": 4.850,
    "args": {
      "External id": 5471,      "device": 2, "stream": 0,      "correlation": 1826, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1826, "pid": 2, "tid": 0, "ts": 7429770667584.641,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770667589.492, "dur": 5.452,
    "args": {
      "External id": 5474,      "device": 2, "stream": 0,      "correlation": 1827, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 1827, "pid": 2, "tid": 0, "ts": 7429770667589.492,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770667594.945, "dur": 10.864,
    "args": {
      "External id": 5486,      "device": 2, "stream": 0,      "correlation": 1831, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1831, "pid": 2, "tid": 0, "ts": 7429770667594.945,
    "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": 7429770667605.810, "dur": 5.010,
    "args": {
      "External id": 5495,      "device": 2, "stream": 0,      "correlation": 1832, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1832, "pid": 2, "tid": 0, "ts": 7429770667605.810,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770667610.821, "dur": 64.508,
    "args": {
      "External id": 5505,      "device": 2, "stream": 0,      "correlation": 1836, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1836, "pid": 2, "tid": 0, "ts": 7429770667610.821,
    "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": 7429770667675.330, "dur": 4.689,
    "args": {
      "External id": 5508,      "device": 2, "stream": 0,      "correlation": 1837, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1837, "pid": 2, "tid": 0, "ts": 7429770667675.330,
    "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": 7429770667680.020, "dur": 4.770,
    "args": {
      "External id": 5510,      "device": 2, "stream": 0,      "correlation": 1838, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1838, "pid": 2, "tid": 0, "ts": 7429770667680.020,
    "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": 7429770667684.791, "dur": 7.857,
    "args": {
      "External id": 5513,      "device": 2, "stream": 0,      "correlation": 1839, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1839, "pid": 2, "tid": 0, "ts": 7429770667684.791,
    "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": 7429770667692.649, "dur": 4.570,
    "args": {
      "External id": 5514,      "device": 2, "stream": 0,      "correlation": 1840, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1840, "pid": 2, "tid": 0, "ts": 7429770667692.649,
    "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": 7429770667697.220, "dur": 5.411,
    "args": {
      "External id": 5515,      "device": 2, "stream": 0,      "correlation": 1841, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1841, "pid": 2, "tid": 0, "ts": 7429770667697.220,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770667702.632, "dur": 5.131,
    "args": {
      "External id": 5516,      "device": 2, "stream": 0,      "correlation": 1842, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1842, "pid": 2, "tid": 0, "ts": 7429770667702.632,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770667707.764, "dur": 4.971,
    "args": {
      "External id": 5518,      "device": 2, "stream": 0,      "correlation": 1843, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1843, "pid": 2, "tid": 0, "ts": 7429770667707.764,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770667712.736, "dur": 252.939,
    "args": {
      "External id": 5526,      "device": 2, "stream": 0,      "correlation": 1847, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1847, "pid": 2, "tid": 0, "ts": 7429770667712.736,
    "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": 7429770667965.676, "dur": 5.372,
    "args": {
      "External id": 5529,      "device": 2, "stream": 0,      "correlation": 1848, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1848, "pid": 2, "tid": 0, "ts": 7429770667965.676,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770667971.049, "dur": 252.860,
    "args": {
      "External id": 5537,      "device": 2, "stream": 0,      "correlation": 1854, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1854, "pid": 2, "tid": 0, "ts": 7429770667971.049,
    "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": 7429770668223.910, "dur": 5.892,
    "args": {
      "External id": 5540,      "device": 2, "stream": 0,      "correlation": 1855, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1855, "pid": 2, "tid": 0, "ts": 7429770668223.910,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770668229.803, "dur": 166.261,
    "args": {
      "External id": 5548,      "device": 2, "stream": 0,      "correlation": 1859, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1859, "pid": 2, "tid": 0, "ts": 7429770668229.803,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770668396.065, "dur": 7.376,
    "args": {
      "External id": 5548,      "device": 2, "stream": 0,      "correlation": 1860, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1860, "pid": 2, "tid": 0, "ts": 7429770668396.065,
    "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": 7429770668403.442, "dur": 4.690,
    "args": {
      "External id": 5551,      "device": 2, "stream": 0,      "correlation": 1861, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1861, "pid": 2, "tid": 0, "ts": 7429770668403.442,
    "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": 7429770668408.133, "dur": 4.810,
    "args": {
      "External id": 5553,      "device": 2, "stream": 0,      "correlation": 1862, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1862, "pid": 2, "tid": 0, "ts": 7429770668408.133,
    "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": 7429770668412.944, "dur": 8.057,
    "args": {
      "External id": 5556,      "device": 2, "stream": 0,      "correlation": 1863, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1863, "pid": 2, "tid": 0, "ts": 7429770668412.944,
    "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": 7429770668421.002, "dur": 4.449,
    "args": {
      "External id": 5557,      "device": 2, "stream": 0,      "correlation": 1864, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1864, "pid": 2, "tid": 0, "ts": 7429770668421.002,
    "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": 7429770668425.452, "dur": 4.570,
    "args": {
      "External id": 5558,      "device": 2, "stream": 0,      "correlation": 1865, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1865, "pid": 2, "tid": 0, "ts": 7429770668425.452,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770668430.023, "dur": 4.730,
    "args": {
      "External id": 5559,      "device": 2, "stream": 0,      "correlation": 1866, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1866, "pid": 2, "tid": 0, "ts": 7429770668430.023,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770668434.754, "dur": 4.970,
    "args": {
      "External id": 5561,      "device": 2, "stream": 0,      "correlation": 1867, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1867, "pid": 2, "tid": 0, "ts": 7429770668434.754,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770668439.725, "dur": 64.628,
    "args": {
      "External id": 5569,      "device": 2, "stream": 0,      "correlation": 1871, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1871, "pid": 2, "tid": 0, "ts": 7429770668439.725,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770668504.354, "dur": 50.715,
    "args": {
      "External id": 5582,      "device": 2, "stream": 0,      "correlation": 1875, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1875, "pid": 2, "tid": 0, "ts": 7429770668504.354,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770668555.070, "dur": 50.876,
    "args": {
      "External id": 5595,      "device": 2, "stream": 0,      "correlation": 1879, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1879, "pid": 2, "tid": 0, "ts": 7429770668555.070,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770668605.947, "dur": 4.690,
    "args": {
      "External id": 5605,      "device": 2, "stream": 0,      "correlation": 1880, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1880, "pid": 2, "tid": 0, "ts": 7429770668605.947,
    "cat": "ac2g", "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": 7429770668610.638, "dur": 4.770,
    "args": {
      "External id": 5610,      "device": 2, "stream": 0,      "correlation": 1881, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1881, "pid": 2, "tid": 0, "ts": 7429770668610.638,
    "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": 7429770668615.409, "dur": 6.414,
    "args": {
      "External id": 5611,      "device": 2, "stream": 0,      "correlation": 1882, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1882, "pid": 2, "tid": 0, "ts": 7429770668615.409,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770668621.824, "dur": 4.729,
    "args": {
      "External id": 5612,      "device": 2, "stream": 0,      "correlation": 1883, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1883, "pid": 2, "tid": 0, "ts": 7429770668621.824,
    "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": 7429770668626.554, "dur": 5.452,
    "args": {
      "External id": 5613,      "device": 2, "stream": 0,      "correlation": 1884, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1884, "pid": 2, "tid": 0, "ts": 7429770668626.554,
    "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": 7429770668632.007, "dur": 4.890,
    "args": {
      "External id": 5614,      "device": 2, "stream": 0,      "correlation": 1885, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1885, "pid": 2, "tid": 0, "ts": 7429770668632.007,
    "cat": "ac2g", "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": 7429770668636.898, "dur": 4.489,
    "args": {
      "External id": 5619,      "device": 2, "stream": 0,      "correlation": 1886, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1886, "pid": 2, "tid": 0, "ts": 7429770668636.898,
    "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": 7429770668641.388, "dur": 6.013,
    "args": {
      "External id": 5620,      "device": 2, "stream": 0,      "correlation": 1887, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1887, "pid": 2, "tid": 0, "ts": 7429770668641.388,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770668647.402, "dur": 5.572,
    "args": {
      "External id": 5621,      "device": 2, "stream": 0,      "correlation": 1888, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1888, "pid": 2, "tid": 0, "ts": 7429770668647.402,
    "cat": "ac2g", "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": 7429770668652.975, "dur": 5.652,
    "args": {
      "External id": 5622,      "device": 2, "stream": 0,      "correlation": 1889, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1889, "pid": 2, "tid": 0, "ts": 7429770668652.975,
    "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::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::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": 7429770668658.628, "dur": 6.334,
    "args": {
      "External id": 5639,      "device": 2, "stream": 0,      "correlation": 1890, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1890, "pid": 2, "tid": 0, "ts": 7429770668658.628,
    "cat": "ac2g", "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": 7429770668664.963, "dur": 5.451,
    "args": {
      "External id": 5657,      "device": 2, "stream": 0,      "correlation": 1891, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1891, "pid": 2, "tid": 0, "ts": 7429770668664.963,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770668670.415, "dur": 9.742,
    "args": {
      "External id": 5670,      "device": 2, "stream": 0,      "correlation": 1897, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1897, "pid": 2, "tid": 0, "ts": 7429770668670.415,
    "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": 7429770668680.158, "dur": 5.210,
    "args": {
      "External id": 5673,      "device": 2, "stream": 0,      "correlation": 1898, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1898, "pid": 2, "tid": 0, "ts": 7429770668680.158,
    "cat": "ac2g", "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": 7429770668685.369, "dur": 5.332,
    "args": {
      "External id": 5680,      "device": 2, "stream": 0,      "correlation": 1899, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1899, "pid": 2, "tid": 0, "ts": 7429770668685.369,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770668690.702, "dur": 5.451,
    "args": {
      "External id": 5683,      "device": 2, "stream": 0,      "correlation": 1900, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 1900, "pid": 2, "tid": 0, "ts": 7429770668690.702,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770668696.154, "dur": 10.944,
    "args": {
      "External id": 5695,      "device": 2, "stream": 0,      "correlation": 1904, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1904, "pid": 2, "tid": 0, "ts": 7429770668696.154,
    "cat": "ac2g", "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": 7429770668707.099, "dur": 5.011,
    "args": {
      "External id": 5704,      "device": 2, "stream": 0,      "correlation": 1905, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1905, "pid": 2, "tid": 0, "ts": 7429770668707.099,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770668712.111, "dur": 64.026,
    "args": {
      "External id": 5714,      "device": 2, "stream": 0,      "correlation": 1909, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1909, "pid": 2, "tid": 0, "ts": 7429770668712.111,
    "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": 7429770668776.138, "dur": 5.331,
    "args": {
      "External id": 5717,      "device": 2, "stream": 0,      "correlation": 1910, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1910, "pid": 2, "tid": 0, "ts": 7429770668776.138,
    "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": 7429770668781.470, "dur": 5.412,
    "args": {
      "External id": 5719,      "device": 2, "stream": 0,      "correlation": 1911, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1911, "pid": 2, "tid": 0, "ts": 7429770668781.470,
    "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": 7429770668786.883, "dur": 7.736,
    "args": {
      "External id": 5722,      "device": 2, "stream": 0,      "correlation": 1912, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1912, "pid": 2, "tid": 0, "ts": 7429770668786.883,
    "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": 7429770668794.620, "dur": 4.490,
    "args": {
      "External id": 5723,      "device": 2, "stream": 0,      "correlation": 1913, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1913, "pid": 2, "tid": 0, "ts": 7429770668794.620,
    "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": 7429770668799.111, "dur": 4.649,
    "args": {
      "External id": 5724,      "device": 2, "stream": 0,      "correlation": 1914, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1914, "pid": 2, "tid": 0, "ts": 7429770668799.111,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770668803.761, "dur": 4.810,
    "args": {
      "External id": 5725,      "device": 2, "stream": 0,      "correlation": 1915, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1915, "pid": 2, "tid": 0, "ts": 7429770668803.761,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770668808.572, "dur": 5.091,
    "args": {
      "External id": 5727,      "device": 2, "stream": 0,      "correlation": 1916, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1916, "pid": 2, "tid": 0, "ts": 7429770668808.572,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770668813.664, "dur": 253.181,
    "args": {
      "External id": 5735,      "device": 2, "stream": 0,      "correlation": 1920, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1920, "pid": 2, "tid": 0, "ts": 7429770668813.664,
    "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": 7429770669066.846, "dur": 5.531,
    "args": {
      "External id": 5738,      "device": 2, "stream": 0,      "correlation": 1921, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1921, "pid": 2, "tid": 0, "ts": 7429770669066.846,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770669072.378, "dur": 252.379,
    "args": {
      "External id": 5746,      "device": 2, "stream": 0,      "correlation": 1925, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1925, "pid": 2, "tid": 0, "ts": 7429770669072.378,
    "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": 7429770669324.758, "dur": 5.852,
    "args": {
      "External id": 5749,      "device": 2, "stream": 0,      "correlation": 1926, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1926, "pid": 2, "tid": 0, "ts": 7429770669324.758,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770669330.611, "dur": 165.259,
    "args": {
      "External id": 5757,      "device": 2, "stream": 0,      "correlation": 1930, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1930, "pid": 2, "tid": 0, "ts": 7429770669330.611,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770669495.871, "dur": 7.697,
    "args": {
      "External id": 5757,      "device": 2, "stream": 0,      "correlation": 1931, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1931, "pid": 2, "tid": 0, "ts": 7429770669495.871,
    "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": 7429770669503.569, "dur": 5.090,
    "args": {
      "External id": 5760,      "device": 2, "stream": 0,      "correlation": 1932, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1932, "pid": 2, "tid": 0, "ts": 7429770669503.569,
    "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": 7429770669508.660, "dur": 4.770,
    "args": {
      "External id": 5762,      "device": 2, "stream": 0,      "correlation": 1933, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1933, "pid": 2, "tid": 0, "ts": 7429770669508.660,
    "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": 7429770669513.431, "dur": 8.178,
    "args": {
      "External id": 5765,      "device": 2, "stream": 0,      "correlation": 1934, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1934, "pid": 2, "tid": 0, "ts": 7429770669513.431,
    "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": 7429770669521.610, "dur": 5.051,
    "args": {
      "External id": 5766,      "device": 2, "stream": 0,      "correlation": 1935, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1935, "pid": 2, "tid": 0, "ts": 7429770669521.610,
    "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": 7429770669526.662, "dur": 5.251,
    "args": {
      "External id": 5767,      "device": 2, "stream": 0,      "correlation": 1936, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1936, "pid": 2, "tid": 0, "ts": 7429770669526.662,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770669531.914, "dur": 4.850,
    "args": {
      "External id": 5768,      "device": 2, "stream": 0,      "correlation": 1937, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1937, "pid": 2, "tid": 0, "ts": 7429770669531.914,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770669536.765, "dur": 5.091,
    "args": {
      "External id": 5770,      "device": 2, "stream": 0,      "correlation": 1938, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1938, "pid": 2, "tid": 0, "ts": 7429770669536.765,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770669541.857, "dur": 63.905,
    "args": {
      "External id": 5778,      "device": 2, "stream": 0,      "correlation": 1942, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1942, "pid": 2, "tid": 0, "ts": 7429770669541.857,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770669605.763, "dur": 50.315,
    "args": {
      "External id": 5791,      "device": 2, "stream": 0,      "correlation": 1946, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1946, "pid": 2, "tid": 0, "ts": 7429770669605.763,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770669656.079, "dur": 49.793,
    "args": {
      "External id": 5804,      "device": 2, "stream": 0,      "correlation": 1950, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1950, "pid": 2, "tid": 0, "ts": 7429770669656.079,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770669705.873, "dur": 5.131,
    "args": {
      "External id": 5814,      "device": 2, "stream": 0,      "correlation": 1951, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1951, "pid": 2, "tid": 0, "ts": 7429770669705.873,
    "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::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": 7429770669711.005, "dur": 5.492,
    "args": {
      "External id": 5819,      "device": 2, "stream": 0,      "correlation": 1952, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1952, "pid": 2, "tid": 0, "ts": 7429770669711.005,
    "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": 7429770669716.498, "dur": 6.213,
    "args": {
      "External id": 5820,      "device": 2, "stream": 0,      "correlation": 1953, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1953, "pid": 2, "tid": 0, "ts": 7429770669716.498,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770669722.712, "dur": 4.770,
    "args": {
      "External id": 5821,      "device": 2, "stream": 0,      "correlation": 1954, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1954, "pid": 2, "tid": 0, "ts": 7429770669722.712,
    "cat": "ac2g", "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": 7429770669727.483, "dur": 5.010,
    "args": {
      "External id": 5822,      "device": 2, "stream": 0,      "correlation": 1955, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1955, "pid": 2, "tid": 0, "ts": 7429770669727.483,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770669732.494, "dur": 5.412,
    "args": {
      "External id": 5823,      "device": 2, "stream": 0,      "correlation": 1956, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1956, "pid": 2, "tid": 0, "ts": 7429770669732.494,
    "cat": "ac2g", "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": 7429770669737.907, "dur": 5.051,
    "args": {
      "External id": 5828,      "device": 2, "stream": 0,      "correlation": 1957, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1957, "pid": 2, "tid": 0, "ts": 7429770669737.907,
    "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": 7429770669742.959, "dur": 6.052,
    "args": {
      "External id": 5829,      "device": 2, "stream": 0,      "correlation": 1958, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1958, "pid": 2, "tid": 0, "ts": 7429770669742.959,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770669749.012, "dur": 4.730,
    "args": {
      "External id": 5830,      "device": 2, "stream": 0,      "correlation": 1959, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1959, "pid": 2, "tid": 0, "ts": 7429770669749.012,
    "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::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": 7429770669753.743, "dur": 4.810,
    "args": {
      "External id": 5831,      "device": 2, "stream": 0,      "correlation": 1960, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1960, "pid": 2, "tid": 0, "ts": 7429770669753.743,
    "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::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::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": 7429770669758.554, "dur": 6.093,
    "args": {
      "External id": 5848,      "device": 2, "stream": 0,      "correlation": 1961, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1961, "pid": 2, "tid": 0, "ts": 7429770669758.554,
    "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::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::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": 7429770669764.648, "dur": 5.452,
    "args": {
      "External id": 5866,      "device": 2, "stream": 0,      "correlation": 1962, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1962, "pid": 2, "tid": 0, "ts": 7429770669764.648,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770669770.101, "dur": 9.260,
    "args": {
      "External id": 5879,      "device": 2, "stream": 0,      "correlation": 1968, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1968, "pid": 2, "tid": 0, "ts": 7429770669770.101,
    "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": 7429770669779.362, "dur": 5.091,
    "args": {
      "External id": 5882,      "device": 2, "stream": 0,      "correlation": 1969, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1969, "pid": 2, "tid": 0, "ts": 7429770669779.362,
    "cat": "ac2g", "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": 7429770669784.454, "dur": 5.371,
    "args": {
      "External id": 5889,      "device": 2, "stream": 0,      "correlation": 1970, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1970, "pid": 2, "tid": 0, "ts": 7429770669784.454,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770669789.826, "dur": 5.171,
    "args": {
      "External id": 5892,      "device": 2, "stream": 0,      "correlation": 1971, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 1971, "pid": 2, "tid": 0, "ts": 7429770669789.826,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770669794.998, "dur": 10.904,
    "args": {
      "External id": 5904,      "device": 2, "stream": 0,      "correlation": 1975, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1975, "pid": 2, "tid": 0, "ts": 7429770669794.998,
    "cat": "ac2g", "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": 7429770669805.903, "dur": 5.331,
    "args": {
      "External id": 5913,      "device": 2, "stream": 0,      "correlation": 1976, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1976, "pid": 2, "tid": 0, "ts": 7429770669805.903,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770669811.235, "dur": 64.227,
    "args": {
      "External id": 5923,      "device": 2, "stream": 0,      "correlation": 1980, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1980, "pid": 2, "tid": 0, "ts": 7429770669811.235,
    "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": 7429770669875.463, "dur": 4.650,
    "args": {
      "External id": 5926,      "device": 2, "stream": 0,      "correlation": 1981, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1981, "pid": 2, "tid": 0, "ts": 7429770669875.463,
    "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": 7429770669880.114, "dur": 5.652,
    "args": {
      "External id": 5928,      "device": 2, "stream": 0,      "correlation": 1982, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1982, "pid": 2, "tid": 0, "ts": 7429770669880.114,
    "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": 7429770669885.767, "dur": 8.057,
    "args": {
      "External id": 5931,      "device": 2, "stream": 0,      "correlation": 1983, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 1983, "pid": 2, "tid": 0, "ts": 7429770669885.767,
    "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": 7429770669893.825, "dur": 4.369,
    "args": {
      "External id": 5932,      "device": 2, "stream": 0,      "correlation": 1984, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1984, "pid": 2, "tid": 0, "ts": 7429770669893.825,
    "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": 7429770669898.195, "dur": 4.529,
    "args": {
      "External id": 5933,      "device": 2, "stream": 0,      "correlation": 1985, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1985, "pid": 2, "tid": 0, "ts": 7429770669898.195,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770669902.725, "dur": 4.770,
    "args": {
      "External id": 5934,      "device": 2, "stream": 0,      "correlation": 1986, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1986, "pid": 2, "tid": 0, "ts": 7429770669902.725,
    "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": 7429770669907.496, "dur": 4.931,
    "args": {
      "External id": 5936,      "device": 2, "stream": 0,      "correlation": 1987, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1987, "pid": 2, "tid": 0, "ts": 7429770669907.496,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770669912.428, "dur": 252.659,
    "args": {
      "External id": 5944,      "device": 2, "stream": 0,      "correlation": 1991, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1991, "pid": 2, "tid": 0, "ts": 7429770669912.428,
    "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": 7429770670165.088, "dur": 5.692,
    "args": {
      "External id": 5947,      "device": 2, "stream": 0,      "correlation": 1992, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1992, "pid": 2, "tid": 0, "ts": 7429770670165.088,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770670170.781, "dur": 252.700,
    "args": {
      "External id": 5955,      "device": 2, "stream": 0,      "correlation": 1996, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1996, "pid": 2, "tid": 0, "ts": 7429770670170.781,
    "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": 7429770670423.482, "dur": 5.692,
    "args": {
      "External id": 5958,      "device": 2, "stream": 0,      "correlation": 1997, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 1997, "pid": 2, "tid": 0, "ts": 7429770670423.482,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770670429.175, "dur": 165.579,
    "args": {
      "External id": 5966,      "device": 2, "stream": 0,      "correlation": 2003, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2003, "pid": 2, "tid": 0, "ts": 7429770670429.175,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770670594.755, "dur": 7.095,
    "args": {
      "External id": 5966,      "device": 2, "stream": 0,      "correlation": 2004, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2004, "pid": 2, "tid": 0, "ts": 7429770670594.755,
    "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": 7429770670601.851, "dur": 5.251,
    "args": {
      "External id": 5969,      "device": 2, "stream": 0,      "correlation": 2005, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2005, "pid": 2, "tid": 0, "ts": 7429770670601.851,
    "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": 7429770670607.103, "dur": 6.093,
    "args": {
      "External id": 5971,      "device": 2, "stream": 0,      "correlation": 2006, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2006, "pid": 2, "tid": 0, "ts": 7429770670607.103,
    "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": 7429770670613.197, "dur": 8.098,
    "args": {
      "External id": 5974,      "device": 2, "stream": 0,      "correlation": 2007, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 2007, "pid": 2, "tid": 0, "ts": 7429770670613.197,
    "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": 7429770670621.296, "dur": 4.489,
    "args": {
      "External id": 5975,      "device": 2, "stream": 0,      "correlation": 2008, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2008, "pid": 2, "tid": 0, "ts": 7429770670621.296,
    "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": 7429770670625.786, "dur": 5.051,
    "args": {
      "External id": 5976,      "device": 2, "stream": 0,      "correlation": 2009, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2009, "pid": 2, "tid": 0, "ts": 7429770670625.786,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770670630.838, "dur": 5.371,
    "args": {
      "External id": 5977,      "device": 2, "stream": 0,      "correlation": 2010, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2010, "pid": 2, "tid": 0, "ts": 7429770670630.838,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770670636.210, "dur": 5.051,
    "args": {
      "External id": 5979,      "device": 2, "stream": 0,      "correlation": 2011, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2011, "pid": 2, "tid": 0, "ts": 7429770670636.210,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770670641.262, "dur": 64.427,
    "args": {
      "External id": 5987,      "device": 2, "stream": 0,      "correlation": 2015, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2015, "pid": 2, "tid": 0, "ts": 7429770670641.262,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770670705.690, "dur": 50.435,
    "args": {
      "External id": 6000,      "device": 2, "stream": 0,      "correlation": 2019, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2019, "pid": 2, "tid": 0, "ts": 7429770670705.690,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770670756.126, "dur": 49.873,
    "args": {
      "External id": 6013,      "device": 2, "stream": 0,      "correlation": 2023, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2023, "pid": 2, "tid": 0, "ts": 7429770670756.126,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770670806.000, "dur": 4.770,
    "args": {
      "External id": 6023,      "device": 2, "stream": 0,      "correlation": 2024, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2024, "pid": 2, "tid": 0, "ts": 7429770670806.000,
    "cat": "ac2g", "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": 7429770670810.771, "dur": 4.730,
    "args": {
      "External id": 6028,      "device": 2, "stream": 0,      "correlation": 2025, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2025, "pid": 2, "tid": 0, "ts": 7429770670810.771,
    "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": 7429770670815.502, "dur": 6.213,
    "args": {
      "External id": 6029,      "device": 2, "stream": 0,      "correlation": 2026, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2026, "pid": 2, "tid": 0, "ts": 7429770670815.502,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770670821.716, "dur": 4.810,
    "args": {
      "External id": 6030,      "device": 2, "stream": 0,      "correlation": 2027, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2027, "pid": 2, "tid": 0, "ts": 7429770670821.716,
    "cat": "ac2g", "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": 7429770670826.527, "dur": 4.971,
    "args": {
      "External id": 6031,      "device": 2, "stream": 0,      "correlation": 2028, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2028, "pid": 2, "tid": 0, "ts": 7429770670826.527,
    "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": 7429770670831.499, "dur": 4.890,
    "args": {
      "External id": 6032,      "device": 2, "stream": 0,      "correlation": 2029, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2029, "pid": 2, "tid": 0, "ts": 7429770670831.499,
    "cat": "ac2g", "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": 7429770670836.390, "dur": 4.730,
    "args": {
      "External id": 6037,      "device": 2, "stream": 0,      "correlation": 2030, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2030, "pid": 2, "tid": 0, "ts": 7429770670836.390,
    "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": 7429770670841.121, "dur": 6.213,
    "args": {
      "External id": 6038,      "device": 2, "stream": 0,      "correlation": 2031, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2031, "pid": 2, "tid": 0, "ts": 7429770670841.121,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770670847.335, "dur": 5.091,
    "args": {
      "External id": 6039,      "device": 2, "stream": 0,      "correlation": 2032, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2032, "pid": 2, "tid": 0, "ts": 7429770670847.335,
    "cat": "ac2g", "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": 7429770670852.427, "dur": 5.652,
    "args": {
      "External id": 6040,      "device": 2, "stream": 0,      "correlation": 2033, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2033, "pid": 2, "tid": 0, "ts": 7429770670852.427,
    "cat": "ac2g", "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": 7429770670858.080, "dur": 6.333,
    "args": {
      "External id": 6057,      "device": 2, "stream": 0,      "correlation": 2034, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2034, "pid": 2, "tid": 0, "ts": 7429770670858.080,
    "cat": "ac2g", "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": 7429770670864.414, "dur": 5.813,
    "args": {
      "External id": 6075,      "device": 2, "stream": 0,      "correlation": 2035, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2035, "pid": 2, "tid": 0, "ts": 7429770670864.414,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770670870.228, "dur": 9.621,
    "args": {
      "External id": 6088,      "device": 2, "stream": 0,      "correlation": 2041, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2041, "pid": 2, "tid": 0, "ts": 7429770670870.228,
    "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": 7429770670879.850, "dur": 5.291,
    "args": {
      "External id": 6091,      "device": 2, "stream": 0,      "correlation": 2042, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2042, "pid": 2, "tid": 0, "ts": 7429770670879.850,
    "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::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": 7429770670885.142, "dur": 5.612,
    "args": {
      "External id": 6098,      "device": 2, "stream": 0,      "correlation": 2043, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2043, "pid": 2, "tid": 0, "ts": 7429770670885.142,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770670890.755, "dur": 5.491,
    "args": {
      "External id": 6101,      "device": 2, "stream": 0,      "correlation": 2044, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 2044, "pid": 2, "tid": 0, "ts": 7429770670890.755,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770670896.247, "dur": 10.824,
    "args": {
      "External id": 6113,      "device": 2, "stream": 0,      "correlation": 2048, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2048, "pid": 2, "tid": 0, "ts": 7429770670896.247,
    "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::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::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": 7429770670907.072, "dur": 4.971,
    "args": {
      "External id": 6122,      "device": 2, "stream": 0,      "correlation": 2049, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2049, "pid": 2, "tid": 0, "ts": 7429770670907.072,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770670912.044, "dur": 63.705,
    "args": {
      "External id": 6132,      "device": 2, "stream": 0,      "correlation": 2053, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2053, "pid": 2, "tid": 0, "ts": 7429770670912.044,
    "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": 7429770670975.750, "dur": 4.770,
    "args": {
      "External id": 6135,      "device": 2, "stream": 0,      "correlation": 2054, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2054, "pid": 2, "tid": 0, "ts": 7429770670975.750,
    "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": 7429770670980.521, "dur": 4.850,
    "args": {
      "External id": 6137,      "device": 2, "stream": 0,      "correlation": 2055, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2055, "pid": 2, "tid": 0, "ts": 7429770670980.521,
    "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": 7429770670985.372, "dur": 8.138,
    "args": {
      "External id": 6140,      "device": 2, "stream": 0,      "correlation": 2056, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 2056, "pid": 2, "tid": 0, "ts": 7429770670985.372,
    "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": 7429770670993.511, "dur": 4.569,
    "args": {
      "External id": 6141,      "device": 2, "stream": 0,      "correlation": 2057, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2057, "pid": 2, "tid": 0, "ts": 7429770670993.511,
    "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": 7429770670998.081, "dur": 4.810,
    "args": {
      "External id": 6142,      "device": 2, "stream": 0,      "correlation": 2058, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2058, "pid": 2, "tid": 0, "ts": 7429770670998.081,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770671002.892, "dur": 4.850,
    "args": {
      "External id": 6143,      "device": 2, "stream": 0,      "correlation": 2059, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2059, "pid": 2, "tid": 0, "ts": 7429770671002.892,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770671007.743, "dur": 4.891,
    "args": {
      "External id": 6145,      "device": 2, "stream": 0,      "correlation": 2060, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2060, "pid": 2, "tid": 0, "ts": 7429770671007.743,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770671012.635, "dur": 253.020,
    "args": {
      "External id": 6153,      "device": 2, "stream": 0,      "correlation": 2064, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2064, "pid": 2, "tid": 0, "ts": 7429770671012.635,
    "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": 7429770671265.656, "dur": 5.451,
    "args": {
      "External id": 6156,      "device": 2, "stream": 0,      "correlation": 2065, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2065, "pid": 2, "tid": 0, "ts": 7429770671265.656,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770671271.108, "dur": 252.860,
    "args": {
      "External id": 6164,      "device": 2, "stream": 0,      "correlation": 2069, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2069, "pid": 2, "tid": 0, "ts": 7429770671271.108,
    "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": 7429770671523.969, "dur": 5.812,
    "args": {
      "External id": 6167,      "device": 2, "stream": 0,      "correlation": 2070, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2070, "pid": 2, "tid": 0, "ts": 7429770671523.969,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770671529.782, "dur": 165.019,
    "args": {
      "External id": 6175,      "device": 2, "stream": 0,      "correlation": 2074, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2074, "pid": 2, "tid": 0, "ts": 7429770671529.782,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770671694.802, "dur": 7.416,
    "args": {
      "External id": 6175,      "device": 2, "stream": 0,      "correlation": 2075, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2075, "pid": 2, "tid": 0, "ts": 7429770671694.802,
    "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": 7429770671702.219, "dur": 4.689,
    "args": {
      "External id": 6178,      "device": 2, "stream": 0,      "correlation": 2076, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2076, "pid": 2, "tid": 0, "ts": 7429770671702.219,
    "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": 7429770671706.909, "dur": 5.251,
    "args": {
      "External id": 6180,      "device": 2, "stream": 0,      "correlation": 2077, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2077, "pid": 2, "tid": 0, "ts": 7429770671706.909,
    "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": 7429770671712.161, "dur": 8.218,
    "args": {
      "External id": 6183,      "device": 2, "stream": 0,      "correlation": 2078, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 2078, "pid": 2, "tid": 0, "ts": 7429770671712.161,
    "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": 7429770671720.380, "dur": 5.011,
    "args": {
      "External id": 6184,      "device": 2, "stream": 0,      "correlation": 2079, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2079, "pid": 2, "tid": 0, "ts": 7429770671720.380,
    "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": 7429770671725.392, "dur": 4.930,
    "args": {
      "External id": 6185,      "device": 2, "stream": 0,      "correlation": 2080, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2080, "pid": 2, "tid": 0, "ts": 7429770671725.392,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770671730.323, "dur": 4.810,
    "args": {
      "External id": 6186,      "device": 2, "stream": 0,      "correlation": 2081, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2081, "pid": 2, "tid": 0, "ts": 7429770671730.323,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770671735.134, "dur": 5.331,
    "args": {
      "External id": 6188,      "device": 2, "stream": 0,      "correlation": 2082, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2082, "pid": 2, "tid": 0, "ts": 7429770671735.134,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770671740.466, "dur": 64.187,
    "args": {
      "External id": 6196,      "device": 2, "stream": 0,      "correlation": 2086, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2086, "pid": 2, "tid": 0, "ts": 7429770671740.466,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770671804.654, "dur": 50.475,
    "args": {
      "External id": 6209,      "device": 2, "stream": 0,      "correlation": 2090, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2090, "pid": 2, "tid": 0, "ts": 7429770671804.654,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770671855.130, "dur": 50.154,
    "args": {
      "External id": 6222,      "device": 2, "stream": 0,      "correlation": 2094, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2094, "pid": 2, "tid": 0, "ts": 7429770671855.130,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770671905.285, "dur": 5.291,
    "args": {
      "External id": 6232,      "device": 2, "stream": 0,      "correlation": 2095, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2095, "pid": 2, "tid": 0, "ts": 7429770671905.285,
    "cat": "ac2g", "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": 7429770671910.577, "dur": 5.291,
    "args": {
      "External id": 6237,      "device": 2, "stream": 0,      "correlation": 2096, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2096, "pid": 2, "tid": 0, "ts": 7429770671910.577,
    "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": 7429770671915.869, "dur": 6.254,
    "args": {
      "External id": 6238,      "device": 2, "stream": 0,      "correlation": 2097, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2097, "pid": 2, "tid": 0, "ts": 7429770671915.869,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770671922.124, "dur": 5.050,
    "args": {
      "External id": 6239,      "device": 2, "stream": 0,      "correlation": 2098, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2098, "pid": 2, "tid": 0, "ts": 7429770671922.124,
    "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::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": 7429770671927.175, "dur": 4.971,
    "args": {
      "External id": 6240,      "device": 2, "stream": 0,      "correlation": 2099, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2099, "pid": 2, "tid": 0, "ts": 7429770671927.175,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770671932.147, "dur": 5.211,
    "args": {
      "External id": 6241,      "device": 2, "stream": 0,      "correlation": 2100, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2100, "pid": 2, "tid": 0, "ts": 7429770671932.147,
    "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": 7429770671937.359, "dur": 4.970,
    "args": {
      "External id": 6246,      "device": 2, "stream": 0,      "correlation": 2101, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2101, "pid": 2, "tid": 0, "ts": 7429770671937.359,
    "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": 7429770671942.330, "dur": 6.213,
    "args": {
      "External id": 6247,      "device": 2, "stream": 0,      "correlation": 2102, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2102, "pid": 2, "tid": 0, "ts": 7429770671942.330,
    "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": 7429770671948.544, "dur": 5.051,
    "args": {
      "External id": 6248,      "device": 2, "stream": 0,      "correlation": 2103, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2103, "pid": 2, "tid": 0, "ts": 7429770671948.544,
    "cat": "ac2g", "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": 7429770671953.596, "dur": 4.850,
    "args": {
      "External id": 6249,      "device": 2, "stream": 0,      "correlation": 2104, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2104, "pid": 2, "tid": 0, "ts": 7429770671953.596,
    "cat": "ac2g", "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": 7429770671958.447, "dur": 6.294,
    "args": {
      "External id": 6266,      "device": 2, "stream": 0,      "correlation": 2105, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2105, "pid": 2, "tid": 0, "ts": 7429770671958.447,
    "cat": "ac2g", "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": 7429770671964.742, "dur": 5.411,
    "args": {
      "External id": 6284,      "device": 2, "stream": 0,      "correlation": 2106, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2106, "pid": 2, "tid": 0, "ts": 7429770671964.742,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770671970.154, "dur": 9.621,
    "args": {
      "External id": 6297,      "device": 2, "stream": 0,      "correlation": 2112, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2112, "pid": 2, "tid": 0, "ts": 7429770671970.154,
    "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": 7429770671979.776, "dur": 4.650,
    "args": {
      "External id": 6300,      "device": 2, "stream": 0,      "correlation": 2113, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2113, "pid": 2, "tid": 0, "ts": 7429770671979.776,
    "cat": "ac2g", "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": 7429770671984.427, "dur": 4.930,
    "args": {
      "External id": 6307,      "device": 2, "stream": 0,      "correlation": 2114, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2114, "pid": 2, "tid": 0, "ts": 7429770671984.427,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770671989.358, "dur": 5.131,
    "args": {
      "External id": 6310,      "device": 2, "stream": 0,      "correlation": 2115, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 2115, "pid": 2, "tid": 0, "ts": 7429770671989.358,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770671994.490, "dur": 10.864,
    "args": {
      "External id": 6322,      "device": 2, "stream": 0,      "correlation": 2119, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2119, "pid": 2, "tid": 0, "ts": 7429770671994.490,
    "cat": "ac2g", "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": 7429770672005.355, "dur": 5.331,
    "args": {
      "External id": 6331,      "device": 2, "stream": 0,      "correlation": 2120, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2120, "pid": 2, "tid": 0, "ts": 7429770672005.355,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770672010.687, "dur": 64.267,
    "args": {
      "External id": 6341,      "device": 2, "stream": 0,      "correlation": 2124, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2124, "pid": 2, "tid": 0, "ts": 7429770672010.687,
    "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": 7429770672074.955, "dur": 4.810,
    "args": {
      "External id": 6344,      "device": 2, "stream": 0,      "correlation": 2125, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2125, "pid": 2, "tid": 0, "ts": 7429770672074.955,
    "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": 7429770672079.766, "dur": 4.810,
    "args": {
      "External id": 6346,      "device": 2, "stream": 0,      "correlation": 2126, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2126, "pid": 2, "tid": 0, "ts": 7429770672079.766,
    "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": 7429770672084.577, "dur": 7.857,
    "args": {
      "External id": 6349,      "device": 2, "stream": 0,      "correlation": 2127, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 2127, "pid": 2, "tid": 0, "ts": 7429770672084.577,
    "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": 7429770672092.435, "dur": 5.050,
    "args": {
      "External id": 6350,      "device": 2, "stream": 0,      "correlation": 2128, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2128, "pid": 2, "tid": 0, "ts": 7429770672092.435,
    "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": 7429770672097.486, "dur": 5.372,
    "args": {
      "External id": 6351,      "device": 2, "stream": 0,      "correlation": 2129, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2129, "pid": 2, "tid": 0, "ts": 7429770672097.486,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770672102.859, "dur": 5.131,
    "args": {
      "External id": 6352,      "device": 2, "stream": 0,      "correlation": 2130, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2130, "pid": 2, "tid": 0, "ts": 7429770672102.859,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770672107.991, "dur": 5.571,
    "args": {
      "External id": 6354,      "device": 2, "stream": 0,      "correlation": 2131, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2131, "pid": 2, "tid": 0, "ts": 7429770672107.991,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770672113.563, "dur": 253.542,
    "args": {
      "External id": 6362,      "device": 2, "stream": 0,      "correlation": 2135, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2135, "pid": 2, "tid": 0, "ts": 7429770672113.563,
    "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": 7429770672367.106, "dur": 5.451,
    "args": {
      "External id": 6365,      "device": 2, "stream": 0,      "correlation": 2136, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2136, "pid": 2, "tid": 0, "ts": 7429770672367.106,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770672372.558, "dur": 253.983,
    "args": {
      "External id": 6373,      "device": 2, "stream": 0,      "correlation": 2140, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2140, "pid": 2, "tid": 0, "ts": 7429770672372.558,
    "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": 7429770672626.542, "dur": 5.812,
    "args": {
      "External id": 6376,      "device": 2, "stream": 0,      "correlation": 2141, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2141, "pid": 2, "tid": 0, "ts": 7429770672626.542,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770672632.355, "dur": 164.537,
    "args": {
      "External id": 6384,      "device": 2, "stream": 0,      "correlation": 2145, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2145, "pid": 2, "tid": 0, "ts": 7429770672632.355,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770672796.893, "dur": 7.817,
    "args": {
      "External id": 6384,      "device": 2, "stream": 0,      "correlation": 2146, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2146, "pid": 2, "tid": 0, "ts": 7429770672796.893,
    "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": 7429770672804.711, "dur": 5.371,
    "args": {
      "External id": 6387,      "device": 2, "stream": 0,      "correlation": 2147, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2147, "pid": 2, "tid": 0, "ts": 7429770672804.711,
    "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": 7429770672810.083, "dur": 5.452,
    "args": {
      "External id": 6389,      "device": 2, "stream": 0,      "correlation": 2148, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2148, "pid": 2, "tid": 0, "ts": 7429770672810.083,
    "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": 7429770672815.536, "dur": 8.057,
    "args": {
      "External id": 6392,      "device": 2, "stream": 0,      "correlation": 2149, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 2149, "pid": 2, "tid": 0, "ts": 7429770672815.536,
    "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": 7429770672823.594, "dur": 5.011,
    "args": {
      "External id": 6393,      "device": 2, "stream": 0,      "correlation": 2150, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2150, "pid": 2, "tid": 0, "ts": 7429770672823.594,
    "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": 7429770672828.606, "dur": 4.890,
    "args": {
      "External id": 6394,      "device": 2, "stream": 0,      "correlation": 2151, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2151, "pid": 2, "tid": 0, "ts": 7429770672828.606,
    "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": 7429770672833.497, "dur": 5.331,
    "args": {
      "External id": 6395,      "device": 2, "stream": 0,      "correlation": 2152, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2152, "pid": 2, "tid": 0, "ts": 7429770672833.497,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770672838.829, "dur": 5.091,
    "args": {
      "External id": 6397,      "device": 2, "stream": 0,      "correlation": 2153, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2153, "pid": 2, "tid": 0, "ts": 7429770672838.829,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770672843.921, "dur": 63.545,
    "args": {
      "External id": 6405,      "device": 2, "stream": 0,      "correlation": 2157, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2157, "pid": 2, "tid": 0, "ts": 7429770672843.921,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770672907.467, "dur": 50.194,
    "args": {
      "External id": 6418,      "device": 2, "stream": 0,      "correlation": 2161, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2161, "pid": 2, "tid": 0, "ts": 7429770672907.467,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770672957.662, "dur": 49.874,
    "args": {
      "External id": 6431,      "device": 2, "stream": 0,      "correlation": 2165, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2165, "pid": 2, "tid": 0, "ts": 7429770672957.662,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770673007.537, "dur": 4.729,
    "args": {
      "External id": 6441,      "device": 2, "stream": 0,      "correlation": 2166, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2166, "pid": 2, "tid": 0, "ts": 7429770673007.537,
    "cat": "ac2g", "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": 7429770673012.267, "dur": 4.770,
    "args": {
      "External id": 6446,      "device": 2, "stream": 0,      "correlation": 2167, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2167, "pid": 2, "tid": 0, "ts": 7429770673012.267,
    "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": 7429770673017.038, "dur": 6.575,
    "args": {
      "External id": 6447,      "device": 2, "stream": 0,      "correlation": 2168, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2168, "pid": 2, "tid": 0, "ts": 7429770673017.038,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770673023.614, "dur": 4.769,
    "args": {
      "External id": 6448,      "device": 2, "stream": 0,      "correlation": 2169, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2169, "pid": 2, "tid": 0, "ts": 7429770673023.614,
    "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::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": 7429770673028.384, "dur": 5.011,
    "args": {
      "External id": 6449,      "device": 2, "stream": 0,      "correlation": 2170, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2170, "pid": 2, "tid": 0, "ts": 7429770673028.384,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770673033.396, "dur": 4.529,
    "args": {
      "External id": 6450,      "device": 2, "stream": 0,      "correlation": 2171, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2171, "pid": 2, "tid": 0, "ts": 7429770673033.396,
    "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::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": 7429770673037.926, "dur": 4.530,
    "args": {
      "External id": 6455,      "device": 2, "stream": 0,      "correlation": 2172, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2172, "pid": 2, "tid": 0, "ts": 7429770673037.926,
    "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": 7429770673042.457, "dur": 5.973,
    "args": {
      "External id": 6456,      "device": 2, "stream": 0,      "correlation": 2173, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2173, "pid": 2, "tid": 0, "ts": 7429770673042.457,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770673048.431, "dur": 5.090,
    "args": {
      "External id": 6457,      "device": 2, "stream": 0,      "correlation": 2174, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2174, "pid": 2, "tid": 0, "ts": 7429770673048.431,
    "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::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": 7429770673053.522, "dur": 4.850,
    "args": {
      "External id": 6458,      "device": 2, "stream": 0,      "correlation": 2175, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2175, "pid": 2, "tid": 0, "ts": 7429770673053.522,
    "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::direct_copy_kernel_cuda(at::TensorIteratorBase&)::{lambda()#3}::operator()() const::{lambda()#7}::operator()() const::{lambda(float)#1}>(at::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": 7429770673058.373, "dur": 6.334,
    "args": {
      "External id": 6475,      "device": 2, "stream": 0,      "correlation": 2176, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2176, "pid": 2, "tid": 0, "ts": 7429770673058.373,
    "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": 7429770673064.708, "dur": 5.572,
    "args": {
      "External id": 6493,      "device": 2, "stream": 0,      "correlation": 2177, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2177, "pid": 2, "tid": 0, "ts": 7429770673064.708,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770673070.281, "dur": 9.982,
    "args": {
      "External id": 6506,      "device": 2, "stream": 0,      "correlation": 2183, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2183, "pid": 2, "tid": 0, "ts": 7429770673070.281,
    "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": 7429770673080.264, "dur": 4.449,
    "args": {
      "External id": 6509,      "device": 2, "stream": 0,      "correlation": 2184, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2184, "pid": 2, "tid": 0, "ts": 7429770673080.264,
    "cat": "ac2g", "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": 7429770673084.714, "dur": 4.730,
    "args": {
      "External id": 6516,      "device": 2, "stream": 0,      "correlation": 2185, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2185, "pid": 2, "tid": 0, "ts": 7429770673084.714,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770673089.445, "dur": 5.612,
    "args": {
      "External id": 6519,      "device": 2, "stream": 0,      "correlation": 2186, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 2186, "pid": 2, "tid": 0, "ts": 7429770673089.445,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770673095.058, "dur": 10.944,
    "args": {
      "External id": 6531,      "device": 2, "stream": 0,      "correlation": 2190, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2190, "pid": 2, "tid": 0, "ts": 7429770673095.058,
    "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": 7429770673106.003, "dur": 5.010,
    "args": {
      "External id": 6540,      "device": 2, "stream": 0,      "correlation": 2191, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2191, "pid": 2, "tid": 0, "ts": 7429770673106.003,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770673111.014, "dur": 63.706,
    "args": {
      "External id": 6550,      "device": 2, "stream": 0,      "correlation": 2195, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2195, "pid": 2, "tid": 0, "ts": 7429770673111.014,
    "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": 7429770673174.721, "dur": 5.291,
    "args": {
      "External id": 6553,      "device": 2, "stream": 0,      "correlation": 2196, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2196, "pid": 2, "tid": 0, "ts": 7429770673174.721,
    "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": 7429770673180.013, "dur": 5.652,
    "args": {
      "External id": 6555,      "device": 2, "stream": 0,      "correlation": 2197, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2197, "pid": 2, "tid": 0, "ts": 7429770673180.013,
    "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": 7429770673185.666, "dur": 7.977,
    "args": {
      "External id": 6558,      "device": 2, "stream": 0,      "correlation": 2198, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 2198, "pid": 2, "tid": 0, "ts": 7429770673185.666,
    "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": 7429770673193.644, "dur": 4.850,
    "args": {
      "External id": 6559,      "device": 2, "stream": 0,      "correlation": 2199, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2199, "pid": 2, "tid": 0, "ts": 7429770673193.644,
    "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": 7429770673198.495, "dur": 5.372,
    "args": {
      "External id": 6560,      "device": 2, "stream": 0,      "correlation": 2200, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2200, "pid": 2, "tid": 0, "ts": 7429770673198.495,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770673203.868, "dur": 5.050,
    "args": {
      "External id": 6561,      "device": 2, "stream": 0,      "correlation": 2201, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2201, "pid": 2, "tid": 0, "ts": 7429770673203.868,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770673208.919, "dur": 5.331,
    "args": {
      "External id": 6563,      "device": 2, "stream": 0,      "correlation": 2202, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2202, "pid": 2, "tid": 0, "ts": 7429770673208.919,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770673214.251, "dur": 253.782,
    "args": {
      "External id": 6571,      "device": 2, "stream": 0,      "correlation": 2208, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2208, "pid": 2, "tid": 0, "ts": 7429770673214.251,
    "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": 7429770673468.034, "dur": 6.534,
    "args": {
      "External id": 6574,      "device": 2, "stream": 0,      "correlation": 2209, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2209, "pid": 2, "tid": 0, "ts": 7429770673468.034,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770673474.569, "dur": 252.619,
    "args": {
      "External id": 6582,      "device": 2, "stream": 0,      "correlation": 2213, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2213, "pid": 2, "tid": 0, "ts": 7429770673474.569,
    "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": 7429770673727.189, "dur": 5.733,
    "args": {
      "External id": 6585,      "device": 2, "stream": 0,      "correlation": 2214, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2214, "pid": 2, "tid": 0, "ts": 7429770673727.189,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770673732.923, "dur": 164.777,
    "args": {
      "External id": 6593,      "device": 2, "stream": 0,      "correlation": 2218, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2218, "pid": 2, "tid": 0, "ts": 7429770673732.923,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770673897.701, "dur": 7.536,
    "args": {
      "External id": 6593,      "device": 2, "stream": 0,      "correlation": 2219, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2219, "pid": 2, "tid": 0, "ts": 7429770673897.701,
    "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": 7429770673905.238, "dur": 5.131,
    "args": {
      "External id": 6596,      "device": 2, "stream": 0,      "correlation": 2220, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2220, "pid": 2, "tid": 0, "ts": 7429770673905.238,
    "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": 7429770673910.370, "dur": 4.730,
    "args": {
      "External id": 6598,      "device": 2, "stream": 0,      "correlation": 2221, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2221, "pid": 2, "tid": 0, "ts": 7429770673910.370,
    "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": 7429770673915.101, "dur": 7.937,
    "args": {
      "External id": 6601,      "device": 2, "stream": 0,      "correlation": 2222, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 2222, "pid": 2, "tid": 0, "ts": 7429770673915.101,
    "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": 7429770673923.039, "dur": 5.131,
    "args": {
      "External id": 6602,      "device": 2, "stream": 0,      "correlation": 2223, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2223, "pid": 2, "tid": 0, "ts": 7429770673923.039,
    "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": 7429770673928.171, "dur": 4.690,
    "args": {
      "External id": 6603,      "device": 2, "stream": 0,      "correlation": 2224, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2224, "pid": 2, "tid": 0, "ts": 7429770673928.171,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770673932.862, "dur": 4.810,
    "args": {
      "External id": 6604,      "device": 2, "stream": 0,      "correlation": 2225, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2225, "pid": 2, "tid": 0, "ts": 7429770673932.862,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770673937.673, "dur": 4.930,
    "args": {
      "External id": 6606,      "device": 2, "stream": 0,      "correlation": 2226, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2226, "pid": 2, "tid": 0, "ts": 7429770673937.673,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770673942.604, "dur": 64.066,
    "args": {
      "External id": 6614,      "device": 2, "stream": 0,      "correlation": 2230, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2230, "pid": 2, "tid": 0, "ts": 7429770673942.604,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770674006.671, "dur": 50.716,
    "args": {
      "External id": 6627,      "device": 2, "stream": 0,      "correlation": 2234, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2234, "pid": 2, "tid": 0, "ts": 7429770674006.671,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770674057.388, "dur": 50.755,
    "args": {
      "External id": 6640,      "device": 2, "stream": 0,      "correlation": 2238, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2238, "pid": 2, "tid": 0, "ts": 7429770674057.388,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770674108.144, "dur": 5.211,
    "args": {
      "External id": 6650,      "device": 2, "stream": 0,      "correlation": 2239, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2239, "pid": 2, "tid": 0, "ts": 7429770674108.144,
    "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::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": 7429770674113.356, "dur": 4.810,
    "args": {
      "External id": 6655,      "device": 2, "stream": 0,      "correlation": 2240, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2240, "pid": 2, "tid": 0, "ts": 7429770674113.356,
    "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": 7429770674118.167, "dur": 5.933,
    "args": {
      "External id": 6656,      "device": 2, "stream": 0,      "correlation": 2241, "kind": "Dispatch Kernel",      "grid":     [84, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2241, "pid": 2, "tid": 0, "ts": 7429770674118.167,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770674129.594, "dur": 4.970,
    "args": {
      "External id": 6657,      "device": 2, "stream": 0,      "correlation": 2242, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2242, "pid": 2, "tid": 0, "ts": 7429770674129.594,
    "cat": "ac2g", "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": 7429770674134.565, "dur": 5.572,
    "args": {
      "External id": 6658,      "device": 2, "stream": 0,      "correlation": 2243, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2243, "pid": 2, "tid": 0, "ts": 7429770674134.565,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770674140.138, "dur": 4.449,
    "args": {
      "External id": 6659,      "device": 2, "stream": 0,      "correlation": 2244, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2244, "pid": 2, "tid": 0, "ts": 7429770674140.138,
    "cat": "ac2g", "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": 7429770674144.588, "dur": 4.489,
    "args": {
      "External id": 6664,      "device": 2, "stream": 0,      "correlation": 2245, "kind": "Dispatch Kernel",      "grid":     [168, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2245, "pid": 2, "tid": 0, "ts": 7429770674144.588,
    "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": 7429770674149.078, "dur": 6.174,
    "args": {
      "External id": 6665,      "device": 2, "stream": 0,      "correlation": 2246, "kind": "Dispatch Kernel",      "grid":     [21, 2, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2246, "pid": 2, "tid": 0, "ts": 7429770674149.078,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770674155.253, "dur": 5.571,
    "args": {
      "External id": 6666,      "device": 2, "stream": 0,      "correlation": 2247, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2247, "pid": 2, "tid": 0, "ts": 7429770674155.253,
    "cat": "ac2g", "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": 7429770674160.825, "dur": 5.532,
    "args": {
      "External id": 6667,      "device": 2, "stream": 0,      "correlation": 2248, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2248, "pid": 2, "tid": 0, "ts": 7429770674160.825,
    "cat": "ac2g", "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": 7429770674166.358, "dur": 6.695,
    "args": {
      "External id": 6684,      "device": 2, "stream": 0,      "correlation": 2249, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2249, "pid": 2, "tid": 0, "ts": 7429770674166.358,
    "cat": "ac2g", "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": 7429770674173.054, "dur": 5.411,
    "args": {
      "External id": 6702,      "device": 2, "stream": 0,      "correlation": 2250, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2250, "pid": 2, "tid": 0, "ts": 7429770674173.054,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT32x128x16_MI32x32x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA4_LPB4_LPM0_LRVW4_LWPMn1_MIAV0_MIWT1_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770674178.466, "dur": 9.822,
    "args": {
      "External id": 6715,      "device": 2, "stream": 0,      "correlation": 2256, "kind": "Dispatch Kernel",      "grid":     [24576, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2256, "pid": 2, "tid": 0, "ts": 7429770674178.466,
    "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": 7429770674188.289, "dur": 4.890,
    "args": {
      "External id": 6718,      "device": 2, "stream": 0,      "correlation": 2257, "kind": "Dispatch Kernel",      "grid":     [221, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2257, "pid": 2, "tid": 0, "ts": 7429770674188.289,
    "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::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": 7429770674193.180, "dur": 5.291,
    "args": {
      "External id": 6725,      "device": 2, "stream": 0,      "correlation": 2258, "kind": "Dispatch Kernel",      "grid":     [882, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2258, "pid": 2, "tid": 0, "ts": 7429770674193.180,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void (anonymous namespace)::softmax_warp_forward<float, float, float, 7, false, false>(float*, float const*, int, int, int, bool const*, int, bool)", "pid": 2, "tid": 0,
    "ts": 7429770674198.472, "dur": 5.010,
    "args": {
      "External id": 6728,      "device": 2, "stream": 0,      "correlation": 2259, "kind": "Dispatch Kernel",      "grid":     [672, 1, 1], "block":     [64, 2, 1]
    }
  },
  {
    "ph": "f", "id": 2259, "pid": 2, "tid": 0, "ts": 7429770674198.472,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Ailk_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x16_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA2_GRVWB2_GSUAMB_ISA942_IU1_K1_LBSPPA0_LBSPPB128_LBSPPM0_LPA0_LPB2_LPM0_LRVW2_LWPMn1_MIAV0_MIWT2_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_WSGRA0_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770674203.483, "dur": 10.864,
    "args": {
      "External id": 6740,      "device": 2, "stream": 0,      "correlation": 2263, "kind": "Dispatch Kernel",      "grid":     [16384, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2263, "pid": 2, "tid": 0, "ts": 7429770674203.483,
    "cat": "ac2g", "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": 7429770674214.348, "dur": 5.292,
    "args": {
      "External id": 6749,      "device": 2, "stream": 0,      "correlation": 2264, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2264, "pid": 2, "tid": 0, "ts": 7429770674214.348,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x32x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_1_MO40_NTn1_NTA0_NTB0_NTC0_NTD0_NTM0_NEPBS2_NLCA1_NLCB1_ONLL1_PGR2_PLR1_PKA1_SIA3_SS1_SPO1_SRVW0_SSO0_SVW1_TLDS1_ULSGRO0_USL1_UIOFGRO0_USFGROn1_VSn1_VWA1_VWB1_WSGRA1_WSGRB1_WS64_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770674219.641, "dur": 64.667,
    "args": {
      "External id": 6759,      "device": 2, "stream": 0,      "correlation": 2268, "kind": "Dispatch Kernel",      "grid":     [49152, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2268, "pid": 2, "tid": 0, "ts": 7429770674219.641,
    "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": 7429770674284.309, "dur": 5.371,
    "args": {
      "External id": 6762,      "device": 2, "stream": 0,      "correlation": 2269, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2269, "pid": 2, "tid": 0, "ts": 7429770674284.309,
    "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": 7429770674289.681, "dur": 5.412,
    "args": {
      "External id": 6764,      "device": 2, "stream": 0,      "correlation": 2270, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2270, "pid": 2, "tid": 0, "ts": 7429770674289.681,
    "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": 7429770674295.094, "dur": 7.977,
    "args": {
      "External id": 6767,      "device": 2, "stream": 0,      "correlation": 2271, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 2271, "pid": 2, "tid": 0, "ts": 7429770674295.094,
    "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": 7429770674303.072, "dur": 4.369,
    "args": {
      "External id": 6768,      "device": 2, "stream": 0,      "correlation": 2272, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2272, "pid": 2, "tid": 0, "ts": 7429770674303.072,
    "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": 7429770674307.442, "dur": 5.131,
    "args": {
      "External id": 6769,      "device": 2, "stream": 0,      "correlation": 2273, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2273, "pid": 2, "tid": 0, "ts": 7429770674307.442,
    "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": 7429770674312.574, "dur": 5.251,
    "args": {
      "External id": 6770,      "device": 2, "stream": 0,      "correlation": 2274, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2274, "pid": 2, "tid": 0, "ts": 7429770674312.574,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770674317.826, "dur": 5.131,
    "args": {
      "External id": 6772,      "device": 2, "stream": 0,      "correlation": 2275, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2275, "pid": 2, "tid": 0, "ts": 7429770674317.826,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770674322.958, "dur": 252.940,
    "args": {
      "External id": 6780,      "device": 2, "stream": 0,      "correlation": 2279, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2279, "pid": 2, "tid": 0, "ts": 7429770674322.958,
    "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": 7429770674575.899, "dur": 5.331,
    "args": {
      "External id": 6783,      "device": 2, "stream": 0,      "correlation": 2280, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2280, "pid": 2, "tid": 0, "ts": 7429770674575.899,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT160x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT5_3_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770674581.231, "dur": 252.740,
    "args": {
      "External id": 6791,      "device": 2, "stream": 0,      "correlation": 2284, "kind": "Dispatch Kernel",      "grid":     [23040, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2284, "pid": 2, "tid": 0, "ts": 7429770674581.231,
    "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": 7429770674833.972, "dur": 5.812,
    "args": {
      "External id": 6794,      "device": 2, "stream": 0,      "correlation": 2285, "kind": "Dispatch Kernel",      "grid":     [1176, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2285, "pid": 2, "tid": 0, "ts": 7429770674833.972,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT96x96x32_MI16x16x1_SN_LDSB0_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT3_3_MO1_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770674839.785, "dur": 164.657,
    "args": {
      "External id": 6802,      "device": 2, "stream": 0,      "correlation": 2289, "kind": "Dispatch Kernel",      "grid":     [165120, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2289, "pid": 2, "tid": 0, "ts": 7429770674839.785,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_SS_BiasS_AS_ScaleAlphaVec_PostGSU16_VW4", "pid": 2, "tid": 0,
    "ts": 7429770675004.443, "dur": 7.376,
    "args": {
      "External id": 6802,      "device": 2, "stream": 0,      "correlation": 2290, "kind": "Dispatch Kernel",      "grid":     [86016, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2290, "pid": 2, "tid": 0, "ts": 7429770675004.443,
    "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": 7429770675011.820, "dur": 4.810,
    "args": {
      "External id": 6805,      "device": 2, "stream": 0,      "correlation": 2291, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2291, "pid": 2, "tid": 0, "ts": 7429770675011.820,
    "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": 7429770675016.631, "dur": 5.051,
    "args": {
      "External id": 6807,      "device": 2, "stream": 0,      "correlation": 2292, "kind": "Dispatch Kernel",      "grid":     [336, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2292, "pid": 2, "tid": 0, "ts": 7429770675016.631,
    "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": 7429770675021.683, "dur": 7.937,
    "args": {
      "External id": 6810,      "device": 2, "stream": 0,      "correlation": 2293, "kind": "Dispatch Kernel",      "grid":     [11, 1, 1], "block":     [64, 8, 1]
    }
  },
  {
    "ph": "f", "id": 2293, "pid": 2, "tid": 0, "ts": 7429770675021.683,
    "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": 7429770675029.621, "dur": 4.409,
    "args": {
      "External id": 6811,      "device": 2, "stream": 0,      "correlation": 2294, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2294, "pid": 2, "tid": 0, "ts": 7429770675029.621,
    "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": 7429770675034.031, "dur": 4.610,
    "args": {
      "External id": 6812,      "device": 2, "stream": 0,      "correlation": 2295, "kind": "Dispatch Kernel",      "grid":     [1, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2295, "pid": 2, "tid": 0, "ts": 7429770675034.031,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "void at::native::elementwise_kernel<128, 2, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770675038.642, "dur": 4.850,
    "args": {
      "External id": 6813,      "device": 2, "stream": 0,      "correlation": 2296, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2296, "pid": 2, "tid": 0, "ts": 7429770675038.642,
    "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::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1}>(int, at::native::gpu_kernel_impl_nocast<at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > >(at::TensorIteratorBase&, at::native::BinaryFunctor<float, float, float, at::native::binary_internal::MulFunctor<float> > const&)::{lambda(int)#1})", "pid": 2, "tid": 0,
    "ts": 7429770675043.493, "dur": 5.010,
    "args": {
      "External id": 6815,      "device": 2, "stream": 0,      "correlation": 2297, "kind": "Dispatch Kernel",      "grid":     [1344, 1, 1], "block":     [128, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2297, "pid": 2, "tid": 0, "ts": 7429770675043.493,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "ph": "X", "cat": "kernel", "name": "Cijk_Alik_Bljk_SB_Bias_AS_SAV_UserArgs_MT64x96x32_MI16x16x1_SN_LDSB1_AFC1_AFEM1_AFEM1_ASEM1_CLR1_CADS0_EPS0_GRVWA4_GRVWB4_GSUAMB_ISA942_IU1_K1_LBSPPA128_LBSPPB128_LBSPPM0_LPA8_LPB8_LPM0_LRVW4_LWPMn1_MIAV0_MIWT2_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_WG32_8_1_WGMXCC1_WGMXCCG0", "pid": 2, "tid": 0,
    "ts": 7429770675048.504, "dur": 1143.385,
    "args": {
      "External id": 6829,      "device": 2, "stream": 0,      "correlation": 2303, "kind": "Dispatch Kernel",      "grid":     [513024, 1, 1], "block":     [256, 1, 1]
    }
  },
  {
    "ph": "f", "id": 2303, "pid": 2, "tid": 0, "ts": 7429770675048.504,
    "cat": "ac2g", "name": "ac2g", "bp": "e"
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429770598901.728, "pid": 870, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429770598901.728, "pid": 870, "tid": 0,
    "args": {
      "labels": "CPU"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429770598901.728, "pid": 870, "tid": 0,
    "args": {
      "sort_index": 870
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429770598901.728, "pid": 0, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429770598901.728, "pid": 0, "tid": 0,
    "args": {
      "labels": "GPU 0"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429770598901.728, "pid": 0, "tid": 0,
    "args": {
      "sort_index": 5000000
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429770598901.728, "pid": 1, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429770598901.728, "pid": 1, "tid": 0,
    "args": {
      "labels": "GPU 1"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429770598901.728, "pid": 1, "tid": 0,
    "args": {
      "sort_index": 5000001
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429770598901.728, "pid": 2, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429770598901.728, "pid": 2, "tid": 0,
    "args": {
      "labels": "GPU 2"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429770598901.728, "pid": 2, "tid": 0,
    "args": {
      "sort_index": 5000002
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429770598901.728, "pid": 3, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429770598901.728, "pid": 3, "tid": 0,
    "args": {
      "labels": "GPU 3"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429770598901.728, "pid": 3, "tid": 0,
    "args": {
      "sort_index": 5000003
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429770598901.728, "pid": 4, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429770598901.728, "pid": 4, "tid": 0,
    "args": {
      "labels": "GPU 4"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429770598901.728, "pid": 4, "tid": 0,
    "args": {
      "sort_index": 5000004
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429770598901.728, "pid": 5, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429770598901.728, "pid": 5, "tid": 0,
    "args": {
      "labels": "GPU 5"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429770598901.728, "pid": 5, "tid": 0,
    "args": {
      "sort_index": 5000005
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429770598901.728, "pid": 6, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429770598901.728, "pid": 6, "tid": 0,
    "args": {
      "labels": "GPU 6"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429770598901.728, "pid": 6, "tid": 0,
    "args": {
      "sort_index": 5000006
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429770598901.728, "pid": 7, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429770598901.728, "pid": 7, "tid": 0,
    "args": {
      "labels": "GPU 7"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429770598901.728, "pid": 7, "tid": 0,
    "args": {
      "sort_index": 5000007
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429770598901.728, "pid": 8, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429770598901.728, "pid": 8, "tid": 0,
    "args": {
      "labels": "GPU 8"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429770598901.728, "pid": 8, "tid": 0,
    "args": {
      "sort_index": 5000008
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429770598901.728, "pid": 9, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429770598901.728, "pid": 9, "tid": 0,
    "args": {
      "labels": "GPU 9"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429770598901.728, "pid": 9, "tid": 0,
    "args": {
      "sort_index": 5000009
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429770598901.728, "pid": 10, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429770598901.728, "pid": 10, "tid": 0,
    "args": {
      "labels": "GPU 10"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429770598901.728, "pid": 10, "tid": 0,
    "args": {
      "sort_index": 5000010
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429770598901.728, "pid": 11, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429770598901.728, "pid": 11, "tid": 0,
    "args": {
      "labels": "GPU 11"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429770598901.728, "pid": 11, "tid": 0,
    "args": {
      "sort_index": 5000011
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429770598901.728, "pid": 12, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429770598901.728, "pid": 12, "tid": 0,
    "args": {
      "labels": "GPU 12"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429770598901.728, "pid": 12, "tid": 0,
    "args": {
      "sort_index": 5000012
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429770598901.728, "pid": 13, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429770598901.728, "pid": 13, "tid": 0,
    "args": {
      "labels": "GPU 13"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429770598901.728, "pid": 13, "tid": 0,
    "args": {
      "sort_index": 5000013
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429770598901.728, "pid": 14, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429770598901.728, "pid": 14, "tid": 0,
    "args": {
      "labels": "GPU 14"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429770598901.728, "pid": 14, "tid": 0,
    "args": {
      "sort_index": 5000014
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 7429770598901.728, "pid": 15, "tid": 0,
    "args": {
      "name": "python"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 7429770598901.728, "pid": 15, "tid": 0,
    "args": {
      "labels": "GPU 15"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429770598901.728, "pid": 15, "tid": 0,
    "args": {
      "sort_index": 5000015
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 7429770598901.728, "pid": 0, "tid": 0,
    "args": {
      "name": "thread 0 (python)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 7429770598901.728, "pid": 0, "tid": 0,
    "args": {
      "sort_index": 0
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 7429770598901.728, "pid": 2, "tid": 0,
    "args": {
      "name": "stream 0 "
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 7429770598901.728, "pid": 2, "tid": 0,
    "args": {
      "sort_index": 0
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 7429770598901.728, "pid": 870, "tid": 870,
    "args": {
      "name": "thread 870 (python)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 7429770598901.728, "pid": 870, "tid": 870,
    "args": {
      "sort_index": 870
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 7429770598901.728, "pid": 870, "tid": 870,
    "args": {
      "name": "thread 870 (python)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 7429770598901.728, "pid": 870, "tid": 870,
    "args": {
      "sort_index": 870
    }
  },
  {
    "ph": "X", "cat": "Trace", "ts": 7429770598768.657, "dur": 77443.564,
    "pid": "Spans", "tid": "PyTorch Profiler",
    "name": "PyTorch Profiler (0)",
    "args": {
      "Op count": 0
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 7429770598768.657,
    "pid": "Spans", "tid": 0,
    "args": {
      "sort_index": 536870912
    }
  },
  {
    "name": "Iteration Start: PyTorch Profiler", "ph": "i", "s": "g",
    "pid": "Traces", "tid": "Trace PyTorch Profiler", "ts": 7429770598768.657
  },
  {
    "name": "Record Window End", "ph": "i", "s": "g",
    "pid": "", "tid": "", "ts": 7429770691329.721
  }
  ],
  "traceName": "./log_llama/felix-test-job-vllm-3-4dknr_870.1750951368726808399.pt.trace.json",
  "displayTimeUnit": "ms",
  "baseTimeNanoseconds": 1743521598000000000
}