{
  "schemaVersion": 1,
  "deviceProperties": [
    {
      "id": 0, "name": "NVIDIA H200", "totalGlobalMem": 149631664128,
      "computeMajor": 9, "computeMinor": 0,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 2048,
      "regsPerBlock": 65536, "warpSize": 32,
      "sharedMemPerBlock": 49152, "numSms": 132
    , "regsPerMultiprocessor": 65536, "sharedMemPerBlockOptin": 232448, "sharedMemPerMultiprocessor": 233472
    },
    {
      "id": 1, "name": "NVIDIA H200", "totalGlobalMem": 149631664128,
      "computeMajor": 9, "computeMinor": 0,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 2048,
      "regsPerBlock": 65536, "warpSize": 32,
      "sharedMemPerBlock": 49152, "numSms": 132
    , "regsPerMultiprocessor": 65536, "sharedMemPerBlockOptin": 232448, "sharedMemPerMultiprocessor": 233472
    },
    {
      "id": 2, "name": "NVIDIA H200", "totalGlobalMem": 149631664128,
      "computeMajor": 9, "computeMinor": 0,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 2048,
      "regsPerBlock": 65536, "warpSize": 32,
      "sharedMemPerBlock": 49152, "numSms": 132
    , "regsPerMultiprocessor": 65536, "sharedMemPerBlockOptin": 232448, "sharedMemPerMultiprocessor": 233472
    },
    {
      "id": 3, "name": "NVIDIA H200", "totalGlobalMem": 149631664128,
      "computeMajor": 9, "computeMinor": 0,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 2048,
      "regsPerBlock": 65536, "warpSize": 32,
      "sharedMemPerBlock": 49152, "numSms": 132
    , "regsPerMultiprocessor": 65536, "sharedMemPerBlockOptin": 232448, "sharedMemPerMultiprocessor": 233472
    },
    {
      "id": 4, "name": "NVIDIA H200", "totalGlobalMem": 149631664128,
      "computeMajor": 9, "computeMinor": 0,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 2048,
      "regsPerBlock": 65536, "warpSize": 32,
      "sharedMemPerBlock": 49152, "numSms": 132
    , "regsPerMultiprocessor": 65536, "sharedMemPerBlockOptin": 232448, "sharedMemPerMultiprocessor": 233472
    },
    {
      "id": 5, "name": "NVIDIA H200", "totalGlobalMem": 149631664128,
      "computeMajor": 9, "computeMinor": 0,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 2048,
      "regsPerBlock": 65536, "warpSize": 32,
      "sharedMemPerBlock": 49152, "numSms": 132
    , "regsPerMultiprocessor": 65536, "sharedMemPerBlockOptin": 232448, "sharedMemPerMultiprocessor": 233472
    },
    {
      "id": 6, "name": "NVIDIA H200", "totalGlobalMem": 149631664128,
      "computeMajor": 9, "computeMinor": 0,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 2048,
      "regsPerBlock": 65536, "warpSize": 32,
      "sharedMemPerBlock": 49152, "numSms": 132
    , "regsPerMultiprocessor": 65536, "sharedMemPerBlockOptin": 232448, "sharedMemPerMultiprocessor": 233472
    },
    {
      "id": 7, "name": "NVIDIA H200", "totalGlobalMem": 149631664128,
      "computeMajor": 9, "computeMinor": 0,
      "maxThreadsPerBlock": 1024, "maxThreadsPerMultiprocessor": 2048,
      "regsPerBlock": 65536, "warpSize": 32,
      "sharedMemPerBlock": 49152, "numSms": 132
    , "regsPerMultiprocessor": 65536, "sharedMemPerBlockOptin": 232448, "sharedMemPerMultiprocessor": 233472
    }
  ],
  "cupti_version": 22,
  "cuda_runtime_version": 12040,
  "cuda_driver_version": 12080,
  "distributedInfo": {"backend": "nccl", "rank": 6, "world_size": 8, "pg_count": 1, "pg_config": [{"pg_name": "0", "pg_desc": "default_pg", "backend_config": "cuda:nccl", "pg_size": 8, "ranks": [0, 1, 2, 3, 4, 5, 6, 7]}], "nccl_version": "2.21.5"},
  "record_shapes": 1,
  "trace_id": "09A2E5B3241E4AFC839F02E682121BD6",
  "traceEvents": [
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: DivBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656709756.364, "dur": 109.015,
    "args": {
      "External id": 2483201,"Record function id": 0, "Sequence number": 24740650, "Fwd thread id": 1, "Ev Idx": 0
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "DivBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656709774.062, "dur": 83.015,
    "args": {
      "External id": 2483202,"Sequence number": 24740650, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 1
    }
  },
  {
    "ph": "f", "id": 1, "pid": 1336759, "tid": 1381173, "ts": 1514656709774.062,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::div", "pid": 1336759, "tid": 1381173,
    "ts": 1514656709781.476, "dur": 73.534,
    "args": {
      "External id": 2483203,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "long int"], "Input Strides": [[1], []], "Input Dims": [[1], []], "Ev Idx": 2
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: AddBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656709878.836, "dur": 259.620,
    "args": {
      "External id": 2483204,"Record function id": 0, "Ev Idx": 3
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656709928.494, "dur": 134.002,
    "args": {
      "External id": 2483205,"Record function id": 0, "Ev Idx": 4
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.27", "pid": 1336759, "tid": 1381173,
    "ts": 1514656709958.454, "dur": 90.768,
    "args": {
      "External id": 2483206,"Record function id": 0, "Ev Idx": 5
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AddBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656710068.100, "dur": 2.034,
    "args": {
      "External id": 2483207,"Sequence number": 24740649, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 6
    }
  },
  {
    "ph": "f", "id": 2, "pid": 1336759, "tid": 1381173, "ts": 1514656710068.100,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1381173,
    "ts": 1514656710073.856, "dur": 59.804,
    "args": {
      "External id": 2483208,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[1], []], "Input Dims": [[1], []], "Ev Idx": 7
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1381173,
    "ts": 1514656710083.774, "dur": 49.418,
    "args": {
      "External id": 2483209,"Record function id": 0, "Concrete Inputs": ["", "[]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[1], [], [], []], "Ev Idx": 8
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656710095.076, "dur": 4.403,
    "args": {
      "External id": 2483210,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: FusedLinearCrossEntropyFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656710148.846, "dur": 38038.597,
    "args": {
      "External id": 2483211,"Record function id": 0, "Sequence number": 24740647, "Fwd thread id": 1, "Ev Idx": 10
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FusedLinearCrossEntropyFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656710150.636, "dur": 38026.109,
    "args": {
      "External id": 2483212,"Sequence number": 24740647, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 11
    }
  },
  {
    "ph": "f", "id": 3, "pid": 1336759, "tid": 1381173, "ts": 1514656710150.636,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656710188.798, "dur": 3.632,
    "args": {
      "External id": 2483213,"Record function id": 0, "Concrete Inputs": ["[]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 12
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514656710198.120, "dur": 37865.317,
    "args": {
      "External id": 2483214,"Record function id": 0, "Concrete Inputs": ["", "", "6", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 13
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514656710200.199, "dur": 37862.752,
    "args": {
      "External id": 2483215,"Record function id": 0, "Concrete Inputs": ["", "6", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], []], "Ev Idx": 14
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656710205.234, "dur": 8.041,
    "args": {
      "External id": 2483216,"Record function id": 0, "Concrete Inputs": ["[]", "[]", "6", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 15
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656710215.119, "dur": 37846.229,
    "args": {
      "External id": 2483217,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 16
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748068.404, "dur": 0.607,
    "args": {
      "External id": 2483218,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 17
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748071.215, "dur": 2.775,
    "args": {
      "External id": 2483219,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 18
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748072.599, "dur": 1.218,
    "args": {
      "External id": 2483220,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 19
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::ne", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748080.257, "dur": 32.845,
    "args": {
      "External id": 2483221,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 20
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::is_nonzero", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748121.608, "dur": 45.372,
    "args": {
      "External id": 2483222,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 21
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748122.890, "dur": 43.890,
    "args": {
      "External id": 2483223,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 22
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748124.430, "dur": 42.035,
    "args": {
      "External id": 2483224,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 23
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748199.277, "dur": 18.500,
    "args": {
      "External id": 2483225,"Record function id": 0, "Sequence number": 24740646, "Fwd thread id": 1, "Ev Idx": 24
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748200.991, "dur": 14.296,
    "args": {
      "External id": 2483226,"Sequence number": 24740646, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 25
    }
  },
  {
    "ph": "f", "id": 4, "pid": 1336759, "tid": 1381173, "ts": 1514656748200.991,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748204.946, "dur": 10.092,
    "args": {
      "External id": 2483227,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 26
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748208.930, "dur": 5.917,
    "args": {
      "External id": 2483228,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 27
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SliceBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748221.820, "dur": 90.228,
    "args": {
      "External id": 2483229,"Record function id": 0, "Sequence number": 24740645, "Fwd thread id": 1, "Ev Idx": 28
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SliceBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748222.842, "dur": 81.933,
    "args": {
      "External id": 2483230,"Sequence number": 24740645, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 29
    }
  },
  {
    "ph": "f", "id": 5, "pid": 1336759, "tid": 1381173, "ts": 1514656748222.842,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748226.563, "dur": 77.723,
    "args": {
      "External id": 2483231,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]", "2", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 2048], [], [], [], [], []], "Ev Idx": 30
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748231.303, "dur": 30.017,
    "args": {
      "External id": 2483232,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 2048]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 31
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748235.762, "dur": 4.351,
    "args": {
      "External id": 2483233,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 2048]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 32
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748241.407, "dur": 19.611,
    "args": {
      "External id": 2483234,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 33
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748243.896, "dur": 16.606,
    "args": {
      "External id": 2483235,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 34
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748266.507, "dur": 4.506,
    "args": {
      "External id": 2483236,"Record function id": 0, "Concrete Inputs": ["", "2", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 1], [], [], [], []], "Input Dims": [[16, 4096, 2048], [], [], [], []], "Ev Idx": 35
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748269.125, "dur": 1.585,
    "args": {
      "External id": 2483237,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]", "[8388608, 2048, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8388608, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 2048], [], [], []], "Ev Idx": 36
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748272.104, "dur": 31.122,
    "args": {
      "External id": 2483238,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 37
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SelectBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748316.297, "dur": 61.524,
    "args": {
      "External id": 2483239,"Record function id": 0, "Sequence number": 24740644, "Fwd thread id": 1, "Ev Idx": 38
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SelectBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748317.326, "dur": 57.010,
    "args": {
      "External id": 2483240,"Sequence number": 24740644, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 39
    }
  },
  {
    "ph": "f", "id": 6, "pid": 1336759, "tid": 1381173, "ts": 1514656748317.326,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748320.283, "dur": 53.753,
    "args": {
      "External id": 2483241,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 2048]", "2", "3"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 2048], [], [], []], "Ev Idx": 40
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748326.112, "dur": 19.031,
    "args": {
      "External id": 2483242,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 2048]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 41
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748326.888, "dur": 2.996,
    "args": {
      "External id": 2483243,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 2048]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 42
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748330.457, "dur": 14.429,
    "args": {
      "External id": 2483244,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[33554432, 8192, 2048, 1]], "Input Dims": [[16, 4096, 4, 2048]], "Ev Idx": 43
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748333.913, "dur": 10.559,
    "args": {
      "External id": 2483245,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], []], "Input Dims": [[16, 4096, 4, 2048], []], "Ev Idx": 44
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748349.237, "dur": 5.105,
    "args": {
      "External id": 2483246,"Record function id": 0, "Concrete Inputs": ["", "2", "3"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], []], "Input Dims": [[16, 4096, 4, 2048], [], []], "Ev Idx": 45
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748352.740, "dur": 0.950,
    "args": {
      "External id": 2483247,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]", "[33554432, 8192, 1]", "6144"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], []], "Ev Idx": 46
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748355.249, "dur": 18.213,
    "args": {
      "External id": 2483248,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 47
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SliceBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748381.564, "dur": 112.790,
    "args": {
      "External id": 2483249,"Record function id": 0, "Sequence number": 24740643, "Fwd thread id": 1, "Ev Idx": 48
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SliceBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748382.498, "dur": 108.711,
    "args": {
      "External id": 2483250,"Sequence number": 24740643, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[33554432, 8192, 2048, 1]], "Input Dims": [[16, 4096, 4, 2048]], "Ev Idx": 49
    }
  },
  {
    "ph": "f", "id": 7, "pid": 1336759, "tid": 1381173, "ts": 1514656748382.498,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748386.268, "dur": 104.562,
    "args": {
      "External id": 2483251,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 2048]", "1", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], [], [], []], "Ev Idx": 50
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748389.901, "dur": 14.264,
    "args": {
      "External id": 2483252,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 2048]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 51
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748390.749, "dur": 2.157,
    "args": {
      "External id": 2483253,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 2048]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 52
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748393.502, "dur": 10.341,
    "args": {
      "External id": 2483254,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[33554432, 8192, 2048, 1]], "Input Dims": [[16, 4096, 4, 2048]], "Ev Idx": 53
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748394.514, "dur": 8.994,
    "args": {
      "External id": 2483255,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], []], "Input Dims": [[16, 4096, 4, 2048], []], "Ev Idx": 54
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748408.078, "dur": 2.148,
    "args": {
      "External id": 2483256,"Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], [], []], "Ev Idx": 55
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748409.179, "dur": 0.874,
    "args": {
      "External id": 2483257,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 2048]", "[33554432, 8192, 2048, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], []], "Ev Idx": 56
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748410.913, "dur": 79.083,
    "args": {
      "External id": 2483258,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [33554432, 8192, 2048, 1], []], "Input Dims": [[16, 4096, 4, 2048], [16, 4096, 4, 2048], []], "Ev Idx": 57
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SliceBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748501.039, "dur": 113.525,
    "args": {
      "External id": 2483259,"Record function id": 0, "Sequence number": 24740642, "Fwd thread id": 1, "Ev Idx": 58
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SliceBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748501.816, "dur": 109.057,
    "args": {
      "External id": 2483260,"Sequence number": 24740642, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[33554432, 8192, 2048, 1]], "Input Dims": [[16, 4096, 4, 2048]], "Ev Idx": 59
    }
  },
  {
    "ph": "f", "id": 8, "pid": 1336759, "tid": 1381173, "ts": 1514656748501.816,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748503.198, "dur": 107.406,
    "args": {
      "External id": 2483261,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 2048]", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], [], [], []], "Ev Idx": 60
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748506.876, "dur": 19.198,
    "args": {
      "External id": 2483262,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 2048]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 61
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748507.862, "dur": 4.166,
    "args": {
      "External id": 2483263,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 2048]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 62
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748512.765, "dur": 13.063,
    "args": {
      "External id": 2483264,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[33554432, 8192, 2048, 1]], "Input Dims": [[16, 4096, 4, 2048]], "Ev Idx": 63
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748513.875, "dur": 11.641,
    "args": {
      "External id": 2483265,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], []], "Input Dims": [[16, 4096, 4, 2048], []], "Ev Idx": 64
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748540.246, "dur": 6.228,
    "args": {
      "External id": 2483266,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], [], []], "Ev Idx": 65
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748545.336, "dur": 0.747,
    "args": {
      "External id": 2483267,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 2048]", "[33554432, 8192, 2048, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], []], "Ev Idx": 66
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748549.305, "dur": 60.498,
    "args": {
      "External id": 2483268,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [33554432, 8192, 2048, 1], []], "Input Dims": [[16, 4096, 4, 2048], [16, 4096, 4, 2048], []], "Ev Idx": 67
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: AddBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748620.477, "dur": 37.624,
    "args": {
      "External id": 2483269,"Record function id": 0, "Sequence number": 24740641, "Fwd thread id": 1, "Ev Idx": 68
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AddBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748622.071, "dur": 1.240,
    "args": {
      "External id": 2483270,"Sequence number": 24740641, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 69
    }
  },
  {
    "ph": "f", "id": 9, "pid": 1336759, "tid": 1381173, "ts": 1514656748622.071,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748625.746, "dur": 29.709,
    "args": {
      "External id": 2483271,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[1], []], "Input Dims": [[1], []], "Ev Idx": 70
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748627.842, "dur": 27.160,
    "args": {
      "External id": 2483272,"Record function id": 0, "Concrete Inputs": ["", "[]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[1], [], [], []], "Ev Idx": 71
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748637.537, "dur": 0.567,
    "args": {
      "External id": 2483273,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 72
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: FusedLinearCrossEntropyFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748663.015, "dur": 2269.044,
    "args": {
      "External id": 2483274,"Record function id": 0, "Sequence number": 24740639, "Fwd thread id": 1, "Ev Idx": 73
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FusedLinearCrossEntropyFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748664.611, "dur": 2236.934,
    "args": {
      "External id": 2483275,"Sequence number": 24740639, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 74
    }
  },
  {
    "ph": "f", "id": 10, "pid": 1336759, "tid": 1381173, "ts": 1514656748664.611,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748701.877, "dur": 3.545,
    "args": {
      "External id": 2483276,"Record function id": 0, "Concrete Inputs": ["[]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 75
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748707.812, "dur": 2105.220,
    "args": {
      "External id": 2483277,"Record function id": 0, "Concrete Inputs": ["", "", "6", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 76
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748709.607, "dur": 2102.962,
    "args": {
      "External id": 2483278,"Record function id": 0, "Concrete Inputs": ["", "6", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], []], "Ev Idx": 77
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748712.226, "dur": 5.295,
    "args": {
      "External id": 2483279,"Record function id": 0, "Concrete Inputs": ["[]", "[]", "6", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 78
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656748718.524, "dur": 2093.201,
    "args": {
      "External id": 2483280,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 79
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1336759, "tid": 1381173,
    "ts": 1514656750816.126, "dur": 0.320,
    "args": {
      "External id": 2483281,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 80
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656750819.996, "dur": 2.430,
    "args": {
      "External id": 2483282,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 81
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656750821.286, "dur": 0.855,
    "args": {
      "External id": 2483283,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 82
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::ne", "pid": 1336759, "tid": 1381173,
    "ts": 1514656750826.416, "dur": 20.512,
    "args": {
      "External id": 2483284,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 83
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::is_nonzero", "pid": 1336759, "tid": 1381173,
    "ts": 1514656750851.676, "dur": 42.496,
    "args": {
      "External id": 2483285,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 84
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1336759, "tid": 1381173,
    "ts": 1514656750852.788, "dur": 41.191,
    "args": {
      "External id": 2483286,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 85
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1336759, "tid": 1381173,
    "ts": 1514656750853.994, "dur": 39.498,
    "args": {
      "External id": 2483287,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 86
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656750909.688, "dur": 19.050,
    "args": {
      "External id": 2483288,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 87
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656750939.518, "dur": 14.715,
    "args": {
      "External id": 2483289,"Record function id": 0, "Sequence number": 24740638, "Fwd thread id": 1, "Ev Idx": 88
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656750940.730, "dur": 11.652,
    "args": {
      "External id": 2483290,"Sequence number": 24740638, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 89
    }
  },
  {
    "ph": "f", "id": 11, "pid": 1336759, "tid": 1381173, "ts": 1514656750940.730,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656750947.475, "dur": 4.687,
    "args": {
      "External id": 2483291,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 90
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656750948.608, "dur": 3.374,
    "args": {
      "External id": 2483292,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 91
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SliceBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656750957.647, "dur": 95.855,
    "args": {
      "External id": 2483293,"Record function id": 0, "Sequence number": 24740637, "Fwd thread id": 1, "Ev Idx": 92
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SliceBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656750958.649, "dur": 89.576,
    "args": {
      "External id": 2483294,"Sequence number": 24740637, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 93
    }
  },
  {
    "ph": "f", "id": 12, "pid": 1336759, "tid": 1381173, "ts": 1514656750958.649,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656750960.280, "dur": 87.502,
    "args": {
      "External id": 2483295,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]", "2", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 2048], [], [], [], [], []], "Ev Idx": 94
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 1336759, "tid": 1381173,
    "ts": 1514656750964.623, "dur": 46.256,
    "args": {
      "External id": 2483296,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 2048]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 95
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656750966.034, "dur": 3.215,
    "args": {
      "External id": 2483297,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 2048]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 96
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656750970.096, "dur": 40.531,
    "args": {
      "External id": 2483298,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 97
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656750971.396, "dur": 38.482,
    "args": {
      "External id": 2483299,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 98
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751013.020, "dur": 5.025,
    "args": {
      "External id": 2483300,"Record function id": 0, "Concrete Inputs": ["", "2", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 1], [], [], [], []], "Input Dims": [[16, 4096, 2048], [], [], [], []], "Ev Idx": 99
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751015.224, "dur": 2.605,
    "args": {
      "External id": 2483301,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]", "[8388608, 2048, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8388608, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 2048], [], [], []], "Ev Idx": 100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751018.823, "dur": 28.081,
    "args": {
      "External id": 2483302,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SelectBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751061.534, "dur": 53.426,
    "args": {
      "External id": 2483303,"Record function id": 0, "Sequence number": 24740636, "Fwd thread id": 1, "Ev Idx": 102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SelectBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751062.842, "dur": 49.075,
    "args": {
      "External id": 2483304,"Sequence number": 24740636, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 103
    }
  },
  {
    "ph": "f", "id": 13, "pid": 1336759, "tid": 1381173, "ts": 1514656751062.842,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751067.217, "dur": 44.414,
    "args": {
      "External id": 2483305,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 2048]", "2", "2"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 2048], [], [], []], "Ev Idx": 104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751069.012, "dur": 19.710,
    "args": {
      "External id": 2483306,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 2048]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751070.082, "dur": 2.927,
    "args": {
      "External id": 2483307,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 2048]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751073.695, "dur": 14.750,
    "args": {
      "External id": 2483308,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[33554432, 8192, 2048, 1]], "Input Dims": [[16, 4096, 4, 2048]], "Ev Idx": 107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751074.679, "dur": 13.323,
    "args": {
      "External id": 2483309,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], []], "Input Dims": [[16, 4096, 4, 2048], []], "Ev Idx": 108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751089.834, "dur": 6.923,
    "args": {
      "External id": 2483310,"Record function id": 0, "Concrete Inputs": ["", "2", "2"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], []], "Input Dims": [[16, 4096, 4, 2048], [], []], "Ev Idx": 109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751095.512, "dur": 0.850,
    "args": {
      "External id": 2483311,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]", "[33554432, 8192, 1]", "4096"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], []], "Ev Idx": 110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751097.424, "dur": 13.677,
    "args": {
      "External id": 2483312,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SliceBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751118.773, "dur": 90.306,
    "args": {
      "External id": 2483313,"Record function id": 0, "Sequence number": 24740635, "Fwd thread id": 1, "Ev Idx": 112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SliceBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751122.178, "dur": 83.759,
    "args": {
      "External id": 2483314,"Sequence number": 24740635, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[33554432, 8192, 2048, 1]], "Input Dims": [[16, 4096, 4, 2048]], "Ev Idx": 113
    }
  },
  {
    "ph": "f", "id": 14, "pid": 1336759, "tid": 1381173, "ts": 1514656751122.178,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751123.609, "dur": 82.033,
    "args": {
      "External id": 2483315,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 2048]", "1", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], [], [], []], "Ev Idx": 114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751124.721, "dur": 15.219,
    "args": {
      "External id": 2483316,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 2048]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751125.847, "dur": 2.017,
    "args": {
      "External id": 2483317,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 2048]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751128.333, "dur": 11.331,
    "args": {
      "External id": 2483318,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[33554432, 8192, 2048, 1]], "Input Dims": [[16, 4096, 4, 2048]], "Ev Idx": 117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751128.921, "dur": 10.385,
    "args": {
      "External id": 2483319,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], []], "Input Dims": [[16, 4096, 4, 2048], []], "Ev Idx": 118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751140.876, "dur": 4.539,
    "args": {
      "External id": 2483320,"Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], [], []], "Ev Idx": 119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751144.343, "dur": 0.919,
    "args": {
      "External id": 2483321,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 2048]", "[33554432, 8192, 2048, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], []], "Ev Idx": 120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751145.998, "dur": 58.956,
    "args": {
      "External id": 2483322,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [33554432, 8192, 2048, 1], []], "Input Dims": [[16, 4096, 4, 2048], [16, 4096, 4, 2048], []], "Ev Idx": 121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SliceBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751213.312, "dur": 90.722,
    "args": {
      "External id": 2483323,"Record function id": 0, "Sequence number": 24740634, "Fwd thread id": 1, "Ev Idx": 122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SliceBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751216.596, "dur": 69.598,
    "args": {
      "External id": 2483324,"Sequence number": 24740634, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[33554432, 8192, 2048, 1]], "Input Dims": [[16, 4096, 4, 2048]], "Ev Idx": 123
    }
  },
  {
    "ph": "f", "id": 15, "pid": 1336759, "tid": 1381173, "ts": 1514656751216.596,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751220.800, "dur": 65.109,
    "args": {
      "External id": 2483325,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 2048]", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], [], [], []], "Ev Idx": 124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751221.686, "dur": 14.689,
    "args": {
      "External id": 2483326,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 2048]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751222.285, "dur": 1.890,
    "args": {
      "External id": 2483327,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 2048]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751224.923, "dur": 11.201,
    "args": {
      "External id": 2483328,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[33554432, 8192, 2048, 1]], "Input Dims": [[16, 4096, 4, 2048]], "Ev Idx": 127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751225.823, "dur": 9.844,
    "args": {
      "External id": 2483329,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], []], "Input Dims": [[16, 4096, 4, 2048], []], "Ev Idx": 128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751237.191, "dur": 1.926,
    "args": {
      "External id": 2483330,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], [], []], "Ev Idx": 129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751238.321, "dur": 0.683,
    "args": {
      "External id": 2483331,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 2048]", "[33554432, 8192, 2048, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], []], "Ev Idx": 130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751239.632, "dur": 45.544,
    "args": {
      "External id": 2483332,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [33554432, 8192, 2048, 1], []], "Input Dims": [[16, 4096, 4, 2048], [16, 4096, 4, 2048], []], "Ev Idx": 131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751290.083, "dur": 12.669,
    "args": {
      "External id": 2483333,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [33554432, 8192, 2048, 1], []], "Input Dims": [[16, 4096, 4, 2048], [16, 4096, 4, 2048], []], "Ev Idx": 132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: AddBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751308.091, "dur": 34.227,
    "args": {
      "External id": 2483334,"Record function id": 0, "Sequence number": 24740633, "Fwd thread id": 1, "Ev Idx": 133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AddBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751311.663, "dur": 0.658,
    "args": {
      "External id": 2483335,"Sequence number": 24740633, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 134
    }
  },
  {
    "ph": "f", "id": 16, "pid": 1336759, "tid": 1381173, "ts": 1514656751311.663,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751314.313, "dur": 25.479,
    "args": {
      "External id": 2483336,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[1], []], "Input Dims": [[1], []], "Ev Idx": 135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751315.961, "dur": 23.462,
    "args": {
      "External id": 2483337,"Record function id": 0, "Concrete Inputs": ["", "[]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[1], [], [], []], "Ev Idx": 136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751322.138, "dur": 3.007,
    "args": {
      "External id": 2483338,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: FusedLinearCrossEntropyFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751346.956, "dur": 3089.693,
    "args": {
      "External id": 2483339,"Record function id": 0, "Sequence number": 24740631, "Fwd thread id": 1, "Ev Idx": 138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FusedLinearCrossEntropyFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751350.569, "dur": 3061.890,
    "args": {
      "External id": 2483340,"Sequence number": 24740631, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 139
    }
  },
  {
    "ph": "f", "id": 17, "pid": 1336759, "tid": 1381173, "ts": 1514656751350.569,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751379.887, "dur": 2.564,
    "args": {
      "External id": 2483341,"Record function id": 0, "Concrete Inputs": ["[]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751384.455, "dur": 2945.793,
    "args": {
      "External id": 2483342,"Record function id": 0, "Concrete Inputs": ["", "", "6", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751385.615, "dur": 2944.354,
    "args": {
      "External id": 2483343,"Record function id": 0, "Concrete Inputs": ["", "6", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], []], "Ev Idx": 142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751387.674, "dur": 3.343,
    "args": {
      "External id": 2483344,"Record function id": 0, "Concrete Inputs": ["[]", "[]", "6", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656751391.818, "dur": 2937.206,
    "args": {
      "External id": 2483345,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754333.265, "dur": 0.391,
    "args": {
      "External id": 2483346,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754334.650, "dur": 2.007,
    "args": {
      "External id": 2483347,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754335.848, "dur": 0.695,
    "args": {
      "External id": 2483348,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::ne", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754339.823, "dur": 20.072,
    "args": {
      "External id": 2483349,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::is_nonzero", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754368.965, "dur": 36.522,
    "args": {
      "External id": 2483350,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754370.083, "dur": 35.186,
    "args": {
      "External id": 2483351,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754371.050, "dur": 33.803,
    "args": {
      "External id": 2483352,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754419.818, "dur": 13.802,
    "args": {
      "External id": 2483353,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754446.630, "dur": 9.454,
    "args": {
      "External id": 2483354,"Record function id": 0, "Sequence number": 24740630, "Fwd thread id": 1, "Ev Idx": 153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754448.014, "dur": 6.473,
    "args": {
      "External id": 2483355,"Sequence number": 24740630, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 154
    }
  },
  {
    "ph": "f", "id": 18, "pid": 1336759, "tid": 1381173, "ts": 1514656754448.014,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754450.165, "dur": 4.133,
    "args": {
      "External id": 2483356,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754451.258, "dur": 2.898,
    "args": {
      "External id": 2483357,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SliceBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754459.512, "dur": 60.043,
    "args": {
      "External id": 2483358,"Record function id": 0, "Sequence number": 24740629, "Fwd thread id": 1, "Ev Idx": 157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SliceBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754460.509, "dur": 54.996,
    "args": {
      "External id": 2483359,"Sequence number": 24740629, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 158
    }
  },
  {
    "ph": "f", "id": 19, "pid": 1336759, "tid": 1381173, "ts": 1514656754460.509,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754463.657, "dur": 51.333,
    "args": {
      "External id": 2483360,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]", "2", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 2048], [], [], [], [], []], "Ev Idx": 159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754468.531, "dur": 18.494,
    "args": {
      "External id": 2483361,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 2048]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754469.900, "dur": 2.924,
    "args": {
      "External id": 2483362,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 2048]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754473.706, "dur": 13.076,
    "args": {
      "External id": 2483363,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754474.778, "dur": 11.451,
    "args": {
      "External id": 2483364,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754488.170, "dur": 4.775,
    "args": {
      "External id": 2483365,"Record function id": 0, "Concrete Inputs": ["", "2", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 1], [], [], [], []], "Input Dims": [[16, 4096, 2048], [], [], [], []], "Ev Idx": 164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754491.647, "dur": 0.993,
    "args": {
      "External id": 2483366,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]", "[8388608, 2048, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8388608, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 2048], [], [], []], "Ev Idx": 165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754493.689, "dur": 20.621,
    "args": {
      "External id": 2483367,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SelectBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754523.461, "dur": 69.936,
    "args": {
      "External id": 2483368,"Record function id": 0, "Sequence number": 24740628, "Fwd thread id": 1, "Ev Idx": 167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SelectBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754524.614, "dur": 65.695,
    "args": {
      "External id": 2483369,"Sequence number": 24740628, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 168
    }
  },
  {
    "ph": "f", "id": 20, "pid": 1336759, "tid": 1381173, "ts": 1514656754524.614,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754526.408, "dur": 63.618,
    "args": {
      "External id": 2483370,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 2048]", "2", "1"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 2048], [], [], []], "Ev Idx": 169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754540.331, "dur": 24.106,
    "args": {
      "External id": 2483371,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 2048]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754544.111, "dur": 3.543,
    "args": {
      "External id": 2483372,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 2048]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754548.879, "dur": 15.312,
    "args": {
      "External id": 2483373,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[33554432, 8192, 2048, 1]], "Input Dims": [[16, 4096, 4, 2048]], "Ev Idx": 172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754549.782, "dur": 13.977,
    "args": {
      "External id": 2483374,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], []], "Input Dims": [[16, 4096, 4, 2048], []], "Ev Idx": 173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754565.591, "dur": 9.400,
    "args": {
      "External id": 2483375,"Record function id": 0, "Concrete Inputs": ["", "2", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], []], "Input Dims": [[16, 4096, 4, 2048], [], []], "Ev Idx": 174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754571.429, "dur": 3.100,
    "args": {
      "External id": 2483376,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]", "[33554432, 8192, 1]", "2048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], []], "Ev Idx": 175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754575.531, "dur": 13.853,
    "args": {
      "External id": 2483377,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SliceBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754598.368, "dur": 90.883,
    "args": {
      "External id": 2483378,"Record function id": 0, "Sequence number": 24740627, "Fwd thread id": 1, "Ev Idx": 177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SliceBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754599.944, "dur": 86.975,
    "args": {
      "External id": 2483379,"Sequence number": 24740627, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[33554432, 8192, 2048, 1]], "Input Dims": [[16, 4096, 4, 2048]], "Ev Idx": 178
    }
  },
  {
    "ph": "f", "id": 21, "pid": 1336759, "tid": 1381173, "ts": 1514656754599.944,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754601.121, "dur": 85.512,
    "args": {
      "External id": 2483380,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 2048]", "1", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], [], [], []], "Ev Idx": 179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754602.262, "dur": 20.462,
    "args": {
      "External id": 2483381,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 2048]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754603.062, "dur": 2.004,
    "args": {
      "External id": 2483382,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 2048]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754607.924, "dur": 14.538,
    "args": {
      "External id": 2483383,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[33554432, 8192, 2048, 1]], "Input Dims": [[16, 4096, 4, 2048]], "Ev Idx": 182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754608.346, "dur": 13.738,
    "args": {
      "External id": 2483384,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], []], "Input Dims": [[16, 4096, 4, 2048], []], "Ev Idx": 183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754623.539, "dur": 4.509,
    "args": {
      "External id": 2483385,"Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], [], []], "Ev Idx": 184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754627.037, "dur": 0.746,
    "args": {
      "External id": 2483386,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 2048]", "[33554432, 8192, 2048, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], []], "Ev Idx": 185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754628.779, "dur": 57.180,
    "args": {
      "External id": 2483387,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [33554432, 8192, 2048, 1], []], "Input Dims": [[16, 4096, 4, 2048], [16, 4096, 4, 2048], []], "Ev Idx": 186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SliceBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754693.170, "dur": 87.325,
    "args": {
      "External id": 2483388,"Record function id": 0, "Sequence number": 24740626, "Fwd thread id": 1, "Ev Idx": 187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SliceBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754694.147, "dur": 70.761,
    "args": {
      "External id": 2483389,"Sequence number": 24740626, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[33554432, 8192, 2048, 1]], "Input Dims": [[16, 4096, 4, 2048]], "Ev Idx": 188
    }
  },
  {
    "ph": "f", "id": 22, "pid": 1336759, "tid": 1381173, "ts": 1514656754694.147,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754697.919, "dur": 66.766,
    "args": {
      "External id": 2483390,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 2048]", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], [], [], []], "Ev Idx": 189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754698.613, "dur": 15.334,
    "args": {
      "External id": 2483391,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 2048]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754699.211, "dur": 1.752,
    "args": {
      "External id": 2483392,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 2048]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754701.486, "dur": 12.187,
    "args": {
      "External id": 2483393,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[33554432, 8192, 2048, 1]], "Input Dims": [[16, 4096, 4, 2048]], "Ev Idx": 192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754704.509, "dur": 8.842,
    "args": {
      "External id": 2483394,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], []], "Input Dims": [[16, 4096, 4, 2048], []], "Ev Idx": 193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754714.764, "dur": 3.947,
    "args": {
      "External id": 2483395,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], [], []], "Ev Idx": 194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754717.901, "dur": 0.644,
    "args": {
      "External id": 2483396,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 2048]", "[33554432, 8192, 2048, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], []], "Ev Idx": 195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754719.466, "dur": 44.644,
    "args": {
      "External id": 2483397,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [33554432, 8192, 2048, 1], []], "Input Dims": [[16, 4096, 4, 2048], [16, 4096, 4, 2048], []], "Ev Idx": 196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754768.333, "dur": 11.031,
    "args": {
      "External id": 2483398,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [33554432, 8192, 2048, 1], []], "Input Dims": [[16, 4096, 4, 2048], [16, 4096, 4, 2048], []], "Ev Idx": 197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: AddBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754784.420, "dur": 28.053,
    "args": {
      "External id": 2483399,"Record function id": 0, "Sequence number": 24740625, "Fwd thread id": 1, "Ev Idx": 198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AddBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754785.851, "dur": 1.023,
    "args": {
      "External id": 2483400,"Sequence number": 24740625, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 199
    }
  },
  {
    "ph": "f", "id": 23, "pid": 1336759, "tid": 1381173, "ts": 1514656754785.851,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754788.292, "dur": 19.565,
    "args": {
      "External id": 2483401,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[1], []], "Input Dims": [[1], []], "Ev Idx": 200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754789.812, "dur": 17.711,
    "args": {
      "External id": 2483402,"Record function id": 0, "Concrete Inputs": ["", "[]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[1], [], [], []], "Ev Idx": 201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754793.654, "dur": 0.431,
    "args": {
      "External id": 2483403,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: FusedLinearCrossEntropyFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754816.314, "dur": 3136.904,
    "args": {
      "External id": 2483404,"Record function id": 0, "Sequence number": 24740624, "Fwd thread id": 1, "Ev Idx": 203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FusedLinearCrossEntropyFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754827.274, "dur": 3096.031,
    "args": {
      "External id": 2483405,"Sequence number": 24740624, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 204
    }
  },
  {
    "ph": "f", "id": 24, "pid": 1336759, "tid": 1381173, "ts": 1514656754827.274,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754854.278, "dur": 2.793,
    "args": {
      "External id": 2483406,"Record function id": 0, "Concrete Inputs": ["[]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754859.232, "dur": 2978.171,
    "args": {
      "External id": 2483407,"Record function id": 0, "Concrete Inputs": ["", "", "6", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754860.403, "dur": 2976.682,
    "args": {
      "External id": 2483408,"Record function id": 0, "Concrete Inputs": ["", "6", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], []], "Ev Idx": 207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754864.609, "dur": 2.864,
    "args": {
      "External id": 2483409,"Record function id": 0, "Concrete Inputs": ["[]", "[]", "6", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656754868.080, "dur": 2967.534,
    "args": {
      "External id": 2483410,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::lift_fresh", "pid": 1336759, "tid": 1381173,
    "ts": 1514656757840.896, "dur": 0.529,
    "args": {
      "External id": 2483411,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656757842.596, "dur": 2.270,
    "args": {
      "External id": 2483412,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656757843.958, "dur": 0.795,
    "args": {
      "External id": 2483413,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::ne", "pid": 1336759, "tid": 1381173,
    "ts": 1514656757848.635, "dur": 22.653,
    "args": {
      "External id": 2483414,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::is_nonzero", "pid": 1336759, "tid": 1381173,
    "ts": 1514656757875.703, "dur": 41.093,
    "args": {
      "External id": 2483415,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1336759, "tid": 1381173,
    "ts": 1514656757876.563, "dur": 40.063,
    "args": {
      "External id": 2483416,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1336759, "tid": 1381173,
    "ts": 1514656757878.272, "dur": 38.013,
    "args": {
      "External id": 2483417,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656757934.423, "dur": 14.316,
    "args": {
      "External id": 2483418,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656757966.225, "dur": 11.345,
    "args": {
      "External id": 2483419,"Record function id": 0, "Ev Idx": 218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656757968.435, "dur": 7.581,
    "args": {
      "External id": 2483420,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656757971.312, "dur": 3.797,
    "args": {
      "External id": 2483421,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656757972.078, "dur": 2.906,
    "args": {
      "External id": 2483422,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758004.037, "dur": 14.428,
    "args": {
      "External id": 2483423,"Record function id": 0, "Sequence number": 24740623, "Fwd thread id": 1, "Ev Idx": 222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758006.063, "dur": 9.886,
    "args": {
      "External id": 2483424,"Sequence number": 24740623, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 223
    }
  },
  {
    "ph": "f", "id": 25, "pid": 1336759, "tid": 1381173, "ts": 1514656758006.063,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758008.769, "dur": 6.982,
    "args": {
      "External id": 2483425,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758012.448, "dur": 3.030,
    "args": {
      "External id": 2483426,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SliceBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758023.489, "dur": 71.699,
    "args": {
      "External id": 2483427,"Record function id": 0, "Sequence number": 24740622, "Fwd thread id": 1, "Ev Idx": 226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SliceBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758024.438, "dur": 66.174,
    "args": {
      "External id": 2483428,"Sequence number": 24740622, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 227
    }
  },
  {
    "ph": "f", "id": 26, "pid": 1336759, "tid": 1381173, "ts": 1514656758024.438,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758026.304, "dur": 63.935,
    "args": {
      "External id": 2483429,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]", "2", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 2048], [], [], [], [], []], "Ev Idx": 228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758031.124, "dur": 25.014,
    "args": {
      "External id": 2483430,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 2048]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758032.448, "dur": 3.120,
    "args": {
      "External id": 2483431,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 2048]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758036.372, "dur": 19.480,
    "args": {
      "External id": 2483432,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758038.133, "dur": 17.198,
    "args": {
      "External id": 2483433,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758057.517, "dur": 5.838,
    "args": {
      "External id": 2483434,"Record function id": 0, "Concrete Inputs": ["", "2", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 1], [], [], [], []], "Input Dims": [[16, 4096, 2048], [], [], [], []], "Ev Idx": 233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758062.098, "dur": 0.955,
    "args": {
      "External id": 2483435,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]", "[8388608, 2048, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8388608, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 2048], [], [], []], "Ev Idx": 234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758064.398, "dur": 24.948,
    "args": {
      "External id": 2483436,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SelectBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758101.762, "dur": 50.186,
    "args": {
      "External id": 2483437,"Record function id": 0, "Sequence number": 24740621, "Fwd thread id": 1, "Ev Idx": 236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SelectBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758102.624, "dur": 46.659,
    "args": {
      "External id": 2483438,"Sequence number": 24740621, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 237
    }
  },
  {
    "ph": "f", "id": 27, "pid": 1336759, "tid": 1381173, "ts": 1514656758102.624,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758104.235, "dur": 44.777,
    "args": {
      "External id": 2483439,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 2048]", "2", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 2048], [], [], []], "Ev Idx": 238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758105.485, "dur": 23.236,
    "args": {
      "External id": 2483440,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 2048]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758112.544, "dur": 2.527,
    "args": {
      "External id": 2483441,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 2048]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758115.707, "dur": 12.746,
    "args": {
      "External id": 2483442,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[33554432, 8192, 2048, 1]], "Input Dims": [[16, 4096, 4, 2048]], "Ev Idx": 241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758116.700, "dur": 11.375,
    "args": {
      "External id": 2483443,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], []], "Input Dims": [[16, 4096, 4, 2048], []], "Ev Idx": 242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758129.813, "dur": 4.514,
    "args": {
      "External id": 2483444,"Record function id": 0, "Concrete Inputs": ["", "2", "0"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], []], "Input Dims": [[16, 4096, 4, 2048], [], []], "Ev Idx": 243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758132.828, "dur": 1.049,
    "args": {
      "External id": 2483445,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]", "[33554432, 8192, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], []], "Ev Idx": 244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758134.980, "dur": 13.379,
    "args": {
      "External id": 2483446,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SliceBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758155.802, "dur": 91.216,
    "args": {
      "External id": 2483447,"Record function id": 0, "Sequence number": 24740620, "Fwd thread id": 1, "Ev Idx": 246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SliceBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758156.768, "dur": 87.549,
    "args": {
      "External id": 2483448,"Sequence number": 24740620, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[33554432, 8192, 2048, 1]], "Input Dims": [[16, 4096, 4, 2048]], "Ev Idx": 247
    }
  },
  {
    "ph": "f", "id": 28, "pid": 1336759, "tid": 1381173, "ts": 1514656758156.768,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758158.266, "dur": 85.717,
    "args": {
      "External id": 2483449,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 2048]", "1", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], [], [], []], "Ev Idx": 248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758161.538, "dur": 18.228,
    "args": {
      "External id": 2483450,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 2048]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758162.230, "dur": 2.136,
    "args": {
      "External id": 2483451,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 2048]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758167.595, "dur": 11.915,
    "args": {
      "External id": 2483452,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[33554432, 8192, 2048, 1]], "Input Dims": [[16, 4096, 4, 2048]], "Ev Idx": 251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758168.370, "dur": 10.770,
    "args": {
      "External id": 2483453,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], []], "Input Dims": [[16, 4096, 4, 2048], []], "Ev Idx": 252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758180.712, "dur": 4.094,
    "args": {
      "External id": 2483454,"Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], [], []], "Ev Idx": 253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758181.946, "dur": 2.685,
    "args": {
      "External id": 2483455,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 2048]", "[33554432, 8192, 2048, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], []], "Ev Idx": 254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758185.521, "dur": 57.671,
    "args": {
      "External id": 2483456,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [33554432, 8192, 2048, 1], []], "Input Dims": [[16, 4096, 4, 2048], [16, 4096, 4, 2048], []], "Ev Idx": 255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SliceBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758250.891, "dur": 97.061,
    "args": {
      "External id": 2483457,"Record function id": 0, "Sequence number": 24740619, "Fwd thread id": 1, "Ev Idx": 256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SliceBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758251.671, "dur": 78.836,
    "args": {
      "External id": 2483458,"Sequence number": 24740619, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[33554432, 8192, 2048, 1]], "Input Dims": [[16, 4096, 4, 2048]], "Ev Idx": 257
    }
  },
  {
    "ph": "f", "id": 29, "pid": 1336759, "tid": 1381173, "ts": 1514656758251.671,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758252.778, "dur": 77.475,
    "args": {
      "External id": 2483459,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 2048]", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], [], [], []], "Ev Idx": 258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758253.863, "dur": 26.387,
    "args": {
      "External id": 2483460,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 2048]", "15", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758258.284, "dur": 2.032,
    "args": {
      "External id": 2483461,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 4, 2048]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758264.566, "dur": 15.423,
    "args": {
      "External id": 2483462,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[33554432, 8192, 2048, 1]], "Input Dims": [[16, 4096, 4, 2048]], "Ev Idx": 261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758269.847, "dur": 9.799,
    "args": {
      "External id": 2483463,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], []], "Input Dims": [[16, 4096, 4, 2048], []], "Ev Idx": 262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758281.112, "dur": 1.905,
    "args": {
      "External id": 2483464,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], [], []], "Ev Idx": 263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758282.190, "dur": 0.695,
    "args": {
      "External id": 2483465,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 2048]", "[33554432, 8192, 2048, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], []], "Ev Idx": 264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758283.647, "dur": 45.749,
    "args": {
      "External id": 2483466,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [33554432, 8192, 2048, 1], []], "Input Dims": [[16, 4096, 4, 2048], [16, 4096, 4, 2048], []], "Ev Idx": 265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758334.067, "dur": 12.073,
    "args": {
      "External id": 2483467,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [33554432, 8192, 2048, 1], []], "Input Dims": [[16, 4096, 4, 2048], [16, 4096, 4, 2048], []], "Ev Idx": 266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758353.095, "dur": 334.853,
    "args": {
      "External id": 2483468,"Record function id": 0, "Sequence number": 24740618, "Fwd thread id": 1, "Ev Idx": 267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758354.494, "dur": 325.096,
    "args": {
      "External id": 2483469,"Sequence number": 24740618, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[33554432, 8192, 2048, 1]], "Input Dims": [[16, 4096, 4, 2048]], "Ev Idx": 268
    }
  },
  {
    "ph": "f", "id": 30, "pid": 1336759, "tid": 1381173, "ts": 1514656758354.494,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758506.774, "dur": 52.880,
    "args": {
      "External id": 2483470,"kernel_hash": "cntjjtfhvxdzpnscrpib2i3ct2hk6s7qd3lqykazlircwdkp676d", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "262144", "2048", "1", "1986", "132", "True", "2048", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/nt/cntjjtfhvxdzpnscrpib2i3ct2hk6s7qd3lqykazlircwdkp676d.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[262144, 2048], [2048], [262144, 2048], [262144, 2048], [132, 2048], [262144], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758599.518, "dur": 26.368,
    "args": {
      "External id": 2483471,"kernel_hash": "cofxh4a56o46ifoz3xsku2krq37ug6i5slnssef2z53eyj4ipzd2", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/of/cofxh4a56o46ifoz3xsku2krq37ug6i5slnssef2z53eyj4ipzd2.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758642.249, "dur": 15.612,
    "args": {
      "External id": 2483472,"kernel_hash": "ckjxjmz2vylbnrnans6gouo3or7yw4dvfka55hu6q4kyfb7xeari", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/kj/ckjxjmz2vylbnrnans6gouo3or7yw4dvfka55hu6q4kyfb7xeari.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758698.560, "dur": 13.813,
    "args": {
      "External id": 2483473,"Record function id": 0, "Ev Idx": 272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758700.655, "dur": 10.805,
    "args": {
      "External id": 2483474,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758703.933, "dur": 6.591,
    "args": {
      "External id": 2483475,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758707.311, "dur": 3.073,
    "args": {
      "External id": 2483476,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: StackBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758718.371, "dur": 32.094,
    "args": {
      "External id": 2483477,"Record function id": 0, "Sequence number": 24740617, "Fwd thread id": 1, "Ev Idx": 276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "StackBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758719.422, "dur": 25.316,
    "args": {
      "External id": 2483478,"Sequence number": 24740617, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[33554432, 8192, 2048, 1]], "Input Dims": [[16, 4096, 4, 2048]], "Ev Idx": 277
    }
  },
  {
    "ph": "f", "id": 31, "pid": 1336759, "tid": 1381173, "ts": 1514656758719.422,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758721.561, "dur": 7.637,
    "args": {
      "External id": 2483479,"Record function id": 0, "Concrete Inputs": ["", "-2", "0"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], []], "Input Dims": [[16, 4096, 4, 2048], [], []], "Ev Idx": 278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758726.284, "dur": 1.197,
    "args": {
      "External id": 2483480,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]", "[33554432, 8192, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], []], "Ev Idx": 279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758729.881, "dur": 6.919,
    "args": {
      "External id": 2483481,"Record function id": 0, "Concrete Inputs": ["", "-2", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], []], "Input Dims": [[16, 4096, 4, 2048], [], []], "Ev Idx": 280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758735.657, "dur": 0.366,
    "args": {
      "External id": 2483482,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]", "[33554432, 8192, 1]", "2048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], []], "Ev Idx": 281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758737.272, "dur": 2.162,
    "args": {
      "External id": 2483483,"Record function id": 0, "Concrete Inputs": ["", "-2", "2"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], []], "Input Dims": [[16, 4096, 4, 2048], [], []], "Ev Idx": 282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758738.212, "dur": 0.656,
    "args": {
      "External id": 2483484,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]", "[33554432, 8192, 1]", "4096"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], []], "Ev Idx": 283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758739.857, "dur": 4.175,
    "args": {
      "External id": 2483485,"Record function id": 0, "Concrete Inputs": ["", "-2", "3"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], []], "Input Dims": [[16, 4096, 4, 2048], [], []], "Ev Idx": 284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758742.672, "dur": 0.655,
    "args": {
      "External id": 2483486,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]", "[33554432, 8192, 1]", "6144"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], []], "Ev Idx": 285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: AddBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758754.133, "dur": 5.094,
    "args": {
      "External id": 2483487,"Record function id": 0, "Sequence number": 24740616, "Fwd thread id": 1, "Ev Idx": 286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AddBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758755.056, "dur": 1.003,
    "args": {
      "External id": 2483488,"Sequence number": 24740616, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[33554432, 8192, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 287
    }
  },
  {
    "ph": "f", "id": 32, "pid": 1336759, "tid": 1381173, "ts": 1514656758755.056,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SwiGLULinearFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758762.872, "dur": 545.056,
    "args": {
      "External id": 2483489,"Record function id": 0, "Sequence number": 24740615, "Fwd thread id": 1, "Ev Idx": 288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SwiGLULinearFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758763.952, "dur": 532.290,
    "args": {
      "External id": 2483490,"Sequence number": 24740615, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[33554432, 8192, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 289
    }
  },
  {
    "ph": "f", "id": 33, "pid": 1336759, "tid": 1381173, "ts": 1514656758763.952,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758803.214, "dur": 13.742,
    "args": {
      "External id": 2483491,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[33554432, 8192, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758812.284, "dur": 4.394,
    "args": {
      "External id": 2483492,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]", "[8192, 1]"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList"], "Input Strides": [[33554432, 8192, 1], [], []], "Input Dims": [[16, 4096, 2048], [], []], "Ev Idx": 291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758821.253, "dur": 6.322,
    "args": {
      "External id": 2483493,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758823.903, "dur": 2.553,
    "args": {
      "External id": 2483494,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[5632, 1], [], []], "Input Dims": [[2048, 5632], [], []], "Ev Idx": 293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758825.556, "dur": 0.673,
    "args": {
      "External id": 2483495,"Record function id": 0, "Concrete Inputs": ["", "[5632, 2048]", "[1, 5632]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[5632, 1], [], [], []], "Input Dims": [[2048, 5632], [], [], []], "Ev Idx": 294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758833.560, "dur": 114.379,
    "args": {
      "External id": 2483496,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[8192, 1], [1, 5632], []], "Input Dims": [[65536, 2048], [5632, 2048], []], "Ev Idx": 295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758834.365, "dur": 6.602,
    "args": {
      "External id": 2483497,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 5632]], "Input Dims": [[5632, 2048]], "Ev Idx": 296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758835.017, "dur": 5.250,
    "args": {
      "External id": 2483498,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 5632], [], []], "Input Dims": [[5632, 2048], [], []], "Ev Idx": 297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758837.766, "dur": 2.431,
    "args": {
      "External id": 2483499,"Record function id": 0, "Concrete Inputs": ["", "[2048, 5632]", "[5632, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 5632], [], [], []], "Input Dims": [[5632, 2048], [], [], []], "Ev Idx": 298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758842.280, "dur": 104.954,
    "args": {
      "External id": 2483500,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8192, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632]], "Ev Idx": 299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758844.382, "dur": 101.334,
    "args": {
      "External id": 2483501,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8192, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632]], "Ev Idx": 300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758952.771, "dur": 7.032,
    "args": {
      "External id": 2483502,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [23068672, 5632, 1]], "Input Dims": [[65536, 5632], [16, 4096, 5632]], "Ev Idx": 301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656758957.019, "dur": 2.612,
    "args": {
      "External id": 2483503,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 1], []], "Input Dims": [[65536, 5632], []], "Ev Idx": 302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759038.659, "dur": 6.503,
    "args": {
      "External id": 2483504,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 5632]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759046.659, "dur": 3.381,
    "args": {
      "External id": 2483505,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 5632]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759053.120, "dur": 1.902,
    "args": {
      "External id": 2483506,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 5632]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759095.453, "dur": 3.677,
    "args": {
      "External id": 2483507,"Record function id": 0, "Concrete Inputs": ["", "[-1, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759096.585, "dur": 2.404,
    "args": {
      "External id": 2483508,"Record function id": 0, "Concrete Inputs": ["", "[-1, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::einsum", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759129.193, "dur": 143.878,
    "args": {
      "External id": 2483509,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["", "TensorList", ""], "Input Strides": [[], [[8192, 1], [5632, 1]], []], "Input Dims": [[], [[65536, 2048], [65536, 5632]], []], "Ev Idx": 308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759135.401, "dur": 8.671,
    "args": {
      "External id": 2483510,"Record function id": 0, "Concrete Inputs": ["", "2"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759141.408, "dur": 0.773,
    "args": {
      "External id": 2483511,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048, 1]", "[8192, 1, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[8192, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759145.890, "dur": 8.964,
    "args": {
      "External id": 2483512,"Record function id": 0, "Concrete Inputs": ["", "[1, 2, 0]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8192, 1, 1], []], "Input Dims": [[65536, 2048, 1], []], "Ev Idx": 311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759151.333, "dur": 2.784,
    "args": {
      "External id": 2483513,"Record function id": 0, "Concrete Inputs": ["", "[2048, 1, 65536]", "[1, 1, 8192]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[65536, 2048, 1], [], [], []], "Ev Idx": 312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759156.880, "dur": 3.970,
    "args": {
      "External id": 2483514,"Record function id": 0, "Concrete Inputs": ["", "2"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], []], "Input Dims": [[65536, 5632], []], "Ev Idx": 313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759159.973, "dur": 0.508,
    "args": {
      "External id": 2483515,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632, 1]", "[5632, 1, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[5632, 1], [], [], []], "Input Dims": [[65536, 5632], [], [], []], "Ev Idx": 314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759161.390, "dur": 5.127,
    "args": {
      "External id": 2483516,"Record function id": 0, "Concrete Inputs": ["", "[2, 1, 0]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 1, 1], []], "Input Dims": [[65536, 5632, 1], []], "Ev Idx": 315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759165.457, "dur": 0.586,
    "args": {
      "External id": 2483517,"Record function id": 0, "Concrete Inputs": ["", "[1, 5632, 65536]", "[1, 1, 5632]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[5632, 1, 1], [], [], []], "Input Dims": [[65536, 5632, 1], [], [], []], "Ev Idx": 316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759171.310, "dur": 4.099,
    "args": {
      "External id": 2483518,"Record function id": 0, "Concrete Inputs": ["", "[0, 2, 1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1, 1, 8192], []], "Input Dims": [[2048, 1, 65536], []], "Ev Idx": 317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759174.688, "dur": 0.431,
    "args": {
      "External id": 2483519,"Record function id": 0, "Concrete Inputs": ["", "[2048, 65536, 1]", "[1, 8192, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1, 8192], [], [], []], "Input Dims": [[2048, 1, 65536], [], [], []], "Ev Idx": 318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759178.428, "dur": 7.424,
    "args": {
      "External id": 2483520,"Record function id": 0, "Concrete Inputs": ["", "[1, 2048, 65536]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1, 8192, 1], []], "Input Dims": [[2048, 65536, 1], []], "Ev Idx": 319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759183.839, "dur": 1.853,
    "args": {
      "External id": 2483521,"Record function id": 0, "Concrete Inputs": ["", "[1, 2048, 65536]", "[2048, 1, 8192]"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList"], "Input Strides": [[1, 8192, 1], [], []], "Input Dims": [[2048, 65536, 1], [], []], "Ev Idx": 320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759186.556, "dur": 3.532,
    "args": {
      "External id": 2483522,"Record function id": 0, "Concrete Inputs": ["", "[2, 1, 0]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1, 1, 5632], []], "Input Dims": [[1, 5632, 65536], []], "Ev Idx": 321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759189.285, "dur": 0.487,
    "args": {
      "External id": 2483523,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632, 1]", "[5632, 1, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1, 5632], [], [], []], "Input Dims": [[1, 5632, 65536], [], [], []], "Ev Idx": 322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759190.657, "dur": 2.353,
    "args": {
      "External id": 2483524,"Record function id": 0, "Concrete Inputs": ["", "[1, 65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 1, 1], []], "Input Dims": [[65536, 5632, 1], []], "Ev Idx": 323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759191.655, "dur": 1.275,
    "args": {
      "External id": 2483525,"Record function id": 0, "Concrete Inputs": ["", "[1, 65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 1, 1], []], "Input Dims": [[65536, 5632, 1], []], "Ev Idx": 324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759194.870, "dur": 65.309,
    "args": {
      "External id": 2483526,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1, 8192], [369098752, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632]], "Ev Idx": 325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759262.019, "dur": 2.896,
    "args": {
      "External id": 2483527,"Record function id": 0, "Concrete Inputs": ["", "[2048, 1, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 5632, 1], []], "Input Dims": [[1, 2048, 5632], []], "Ev Idx": 326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759267.153, "dur": 2.173,
    "args": {
      "External id": 2483528,"Record function id": 0, "Concrete Inputs": ["", "[0, 2, 1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 5632, 1], []], "Input Dims": [[2048, 1, 5632], []], "Ev Idx": 327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759268.437, "dur": 0.573,
    "args": {
      "External id": 2483529,"Record function id": 0, "Concrete Inputs": ["", "[2048, 5632, 1]", "[5632, 1, 5632]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[5632, 5632, 1], [], [], []], "Input Dims": [[2048, 1, 5632], [], [], []], "Ev Idx": 328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759271.214, "dur": 0.787,
    "args": {
      "External id": 2483530,"Record function id": 0, "Concrete Inputs": ["", "[2048, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 1, 5632], []], "Input Dims": [[2048, 5632, 1], []], "Ev Idx": 329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759321.018, "dur": 9.998,
    "args": {
      "External id": 2483531,"Record function id": 0, "Ev Idx": 330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759323.162, "dur": 6.961,
    "args": {
      "External id": 2483532,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759325.810, "dur": 3.251,
    "args": {
      "External id": 2483533,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759326.587, "dur": 2.375,
    "args": {
      "External id": 2483534,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759335.140, "dur": 8.514,
    "args": {
      "External id": 2483535,"Record function id": 0, "Sequence number": 24740614, "Fwd thread id": 1, "Ev Idx": 334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759336.065, "dur": 5.473,
    "args": {
      "External id": 2483536,"Sequence number": 24740614, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[23068672, 5632, 1]], "Input Dims": [[16, 4096, 5632]], "Ev Idx": 335
    }
  },
  {
    "ph": "f", "id": 34, "pid": 1336759, "tid": 1381173, "ts": 1514656759336.065,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759339.563, "dur": 1.769,
    "args": {
      "External id": 2483537,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759340.343, "dur": 0.849,
    "args": {
      "External id": 2483538,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759347.663, "dur": 170.319,
    "args": {
      "External id": 2483539,"Record function id": 0, "Sequence number": 24740613, "Fwd thread id": 1, "Ev Idx": 338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759348.529, "dur": 162.854,
    "args": {
      "External id": 2483540,"Sequence number": 24740613, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[65536, 5632]], "Ev Idx": 339
    }
  },
  {
    "ph": "f", "id": 35, "pid": 1336759, "tid": 1381173, "ts": 1514656759348.529,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759354.044, "dur": 4.088,
    "args": {
      "External id": 2483541,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[65536, 5632]], "Ev Idx": 340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759355.442, "dur": 2.082,
    "args": {
      "External id": 2483542,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[5632, 1], [], []], "Input Dims": [[65536, 5632], [], []], "Ev Idx": 341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759356.701, "dur": 0.612,
    "args": {
      "External id": 2483543,"Record function id": 0, "Concrete Inputs": ["", "[5632, 65536]", "[1, 5632]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[5632, 1], [], [], []], "Input Dims": [[65536, 5632], [], [], []], "Ev Idx": 342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759359.223, "dur": 59.648,
    "args": {
      "External id": 2483544,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048]], "Ev Idx": 343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759420.020, "dur": 6.072,
    "args": {
      "External id": 2483545,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759420.787, "dur": 4.747,
    "args": {
      "External id": 2483546,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[5632, 2048], [], []], "Ev Idx": 345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759424.527, "dur": 0.817,
    "args": {
      "External id": 2483547,"Record function id": 0, "Concrete Inputs": ["", "[2048, 5632]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[5632, 2048], [], [], []], "Ev Idx": 346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759427.435, "dur": 4.779,
    "args": {
      "External id": 2483548,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 5632]], "Ev Idx": 347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759428.544, "dur": 3.214,
    "args": {
      "External id": 2483549,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 5632], [], []], "Ev Idx": 348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759431.182, "dur": 0.477,
    "args": {
      "External id": 2483550,"Record function id": 0, "Concrete Inputs": ["", "[5632, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 5632], [], [], []], "Ev Idx": 349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759432.906, "dur": 77.451,
    "args": {
      "External id": 2483551,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048]], "Ev Idx": 350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759522.475, "dur": 23.375,
    "args": {
      "External id": 2483552,"Record function id": 0, "Sequence number": 24740612, "Fwd thread id": 1, "Ev Idx": 351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759523.630, "dur": 19.714,
    "args": {
      "External id": 2483553,"Sequence number": 24740612, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 352
    }
  },
  {
    "ph": "f", "id": 36, "pid": 1336759, "tid": 1381173, "ts": 1514656759523.630,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759525.198, "dur": 17.959,
    "args": {
      "External id": 2483554,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759525.789, "dur": 16.917,
    "args": {
      "External id": 2483555,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759551.655, "dur": 13.094,
    "args": {
      "External id": 2483556,"Record function id": 0, "Sequence number": 24740611, "Fwd thread id": 1, "Ev Idx": 355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759554.948, "dur": 7.337,
    "args": {
      "External id": 2483557,"Sequence number": 24740611, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 5632]], "Ev Idx": 356
    }
  },
  {
    "ph": "f", "id": 37, "pid": 1336759, "tid": 1381173, "ts": 1514656759554.948,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759556.101, "dur": 5.966,
    "args": {
      "External id": 2483558,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 5632]], "Ev Idx": 357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759557.019, "dur": 4.566,
    "args": {
      "External id": 2483559,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 5632], [], []], "Ev Idx": 358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759560.740, "dur": 0.719,
    "args": {
      "External id": 2483560,"Record function id": 0, "Concrete Inputs": ["", "[5632, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 5632], [], [], []], "Ev Idx": 359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759568.979, "dur": 5.718,
    "args": {
      "External id": 2483561,"Record function id": 0, "Ev Idx": 360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759570.102, "dur": 4.118,
    "args": {
      "External id": 2483562,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759571.763, "dur": 2.155,
    "args": {
      "External id": 2483563,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759572.552, "dur": 1.289,
    "args": {
      "External id": 2483564,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759578.034, "dur": 8.163,
    "args": {
      "External id": 2483565,"Record function id": 0, "Sequence number": 24740610, "Fwd thread id": 1, "Ev Idx": 364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759578.868, "dur": 5.385,
    "args": {
      "External id": 2483566,"Sequence number": 24740610, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[23068672, 5632, 1]], "Input Dims": [[16, 4096, 5632]], "Ev Idx": 365
    }
  },
  {
    "ph": "f", "id": 38, "pid": 1336759, "tid": 1381173, "ts": 1514656759578.868,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759580.053, "dur": 4.048,
    "args": {
      "External id": 2483567,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759583.163, "dur": 0.790,
    "args": {
      "External id": 2483568,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759589.274, "dur": 148.672,
    "args": {
      "External id": 2483569,"Record function id": 0, "Sequence number": 24740609, "Fwd thread id": 1, "Ev Idx": 368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759589.946, "dur": 140.200,
    "args": {
      "External id": 2483570,"Sequence number": 24740609, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[65536, 5632]], "Ev Idx": 369
    }
  },
  {
    "ph": "f", "id": 39, "pid": 1336759, "tid": 1381173, "ts": 1514656759589.946,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759591.823, "dur": 2.604,
    "args": {
      "External id": 2483571,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[65536, 5632]], "Ev Idx": 370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759592.437, "dur": 1.593,
    "args": {
      "External id": 2483572,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[5632, 1], [], []], "Input Dims": [[65536, 5632], [], []], "Ev Idx": 371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759593.367, "dur": 0.558,
    "args": {
      "External id": 2483573,"Record function id": 0, "Concrete Inputs": ["", "[5632, 65536]", "[1, 5632]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[5632, 1], [], [], []], "Input Dims": [[65536, 5632], [], [], []], "Ev Idx": 372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759597.165, "dur": 48.134,
    "args": {
      "External id": 2483574,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048]], "Ev Idx": 373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759646.152, "dur": 4.971,
    "args": {
      "External id": 2483575,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759646.885, "dur": 3.691,
    "args": {
      "External id": 2483576,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[5632, 2048], [], []], "Ev Idx": 375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759649.671, "dur": 0.795,
    "args": {
      "External id": 2483577,"Record function id": 0, "Concrete Inputs": ["", "[2048, 5632]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[5632, 2048], [], [], []], "Ev Idx": 376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759652.219, "dur": 6.317,
    "args": {
      "External id": 2483578,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 5632]], "Ev Idx": 377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759652.902, "dur": 5.240,
    "args": {
      "External id": 2483579,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 5632], [], []], "Ev Idx": 378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759655.701, "dur": 2.361,
    "args": {
      "External id": 2483580,"Record function id": 0, "Concrete Inputs": ["", "[5632, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 5632], [], [], []], "Ev Idx": 379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759659.027, "dur": 70.460,
    "args": {
      "External id": 2483581,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048]], "Ev Idx": 380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759744.150, "dur": 33.328,
    "args": {
      "External id": 2483582,"Record function id": 0, "Sequence number": 24740608, "Fwd thread id": 1, "Ev Idx": 381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759745.198, "dur": 5.835,
    "args": {
      "External id": 2483583,"Sequence number": 24740608, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 382
    }
  },
  {
    "ph": "f", "id": 40, "pid": 1336759, "tid": 1381173, "ts": 1514656759745.198,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759746.610, "dur": 4.281,
    "args": {
      "External id": 2483584,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759749.556, "dur": 1.207,
    "args": {
      "External id": 2483585,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759754.184, "dur": 20.852,
    "args": {
      "External id": 2483586,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759781.093, "dur": 9.771,
    "args": {
      "External id": 2483587,"Record function id": 0, "Sequence number": 24740607, "Fwd thread id": 1, "Ev Idx": 386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759782.079, "dur": 6.746,
    "args": {
      "External id": 2483588,"Sequence number": 24740607, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 5632]], "Ev Idx": 387
    }
  },
  {
    "ph": "f", "id": 41, "pid": 1336759, "tid": 1381173, "ts": 1514656759782.079,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759783.047, "dur": 5.587,
    "args": {
      "External id": 2483589,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 5632]], "Ev Idx": 388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759786.222, "dur": 1.895,
    "args": {
      "External id": 2483590,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 5632], [], []], "Ev Idx": 389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759787.370, "dur": 0.639,
    "args": {
      "External id": 2483591,"Record function id": 0, "Concrete Inputs": ["", "[5632, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 5632], [], [], []], "Ev Idx": 390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759794.789, "dur": 5.230,
    "args": {
      "External id": 2483592,"Record function id": 0, "Ev Idx": 391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759796.473, "dur": 2.995,
    "args": {
      "External id": 2483593,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759797.583, "dur": 1.590,
    "args": {
      "External id": 2483594,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759798.259, "dur": 0.801,
    "args": {
      "External id": 2483595,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: LayerNormFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759804.080, "dur": 455.852,
    "args": {
      "External id": 2483596,"Record function id": 0, "Sequence number": 24740606, "Fwd thread id": 1, "Ev Idx": 395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759805.391, "dur": 440.116,
    "args": {
      "External id": 2483597,"Sequence number": 24740606, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [33554432, 8192, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 396
    }
  },
  {
    "ph": "f", "id": 42, "pid": 1336759, "tid": 1381173, "ts": 1514656759805.391,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759831.624, "dur": 34.015,
    "args": {
      "External id": 2483598,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759833.198, "dur": 32.240,
    "args": {
      "External id": 2483599,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759835.687, "dur": 5.930,
    "args": {
      "External id": 2483600,"Record function id": 0, "Concrete Inputs": ["", "15", "0", "", "", "0"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[33554432, 8192, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 2048], [], [], [], [], []], "Ev Idx": 399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759838.303, "dur": 2.718,
    "args": {
      "External id": 2483601,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 2048]", "15", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759843.070, "dur": 21.922,
    "args": {
      "External id": 2483602,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [33554432, 8192, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759876.448, "dur": 4.190,
    "args": {
      "External id": 2483603,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759879.428, "dur": 1.107,
    "args": {
      "External id": 2483604,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759884.172, "dur": 3.716,
    "args": {
      "External id": 2483605,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759884.911, "dur": 2.871,
    "args": {
      "External id": 2483606,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759901.848, "dur": 1.993,
    "args": {
      "External id": 2483607,"Record function id": 0, "Concrete Inputs": ["[65536, 2048]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656759917.601, "dur": 3.334,
    "args": {
      "External id": 2483608,"Record function id": 0, "Concrete Inputs": ["[132, 2048]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760134.470, "dur": 3.397,
    "args": {
      "External id": 2483609,"Record function id": 0, "Concrete Inputs": ["", "[1, -1, 2048]"], "Input type": ["float", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[132, 2048], []], "Ev Idx": 408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760142.226, "dur": 34.840,
    "args": {
      "External id": 2483610,"Record function id": 0, "Concrete Inputs": ["", "[1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[270336, 2048, 1], [], [], []], "Input Dims": [[1, 132, 2048], [], [], []], "Ev Idx": 409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760153.500, "dur": 1.089,
    "args": {
      "External id": 2483611,"Record function id": 0, "Concrete Inputs": ["", "[1, 1, 2048]", "[2048, 0, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[1, 2048], [], [], []], "Ev Idx": 410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760182.623, "dur": 26.959,
    "args": {
      "External id": 2483612,"Record function id": 0, "Concrete Inputs": ["", "", "15", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[1, 2048], [], [], [], [], []], "Ev Idx": 411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760184.469, "dur": 24.921,
    "args": {
      "External id": 2483613,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], [], []], "Input Dims": [[1, 2048], [], [], [], [], [], []], "Ev Idx": 412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760188.775, "dur": 4.211,
    "args": {
      "External id": 2483614,"Record function id": 0, "Concrete Inputs": ["[1, 2048]", "[2048, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760194.496, "dur": 14.321,
    "args": {
      "External id": 2483615,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "float", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[1, 2048], [1, 2048], []], "Ev Idx": 414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760216.950, "dur": 2.576,
    "args": {
      "External id": 2483616,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1]], "Input Dims": [[1, 2048], [2048]], "Ev Idx": 415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760218.124, "dur": 1.268,
    "args": {
      "External id": 2483617,"Record function id": 0, "Concrete Inputs": ["", "[2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[1, 2048], []], "Ev Idx": 416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760225.751, "dur": 4.871,
    "args": {
      "External id": 2483618,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760228.990, "dur": 1.526,
    "args": {
      "External id": 2483619,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760232.623, "dur": 3.887,
    "args": {
      "External id": 2483620,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760233.453, "dur": 2.980,
    "args": {
      "External id": 2483621,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760273.117, "dur": 8.559,
    "args": {
      "External id": 2483622,"Record function id": 0, "Ev Idx": 421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760274.979, "dur": 6.038,
    "args": {
      "External id": 2483623,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760276.917, "dur": 3.201,
    "args": {
      "External id": 2483624,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760278.048, "dur": 1.977,
    "args": {
      "External id": 2483625,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760285.509, "dur": 9.768,
    "args": {
      "External id": 2483626,"Record function id": 0, "Sequence number": 24740605, "Fwd thread id": 1, "Ev Idx": 425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760286.407, "dur": 5.894,
    "args": {
      "External id": 2483627,"Sequence number": 24740605, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 426
    }
  },
  {
    "ph": "f", "id": 43, "pid": 1336759, "tid": 1381173, "ts": 1514656760286.407,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760289.971, "dur": 2.143,
    "args": {
      "External id": 2483628,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760290.831, "dur": 1.174,
    "args": {
      "External id": 2483629,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760300.910, "dur": 137.093,
    "args": {
      "External id": 2483630,"Record function id": 0, "Sequence number": 24740604, "Fwd thread id": 1, "Ev Idx": 429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760301.878, "dur": 131.600,
    "args": {
      "External id": 2483631,"Sequence number": 24740604, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 430
    }
  },
  {
    "ph": "f", "id": 44, "pid": 1336759, "tid": 1381173, "ts": 1514656760301.878,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760304.414, "dur": 4.864,
    "args": {
      "External id": 2483632,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760306.030, "dur": 2.698,
    "args": {
      "External id": 2483633,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[65536, 2048], [], []], "Ev Idx": 432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760307.668, "dur": 0.864,
    "args": {
      "External id": 2483634,"Record function id": 0, "Concrete Inputs": ["", "[2048, 65536]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760310.530, "dur": 55.981,
    "args": {
      "External id": 2483635,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048]], "Ev Idx": 434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760367.627, "dur": 7.758,
    "args": {
      "External id": 2483636,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760370.492, "dur": 4.280,
    "args": {
      "External id": 2483637,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760371.598, "dur": 2.982,
    "args": {
      "External id": 2483638,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760376.494, "dur": 4.713,
    "args": {
      "External id": 2483639,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760377.220, "dur": 3.610,
    "args": {
      "External id": 2483640,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760380.405, "dur": 0.357,
    "args": {
      "External id": 2483641,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760381.702, "dur": 50.713,
    "args": {
      "External id": 2483642,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760442.649, "dur": 10.782,
    "args": {
      "External id": 2483643,"Record function id": 0, "Sequence number": 24740603, "Fwd thread id": 1, "Ev Idx": 442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760443.602, "dur": 7.840,
    "args": {
      "External id": 2483644,"Sequence number": 24740603, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 443
    }
  },
  {
    "ph": "f", "id": 45, "pid": 1336759, "tid": 1381173, "ts": 1514656760443.602,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760446.596, "dur": 4.693,
    "args": {
      "External id": 2483645,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760447.222, "dur": 3.968,
    "args": {
      "External id": 2483646,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760456.977, "dur": 8.484,
    "args": {
      "External id": 2483647,"Record function id": 0, "Sequence number": 24740602, "Fwd thread id": 1, "Ev Idx": 446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760457.882, "dur": 5.715,
    "args": {
      "External id": 2483648,"Sequence number": 24740602, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 447
    }
  },
  {
    "ph": "f", "id": 46, "pid": 1336759, "tid": 1381173, "ts": 1514656760457.882,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760458.555, "dur": 4.821,
    "args": {
      "External id": 2483649,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760459.337, "dur": 3.536,
    "args": {
      "External id": 2483650,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760462.156, "dur": 0.593,
    "args": {
      "External id": 2483651,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760469.408, "dur": 8.167,
    "args": {
      "External id": 2483652,"Record function id": 0, "Ev Idx": 451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760470.756, "dur": 6.348,
    "args": {
      "External id": 2483653,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760472.036, "dur": 4.857,
    "args": {
      "External id": 2483654,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760475.560, "dur": 1.237,
    "args": {
      "External id": 2483655,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760480.929, "dur": 5.976,
    "args": {
      "External id": 2483656,"Record function id": 0, "Sequence number": 24740601, "Fwd thread id": 1, "Ev Idx": 455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760481.736, "dur": 3.152,
    "args": {
      "External id": 2483657,"Sequence number": 24740601, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 456
    }
  },
  {
    "ph": "f", "id": 47, "pid": 1336759, "tid": 1381173, "ts": 1514656760481.736,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760482.455, "dur": 2.275,
    "args": {
      "External id": 2483658,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760483.368, "dur": 1.236,
    "args": {
      "External id": 2483659,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: FlashAttnFuncBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760490.984, "dur": 325.071,
    "args": {
      "External id": 2483660,"Record function id": 0, "Sequence number": 24740600, "Fwd thread id": 1, "Ev Idx": 459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FlashAttnFuncBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760492.466, "dur": 307.616,
    "args": {
      "External id": 2483661,"Sequence number": 24740600, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 460
    }
  },
  {
    "ph": "f", "id": 48, "pid": 1336759, "tid": 1381173, "ts": 1514656760492.466,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760505.735, "dur": 9.954,
    "args": {
      "External id": 2483662,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760512.114, "dur": 3.026,
    "args": {
      "External id": 2483663,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760517.607, "dur": 2.762,
    "args": {
      "External id": 2483664,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760518.291, "dur": 1.819,
    "args": {
      "External id": 2483665,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760521.488, "dur": 21.928,
    "args": {
      "External id": 2483666,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760524.491, "dur": 18.364,
    "args": {
      "External id": 2483667,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760574.823, "dur": 198.911,
    "args": {
      "External id": 2483668,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760665.015, "dur": 4.668,
    "args": {
      "External id": 2483669,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760671.206, "dur": 4.377,
    "args": {
      "External id": 2483670,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760787.020, "dur": 3.873,
    "args": {
      "External id": 2483671,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760793.947, "dur": 0.804,
    "args": {
      "External id": 2483672,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760796.826, "dur": 0.637,
    "args": {
      "External id": 2483673,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RotaryEmbeddingFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760825.977, "dur": 287.780,
    "args": {
      "External id": 2483674,"Record function id": 0, "Sequence number": 24740599, "Fwd thread id": 1, "Ev Idx": 473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760827.565, "dur": 277.206,
    "args": {
      "External id": 2483675,"Sequence number": 24740599, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 474
    }
  },
  {
    "ph": "f", "id": 49, "pid": 1336759, "tid": 1381173, "ts": 1514656760827.565,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760846.945, "dur": 48.728,
    "args": {
      "External id": 2483676,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760853.432, "dur": 3.289,
    "args": {
      "External id": 2483677,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760858.037, "dur": 36.871,
    "args": {
      "External id": 2483678,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], []], "Ev Idx": 477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760905.328, "dur": 4.001,
    "args": {
      "External id": 2483679,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656760906.988, "dur": 1.998,
    "args": {
      "External id": 2483680,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RotaryEmbeddingFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761123.895, "dur": 160.952,
    "args": {
      "External id": 2483681,"Record function id": 0, "Sequence number": 24740598, "Fwd thread id": 1, "Ev Idx": 480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761125.829, "dur": 153.030,
    "args": {
      "External id": 2483682,"Sequence number": 24740598, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 481
    }
  },
  {
    "ph": "f", "id": 50, "pid": 1336759, "tid": 1381173, "ts": 1514656761125.829,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761138.801, "dur": 36.744,
    "args": {
      "External id": 2483683,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761143.663, "dur": 3.963,
    "args": {
      "External id": 2483684,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761148.957, "dur": 26.087,
    "args": {
      "External id": 2483685,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], []], "Ev Idx": 484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761183.051, "dur": 3.943,
    "args": {
      "External id": 2483686,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761184.283, "dur": 2.282,
    "args": {
      "External id": 2483687,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761292.743, "dur": 14.806,
    "args": {
      "External id": 2483688,"Record function id": 0, "Sequence number": 24740597, "Fwd thread id": 1, "Ev Idx": 487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761294.179, "dur": 10.771,
    "args": {
      "External id": 2483689,"Sequence number": 24740597, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 488
    }
  },
  {
    "ph": "f", "id": 51, "pid": 1336759, "tid": 1381173, "ts": 1514656761294.179,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761296.671, "dur": 8.024,
    "args": {
      "External id": 2483690,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761298.040, "dur": 6.443,
    "args": {
      "External id": 2483691,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761311.114, "dur": 7.760,
    "args": {
      "External id": 2483692,"Record function id": 0, "Sequence number": 24740596, "Fwd thread id": 1, "Ev Idx": 491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761312.141, "dur": 5.282,
    "args": {
      "External id": 2483693,"Sequence number": 24740596, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 492
    }
  },
  {
    "ph": "f", "id": 52, "pid": 1336759, "tid": 1381173, "ts": 1514656761312.141,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761315.340, "dur": 1.944,
    "args": {
      "External id": 2483694,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761316.139, "dur": 0.986,
    "args": {
      "External id": 2483695,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761321.834, "dur": 7.484,
    "args": {
      "External id": 2483696,"Record function id": 0, "Sequence number": 24740595, "Fwd thread id": 1, "Ev Idx": 495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761322.732, "dur": 5.107,
    "args": {
      "External id": 2483697,"Sequence number": 24740595, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 496
    }
  },
  {
    "ph": "f", "id": 53, "pid": 1336759, "tid": 1381173, "ts": 1514656761322.732,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761325.847, "dur": 1.848,
    "args": {
      "External id": 2483698,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761326.583, "dur": 0.980,
    "args": {
      "External id": 2483699,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761334.654, "dur": 5.596,
    "args": {
      "External id": 2483700,"Record function id": 0, "Sequence number": 24740594, "Fwd thread id": 1, "Ev Idx": 499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761335.656, "dur": 2.749,
    "args": {
      "External id": 2483701,"Sequence number": 24740594, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 500
    }
  },
  {
    "ph": "f", "id": 54, "pid": 1336759, "tid": 1381173, "ts": 1514656761335.656,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761336.694, "dur": 1.562,
    "args": {
      "External id": 2483702,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761337.319, "dur": 0.802,
    "args": {
      "External id": 2483703,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761343.625, "dur": 200.955,
    "args": {
      "External id": 2483704,"Record function id": 0, "Sequence number": 24740593, "Fwd thread id": 1, "Ev Idx": 503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761344.493, "dur": 175.767,
    "args": {
      "External id": 2483705,"Sequence number": 24740593, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 504
    }
  },
  {
    "ph": "f", "id": 55, "pid": 1336759, "tid": 1381173, "ts": 1514656761344.493,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761348.269, "dur": 9.663,
    "args": {
      "External id": 2483706,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761353.046, "dur": 4.266,
    "args": {
      "External id": 2483707,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[65536, 2048], [], []], "Ev Idx": 506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761355.193, "dur": 1.822,
    "args": {
      "External id": 2483708,"Record function id": 0, "Concrete Inputs": ["", "[2048, 65536]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761359.727, "dur": 70.599,
    "args": {
      "External id": 2483709,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048]], "Ev Idx": 508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761431.635, "dur": 6.865,
    "args": {
      "External id": 2483710,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761432.405, "dur": 5.384,
    "args": {
      "External id": 2483711,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761436.025, "dur": 1.566,
    "args": {
      "External id": 2483712,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761466.091, "dur": 4.751,
    "args": {
      "External id": 2483713,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761469.386, "dur": 1.056,
    "args": {
      "External id": 2483714,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761469.978, "dur": 0.377,
    "args": {
      "External id": 2483715,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761471.592, "dur": 47.545,
    "args": {
      "External id": 2483716,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761553.906, "dur": 7.830,
    "args": {
      "External id": 2483717,"Record function id": 0, "Sequence number": 24740592, "Fwd thread id": 1, "Ev Idx": 516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761555.345, "dur": 4.840,
    "args": {
      "External id": 2483718,"Sequence number": 24740592, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 517
    }
  },
  {
    "ph": "f", "id": 56, "pid": 1336759, "tid": 1381173, "ts": 1514656761555.345,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761557.022, "dur": 3.003,
    "args": {
      "External id": 2483719,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761557.873, "dur": 2.027,
    "args": {
      "External id": 2483720,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761565.249, "dur": 10.953,
    "args": {
      "External id": 2483721,"Record function id": 0, "Sequence number": 24740591, "Fwd thread id": 1, "Ev Idx": 520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761568.082, "dur": 5.931,
    "args": {
      "External id": 2483722,"Sequence number": 24740591, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 521
    }
  },
  {
    "ph": "f", "id": 57, "pid": 1336759, "tid": 1381173, "ts": 1514656761568.082,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761568.729, "dur": 5.079,
    "args": {
      "External id": 2483723,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761569.485, "dur": 3.767,
    "args": {
      "External id": 2483724,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761572.573, "dur": 0.553,
    "args": {
      "External id": 2483725,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761582.046, "dur": 8.995,
    "args": {
      "External id": 2483726,"Record function id": 0, "Ev Idx": 525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761583.336, "dur": 6.777,
    "args": {
      "External id": 2483727,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761585.917, "dur": 3.762,
    "args": {
      "External id": 2483728,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761586.946, "dur": 2.591,
    "args": {
      "External id": 2483729,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761596.577, "dur": 6.663,
    "args": {
      "External id": 2483730,"Record function id": 0, "Sequence number": 24740590, "Fwd thread id": 1, "Ev Idx": 529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761597.337, "dur": 4.285,
    "args": {
      "External id": 2483731,"Sequence number": 24740590, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 530
    }
  },
  {
    "ph": "f", "id": 58, "pid": 1336759, "tid": 1381173, "ts": 1514656761597.337,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761598.186, "dur": 3.291,
    "args": {
      "External id": 2483732,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761600.560, "dur": 0.817,
    "args": {
      "External id": 2483733,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761606.439, "dur": 96.668,
    "args": {
      "External id": 2483734,"Record function id": 0, "Sequence number": 24740589, "Fwd thread id": 1, "Ev Idx": 533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761607.092, "dur": 90.050,
    "args": {
      "External id": 2483735,"Sequence number": 24740589, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 534
    }
  },
  {
    "ph": "f", "id": 59, "pid": 1336759, "tid": 1381173, "ts": 1514656761607.092,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761609.193, "dur": 4.197,
    "args": {
      "External id": 2483736,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761609.710, "dur": 3.212,
    "args": {
      "External id": 2483737,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[65536, 2048], [], []], "Ev Idx": 536
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761612.360, "dur": 0.441,
    "args": {
      "External id": 2483738,"Record function id": 0, "Concrete Inputs": ["", "[2048, 65536]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761615.180, "dur": 33.971,
    "args": {
      "External id": 2483739,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048]], "Ev Idx": 538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761649.975, "dur": 3.022,
    "args": {
      "External id": 2483740,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761650.701, "dur": 1.803,
    "args": {
      "External id": 2483741,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761651.595, "dur": 0.783,
    "args": {
      "External id": 2483742,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761654.155, "dur": 7.290,
    "args": {
      "External id": 2483743,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761654.918, "dur": 6.073,
    "args": {
      "External id": 2483744,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761658.758, "dur": 2.130,
    "args": {
      "External id": 2483745,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761661.931, "dur": 34.374,
    "args": {
      "External id": 2483746,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761707.359, "dur": 32.374,
    "args": {
      "External id": 2483747,"Record function id": 0, "Sequence number": 24740588, "Fwd thread id": 1, "Ev Idx": 546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761708.272, "dur": 4.854,
    "args": {
      "External id": 2483748,"Sequence number": 24740588, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 547
    }
  },
  {
    "ph": "f", "id": 60, "pid": 1336759, "tid": 1381173, "ts": 1514656761708.272,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761711.362, "dur": 1.623,
    "args": {
      "External id": 2483749,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761711.846, "dur": 1.049,
    "args": {
      "External id": 2483750,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761716.097, "dur": 20.977,
    "args": {
      "External id": 2483751,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761743.573, "dur": 8.740,
    "args": {
      "External id": 2483752,"Record function id": 0, "Sequence number": 24740587, "Fwd thread id": 1, "Ev Idx": 551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761744.443, "dur": 5.975,
    "args": {
      "External id": 2483753,"Sequence number": 24740587, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 552
    }
  },
  {
    "ph": "f", "id": 61, "pid": 1336759, "tid": 1381173, "ts": 1514656761744.443,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761745.363, "dur": 4.849,
    "args": {
      "External id": 2483754,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761746.090, "dur": 3.632,
    "args": {
      "External id": 2483755,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761749.175, "dur": 0.438,
    "args": {
      "External id": 2483756,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761756.080, "dur": 6.855,
    "args": {
      "External id": 2483757,"Record function id": 0, "Ev Idx": 556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761759.219, "dur": 3.188,
    "args": {
      "External id": 2483758,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761760.108, "dur": 1.926,
    "args": {
      "External id": 2483759,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761760.532, "dur": 1.395,
    "args": {
      "External id": 2483760,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761766.140, "dur": 4.985,
    "args": {
      "External id": 2483761,"Record function id": 0, "Sequence number": 24740586, "Fwd thread id": 1, "Ev Idx": 560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761766.949, "dur": 2.443,
    "args": {
      "External id": 2483762,"Sequence number": 24740586, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 561
    }
  },
  {
    "ph": "f", "id": 62, "pid": 1336759, "tid": 1381173, "ts": 1514656761766.949,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761767.705, "dur": 1.531,
    "args": {
      "External id": 2483763,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761768.259, "dur": 0.894,
    "args": {
      "External id": 2483764,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761774.287, "dur": 86.899,
    "args": {
      "External id": 2483765,"Record function id": 0, "Sequence number": 24740585, "Fwd thread id": 1, "Ev Idx": 564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761777.258, "dur": 76.803,
    "args": {
      "External id": 2483766,"Sequence number": 24740585, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 565
    }
  },
  {
    "ph": "f", "id": 63, "pid": 1336759, "tid": 1381173, "ts": 1514656761777.258,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761778.998, "dur": 2.059,
    "args": {
      "External id": 2483767,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761779.405, "dur": 1.232,
    "args": {
      "External id": 2483768,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[65536, 2048], [], []], "Ev Idx": 567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761780.065, "dur": 0.423,
    "args": {
      "External id": 2483769,"Record function id": 0, "Concrete Inputs": ["", "[2048, 65536]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761781.690, "dur": 27.474,
    "args": {
      "External id": 2483770,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048]], "Ev Idx": 569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761810.190, "dur": 7.177,
    "args": {
      "External id": 2483771,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761810.572, "dur": 6.222,
    "args": {
      "External id": 2483772,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761816.163, "dur": 0.509,
    "args": {
      "External id": 2483773,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761818.390, "dur": 2.243,
    "args": {
      "External id": 2483774,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761819.203, "dur": 1.094,
    "args": {
      "External id": 2483775,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761819.802, "dur": 0.427,
    "args": {
      "External id": 2483776,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761821.164, "dur": 32.153,
    "args": {
      "External id": 2483777,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761865.579, "dur": 25.080,
    "args": {
      "External id": 2483778,"Record function id": 0, "Sequence number": 24740584, "Fwd thread id": 1, "Ev Idx": 577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761866.498, "dur": 5.214,
    "args": {
      "External id": 2483779,"Sequence number": 24740584, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 578
    }
  },
  {
    "ph": "f", "id": 64, "pid": 1336759, "tid": 1381173, "ts": 1514656761866.498,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761869.919, "dur": 1.632,
    "args": {
      "External id": 2483780,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761870.313, "dur": 1.145,
    "args": {
      "External id": 2483781,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761873.731, "dur": 14.977,
    "args": {
      "External id": 2483782,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761894.219, "dur": 11.212,
    "args": {
      "External id": 2483783,"Record function id": 0, "Sequence number": 24740583, "Fwd thread id": 1, "Ev Idx": 582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761895.162, "dur": 8.437,
    "args": {
      "External id": 2483784,"Sequence number": 24740583, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 583
    }
  },
  {
    "ph": "f", "id": 65, "pid": 1336759, "tid": 1381173, "ts": 1514656761895.162,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761895.768, "dur": 7.602,
    "args": {
      "External id": 2483785,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761899.069, "dur": 3.830,
    "args": {
      "External id": 2483786,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761902.208, "dur": 0.557,
    "args": {
      "External id": 2483787,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761909.209, "dur": 6.593,
    "args": {
      "External id": 2483788,"Record function id": 0, "Ev Idx": 587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761910.299, "dur": 5.093,
    "args": {
      "External id": 2483789,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761911.169, "dur": 3.917,
    "args": {
      "External id": 2483790,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761913.836, "dur": 1.130,
    "args": {
      "External id": 2483791,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: LayerNormFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761919.360, "dur": 409.964,
    "args": {
      "External id": 2483792,"Record function id": 0, "Sequence number": 24740582, "Fwd thread id": 1, "Ev Idx": 591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761920.664, "dur": 376.571,
    "args": {
      "External id": 2483793,"Sequence number": 24740582, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 592
    }
  },
  {
    "ph": "f", "id": 66, "pid": 1336759, "tid": 1381173, "ts": 1514656761920.664,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761951.822, "dur": 1.789,
    "args": {
      "External id": 2483794,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761952.506, "dur": 0.990,
    "args": {
      "External id": 2483795,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761968.710, "dur": 6.044,
    "args": {
      "External id": 2483796,"Record function id": 0, "Concrete Inputs": ["[65536, 2048]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656761984.068, "dur": 49.298,
    "args": {
      "External id": 2483797,"Record function id": 0, "Concrete Inputs": ["[132, 2048]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762194.766, "dur": 2.797,
    "args": {
      "External id": 2483798,"Record function id": 0, "Concrete Inputs": ["", "[1, -1, 2048]"], "Input type": ["float", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[132, 2048], []], "Ev Idx": 597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762201.950, "dur": 36.403,
    "args": {
      "External id": 2483799,"Record function id": 0, "Concrete Inputs": ["", "[1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[270336, 2048, 1], [], [], []], "Input Dims": [[1, 132, 2048], [], [], []], "Ev Idx": 598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762213.984, "dur": 1.110,
    "args": {
      "External id": 2483800,"Record function id": 0, "Concrete Inputs": ["", "[1, 1, 2048]", "[2048, 0, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[1, 2048], [], [], []], "Ev Idx": 599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762243.888, "dur": 28.724,
    "args": {
      "External id": 2483801,"Record function id": 0, "Concrete Inputs": ["", "", "15", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[1, 2048], [], [], [], [], []], "Ev Idx": 600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762245.650, "dur": 26.676,
    "args": {
      "External id": 2483802,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], [], []], "Input Dims": [[1, 2048], [], [], [], [], [], []], "Ev Idx": 601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762249.595, "dur": 5.092,
    "args": {
      "External id": 2483803,"Record function id": 0, "Concrete Inputs": ["[1, 2048]", "[2048, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762256.172, "dur": 15.654,
    "args": {
      "External id": 2483804,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "float", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[1, 2048], [1, 2048], []], "Ev Idx": 603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762276.973, "dur": 2.309,
    "args": {
      "External id": 2483805,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1]], "Input Dims": [[1, 2048], [2048]], "Ev Idx": 604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762278.128, "dur": 1.066,
    "args": {
      "External id": 2483806,"Record function id": 0, "Concrete Inputs": ["", "[2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[1, 2048], []], "Ev Idx": 605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762286.815, "dur": 4.016,
    "args": {
      "External id": 2483807,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762289.889, "dur": 0.824,
    "args": {
      "External id": 2483808,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762308.441, "dur": 15.450,
    "args": {
      "External id": 2483809,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762341.065, "dur": 8.130,
    "args": {
      "External id": 2483810,"Record function id": 0, "Ev Idx": 609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762343.019, "dur": 5.525,
    "args": {
      "External id": 2483811,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762344.904, "dur": 2.780,
    "args": {
      "External id": 2483812,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762345.891, "dur": 1.703,
    "args": {
      "External id": 2483813,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: AddBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762355.098, "dur": 5.531,
    "args": {
      "External id": 2483814,"Record function id": 0, "Sequence number": 24740581, "Fwd thread id": 1, "Ev Idx": 613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AddBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762356.381, "dur": 1.364,
    "args": {
      "External id": 2483815,"Sequence number": 24740581, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[33554432, 8192, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 614
    }
  },
  {
    "ph": "f", "id": 67, "pid": 1336759, "tid": 1381173, "ts": 1514656762356.381,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SwiGLULinearFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762364.081, "dur": 423.581,
    "args": {
      "External id": 2483816,"Record function id": 0, "Sequence number": 24740580, "Fwd thread id": 1, "Ev Idx": 615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SwiGLULinearFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762365.257, "dur": 412.254,
    "args": {
      "External id": 2483817,"Sequence number": 24740580, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[33554432, 8192, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 616
    }
  },
  {
    "ph": "f", "id": 68, "pid": 1336759, "tid": 1381173, "ts": 1514656762365.257,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762394.556, "dur": 7.603,
    "args": {
      "External id": 2483818,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[33554432, 8192, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762399.081, "dur": 2.810,
    "args": {
      "External id": 2483819,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]", "[8192, 1]"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList"], "Input Strides": [[33554432, 8192, 1], [], []], "Input Dims": [[16, 4096, 2048], [], []], "Ev Idx": 618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762404.894, "dur": 6.762,
    "args": {
      "External id": 2483820,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762406.370, "dur": 4.731,
    "args": {
      "External id": 2483821,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[5632, 1], [], []], "Input Dims": [[2048, 5632], [], []], "Ev Idx": 620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762410.131, "dur": 0.791,
    "args": {
      "External id": 2483822,"Record function id": 0, "Concrete Inputs": ["", "[5632, 2048]", "[1, 5632]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[5632, 1], [], [], []], "Input Dims": [[2048, 5632], [], [], []], "Ev Idx": 621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762417.443, "dur": 81.714,
    "args": {
      "External id": 2483823,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[8192, 1], [1, 5632], []], "Input Dims": [[65536, 2048], [5632, 2048], []], "Ev Idx": 622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762418.327, "dur": 4.509,
    "args": {
      "External id": 2483824,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 5632]], "Input Dims": [[5632, 2048]], "Ev Idx": 623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762418.845, "dur": 3.508,
    "args": {
      "External id": 2483825,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 5632], [], []], "Input Dims": [[5632, 2048], [], []], "Ev Idx": 624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762419.429, "dur": 2.751,
    "args": {
      "External id": 2483826,"Record function id": 0, "Concrete Inputs": ["", "[2048, 5632]", "[5632, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 5632], [], [], []], "Input Dims": [[5632, 2048], [], [], []], "Ev Idx": 625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762423.976, "dur": 74.620,
    "args": {
      "External id": 2483827,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8192, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632]], "Ev Idx": 626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762425.349, "dur": 72.334,
    "args": {
      "External id": 2483828,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8192, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632]], "Ev Idx": 627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762502.732, "dur": 5.135,
    "args": {
      "External id": 2483829,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [23068672, 5632, 1]], "Input Dims": [[65536, 5632], [16, 4096, 5632]], "Ev Idx": 628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762506.019, "dur": 1.692,
    "args": {
      "External id": 2483830,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 1], []], "Input Dims": [[65536, 5632], []], "Ev Idx": 629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762557.920, "dur": 6.253,
    "args": {
      "External id": 2483831,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 5632]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762565.354, "dur": 1.872,
    "args": {
      "External id": 2483832,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 5632]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762567.918, "dur": 1.685,
    "args": {
      "External id": 2483833,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 5632]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762604.013, "dur": 2.474,
    "args": {
      "External id": 2483834,"Record function id": 0, "Concrete Inputs": ["", "[-1, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762604.666, "dur": 1.678,
    "args": {
      "External id": 2483835,"Record function id": 0, "Concrete Inputs": ["", "[-1, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::einsum", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762631.037, "dur": 126.190,
    "args": {
      "External id": 2483836,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["", "TensorList", ""], "Input Strides": [[], [[8192, 1], [5632, 1]], []], "Input Dims": [[], [[65536, 2048], [65536, 5632]], []], "Ev Idx": 635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762636.321, "dur": 4.961,
    "args": {
      "External id": 2483837,"Record function id": 0, "Concrete Inputs": ["", "2"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762639.434, "dur": 0.887,
    "args": {
      "External id": 2483838,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048, 1]", "[8192, 1, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[8192, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762642.672, "dur": 9.767,
    "args": {
      "External id": 2483839,"Record function id": 0, "Concrete Inputs": ["", "[1, 2, 0]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8192, 1, 1], []], "Input Dims": [[65536, 2048, 1], []], "Ev Idx": 638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762649.000, "dur": 2.877,
    "args": {
      "External id": 2483840,"Record function id": 0, "Concrete Inputs": ["", "[2048, 1, 65536]", "[1, 1, 8192]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[65536, 2048, 1], [], [], []], "Ev Idx": 639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762653.853, "dur": 3.563,
    "args": {
      "External id": 2483841,"Record function id": 0, "Concrete Inputs": ["", "2"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], []], "Input Dims": [[65536, 5632], []], "Ev Idx": 640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762656.623, "dur": 0.481,
    "args": {
      "External id": 2483842,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632, 1]", "[5632, 1, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[5632, 1], [], [], []], "Input Dims": [[65536, 5632], [], [], []], "Ev Idx": 641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762658.174, "dur": 4.039,
    "args": {
      "External id": 2483843,"Record function id": 0, "Concrete Inputs": ["", "[2, 1, 0]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 1, 1], []], "Input Dims": [[65536, 5632, 1], []], "Ev Idx": 642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762661.183, "dur": 0.621,
    "args": {
      "External id": 2483844,"Record function id": 0, "Concrete Inputs": ["", "[1, 5632, 65536]", "[1, 1, 5632]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[5632, 1, 1], [], [], []], "Input Dims": [[65536, 5632, 1], [], [], []], "Ev Idx": 643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762665.703, "dur": 2.164,
    "args": {
      "External id": 2483845,"Record function id": 0, "Concrete Inputs": ["", "[0, 2, 1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1, 1, 8192], []], "Input Dims": [[2048, 1, 65536], []], "Ev Idx": 644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762666.933, "dur": 0.638,
    "args": {
      "External id": 2483846,"Record function id": 0, "Concrete Inputs": ["", "[2048, 65536, 1]", "[1, 8192, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1, 8192], [], [], []], "Input Dims": [[2048, 1, 65536], [], [], []], "Ev Idx": 645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762673.341, "dur": 5.158,
    "args": {
      "External id": 2483847,"Record function id": 0, "Concrete Inputs": ["", "[1, 2048, 65536]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1, 8192, 1], []], "Input Dims": [[2048, 65536, 1], []], "Ev Idx": 646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762676.386, "dur": 1.954,
    "args": {
      "External id": 2483848,"Record function id": 0, "Concrete Inputs": ["", "[1, 2048, 65536]", "[2048, 1, 8192]"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList"], "Input Strides": [[1, 8192, 1], [], []], "Input Dims": [[2048, 65536, 1], [], []], "Ev Idx": 647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762681.542, "dur": 1.590,
    "args": {
      "External id": 2483849,"Record function id": 0, "Concrete Inputs": ["", "[2, 1, 0]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1, 1, 5632], []], "Input Dims": [[1, 5632, 65536], []], "Ev Idx": 648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762682.407, "dur": 0.454,
    "args": {
      "External id": 2483850,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632, 1]", "[5632, 1, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1, 5632], [], [], []], "Input Dims": [[1, 5632, 65536], [], [], []], "Ev Idx": 649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762683.634, "dur": 5.754,
    "args": {
      "External id": 2483851,"Record function id": 0, "Concrete Inputs": ["", "[1, 65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 1, 1], []], "Input Dims": [[65536, 5632, 1], []], "Ev Idx": 650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762686.607, "dur": 2.677,
    "args": {
      "External id": 2483852,"Record function id": 0, "Concrete Inputs": ["", "[1, 65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 1, 1], []], "Input Dims": [[65536, 5632, 1], []], "Ev Idx": 651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762690.602, "dur": 52.495,
    "args": {
      "External id": 2483853,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1, 8192], [369098752, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632]], "Ev Idx": 652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762744.864, "dur": 3.449,
    "args": {
      "External id": 2483854,"Record function id": 0, "Concrete Inputs": ["", "[2048, 1, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 5632, 1], []], "Input Dims": [[1, 2048, 5632], []], "Ev Idx": 653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762749.072, "dur": 4.141,
    "args": {
      "External id": 2483855,"Record function id": 0, "Concrete Inputs": ["", "[0, 2, 1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 5632, 1], []], "Input Dims": [[2048, 1, 5632], []], "Ev Idx": 654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762752.131, "dur": 0.531,
    "args": {
      "External id": 2483856,"Record function id": 0, "Concrete Inputs": ["", "[2048, 5632, 1]", "[5632, 1, 5632]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[5632, 5632, 1], [], [], []], "Input Dims": [[2048, 1, 5632], [], [], []], "Ev Idx": 655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762755.252, "dur": 0.974,
    "args": {
      "External id": 2483857,"Record function id": 0, "Concrete Inputs": ["", "[2048, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 1, 5632], []], "Input Dims": [[2048, 5632, 1], []], "Ev Idx": 656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762798.249, "dur": 8.691,
    "args": {
      "External id": 2483858,"Record function id": 0, "Ev Idx": 657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762800.062, "dur": 6.275,
    "args": {
      "External id": 2483859,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762801.773, "dur": 3.781,
    "args": {
      "External id": 2483860,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762803.157, "dur": 2.293,
    "args": {
      "External id": 2483861,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762810.414, "dur": 6.068,
    "args": {
      "External id": 2483862,"Record function id": 0, "Sequence number": 24740579, "Fwd thread id": 1, "Ev Idx": 661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762811.211, "dur": 3.250,
    "args": {
      "External id": 2483863,"Sequence number": 24740579, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[23068672, 5632, 1]], "Input Dims": [[16, 4096, 5632]], "Ev Idx": 662
    }
  },
  {
    "ph": "f", "id": 69, "pid": 1336759, "tid": 1381173, "ts": 1514656762811.211,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762812.827, "dur": 1.414,
    "args": {
      "External id": 2483864,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762813.349, "dur": 0.752,
    "args": {
      "External id": 2483865,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762819.759, "dur": 119.985,
    "args": {
      "External id": 2483866,"Record function id": 0, "Sequence number": 24740578, "Fwd thread id": 1, "Ev Idx": 665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762820.514, "dur": 113.062,
    "args": {
      "External id": 2483867,"Sequence number": 24740578, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[65536, 5632]], "Ev Idx": 666
    }
  },
  {
    "ph": "f", "id": 70, "pid": 1336759, "tid": 1381173, "ts": 1514656762820.514,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762825.870, "dur": 5.403,
    "args": {
      "External id": 2483868,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[65536, 5632]], "Ev Idx": 667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762826.965, "dur": 3.773,
    "args": {
      "External id": 2483869,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[5632, 1], [], []], "Input Dims": [[65536, 5632], [], []], "Ev Idx": 668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762830.057, "dur": 0.536,
    "args": {
      "External id": 2483870,"Record function id": 0, "Concrete Inputs": ["", "[5632, 65536]", "[1, 5632]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[5632, 1], [], [], []], "Input Dims": [[65536, 5632], [], [], []], "Ev Idx": 669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762832.584, "dur": 40.082,
    "args": {
      "External id": 2483871,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048]], "Ev Idx": 670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762873.745, "dur": 5.652,
    "args": {
      "External id": 2483872,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762874.558, "dur": 4.207,
    "args": {
      "External id": 2483873,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[5632, 2048], [], []], "Ev Idx": 672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762877.519, "dur": 1.059,
    "args": {
      "External id": 2483874,"Record function id": 0, "Concrete Inputs": ["", "[2048, 5632]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[5632, 2048], [], [], []], "Ev Idx": 673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762880.911, "dur": 2.555,
    "args": {
      "External id": 2483875,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 5632]], "Ev Idx": 674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762881.702, "dur": 1.354,
    "args": {
      "External id": 2483876,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 5632], [], []], "Ev Idx": 675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762882.465, "dur": 0.525,
    "args": {
      "External id": 2483877,"Record function id": 0, "Concrete Inputs": ["", "[5632, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 5632], [], [], []], "Ev Idx": 676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762883.888, "dur": 48.760,
    "args": {
      "External id": 2483878,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048]], "Ev Idx": 677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762944.348, "dur": 7.808,
    "args": {
      "External id": 2483879,"Record function id": 0, "Sequence number": 24740577, "Fwd thread id": 1, "Ev Idx": 678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762945.157, "dur": 5.856,
    "args": {
      "External id": 2483880,"Sequence number": 24740577, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 679
    }
  },
  {
    "ph": "f", "id": 71, "pid": 1336759, "tid": 1381173, "ts": 1514656762945.157,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762948.687, "dur": 2.177,
    "args": {
      "External id": 2483881,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762949.401, "dur": 1.363,
    "args": {
      "External id": 2483882,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762955.329, "dur": 8.200,
    "args": {
      "External id": 2483883,"Record function id": 0, "Sequence number": 24740576, "Fwd thread id": 1, "Ev Idx": 682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762958.336, "dur": 3.365,
    "args": {
      "External id": 2483884,"Sequence number": 24740576, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 5632]], "Ev Idx": 683
    }
  },
  {
    "ph": "f", "id": 72, "pid": 1336759, "tid": 1381173, "ts": 1514656762958.336,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762959.014, "dur": 2.512,
    "args": {
      "External id": 2483885,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 5632]], "Ev Idx": 684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762959.574, "dur": 1.493,
    "args": {
      "External id": 2483886,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 5632], [], []], "Ev Idx": 685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762960.372, "dur": 0.581,
    "args": {
      "External id": 2483887,"Record function id": 0, "Concrete Inputs": ["", "[5632, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 5632], [], [], []], "Ev Idx": 686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762967.181, "dur": 4.452,
    "args": {
      "External id": 2483888,"Record function id": 0, "Ev Idx": 687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762968.319, "dur": 2.862,
    "args": {
      "External id": 2483889,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762969.146, "dur": 1.796,
    "args": {
      "External id": 2483890,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762969.770, "dur": 1.091,
    "args": {
      "External id": 2483891,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762974.501, "dur": 9.554,
    "args": {
      "External id": 2483892,"Record function id": 0, "Sequence number": 24740575, "Fwd thread id": 1, "Ev Idx": 691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762977.804, "dur": 4.442,
    "args": {
      "External id": 2483893,"Sequence number": 24740575, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[23068672, 5632, 1]], "Input Dims": [[16, 4096, 5632]], "Ev Idx": 692
    }
  },
  {
    "ph": "f", "id": 73, "pid": 1336759, "tid": 1381173, "ts": 1514656762977.804,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762978.654, "dur": 3.452,
    "args": {
      "External id": 2483894,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656762981.338, "dur": 0.645,
    "args": {
      "External id": 2483895,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763028.932, "dur": 121.113,
    "args": {
      "External id": 2483896,"Record function id": 0, "Sequence number": 24740574, "Fwd thread id": 1, "Ev Idx": 695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763030.126, "dur": 111.383,
    "args": {
      "External id": 2483897,"Sequence number": 24740574, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[65536, 5632]], "Ev Idx": 696
    }
  },
  {
    "ph": "f", "id": 74, "pid": 1336759, "tid": 1381173, "ts": 1514656763030.126,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763032.410, "dur": 6.126,
    "args": {
      "External id": 2483898,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[65536, 5632]], "Ev Idx": 697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763033.337, "dur": 4.522,
    "args": {
      "External id": 2483899,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[5632, 1], [], []], "Input Dims": [[65536, 5632], [], []], "Ev Idx": 698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763037.128, "dur": 0.558,
    "args": {
      "External id": 2483900,"Record function id": 0, "Concrete Inputs": ["", "[5632, 65536]", "[1, 5632]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[5632, 1], [], [], []], "Input Dims": [[65536, 5632], [], [], []], "Ev Idx": 699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763041.376, "dur": 39.873,
    "args": {
      "External id": 2483901,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048]], "Ev Idx": 700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763082.272, "dur": 4.655,
    "args": {
      "External id": 2483902,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763083.021, "dur": 3.360,
    "args": {
      "External id": 2483903,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[5632, 2048], [], []], "Ev Idx": 702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763084.018, "dur": 2.127,
    "args": {
      "External id": 2483904,"Record function id": 0, "Concrete Inputs": ["", "[2048, 5632]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[5632, 2048], [], [], []], "Ev Idx": 703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763088.283, "dur": 7.762,
    "args": {
      "External id": 2483905,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 5632]], "Ev Idx": 704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763088.714, "dur": 6.958,
    "args": {
      "External id": 2483906,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 5632], [], []], "Ev Idx": 705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763093.691, "dur": 1.915,
    "args": {
      "External id": 2483907,"Record function id": 0, "Concrete Inputs": ["", "[5632, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 5632], [], [], []], "Ev Idx": 706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763096.685, "dur": 44.060,
    "args": {
      "External id": 2483908,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048]], "Ev Idx": 707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763154.474, "dur": 29.340,
    "args": {
      "External id": 2483909,"Record function id": 0, "Sequence number": 24740573, "Fwd thread id": 1, "Ev Idx": 708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763155.513, "dur": 3.571,
    "args": {
      "External id": 2483910,"Sequence number": 24740573, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 709
    }
  },
  {
    "ph": "f", "id": 75, "pid": 1336759, "tid": 1381173, "ts": 1514656763155.513,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763156.910, "dur": 2.020,
    "args": {
      "External id": 2483911,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763157.610, "dur": 1.204,
    "args": {
      "External id": 2483912,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763161.791, "dur": 19.300,
    "args": {
      "External id": 2483913,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763187.504, "dur": 10.474,
    "args": {
      "External id": 2483914,"Record function id": 0, "Sequence number": 24740572, "Fwd thread id": 1, "Ev Idx": 713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763188.524, "dur": 7.767,
    "args": {
      "External id": 2483915,"Sequence number": 24740572, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 5632]], "Ev Idx": 714
    }
  },
  {
    "ph": "f", "id": 76, "pid": 1336759, "tid": 1381173, "ts": 1514656763188.524,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763189.344, "dur": 6.731,
    "args": {
      "External id": 2483916,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 5632]], "Ev Idx": 715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763190.080, "dur": 5.467,
    "args": {
      "External id": 2483917,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 5632], [], []], "Ev Idx": 716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763194.718, "dur": 0.651,
    "args": {
      "External id": 2483918,"Record function id": 0, "Concrete Inputs": ["", "[5632, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 5632], [], [], []], "Ev Idx": 717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763201.915, "dur": 4.878,
    "args": {
      "External id": 2483919,"Record function id": 0, "Ev Idx": 718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763203.091, "dur": 3.223,
    "args": {
      "External id": 2483920,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763204.062, "dur": 1.857,
    "args": {
      "External id": 2483921,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763204.550, "dur": 1.266,
    "args": {
      "External id": 2483922,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: LayerNormFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763212.754, "dur": 415.725,
    "args": {
      "External id": 2483923,"Record function id": 0, "Sequence number": 24740571, "Fwd thread id": 1, "Ev Idx": 722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763214.068, "dur": 376.348,
    "args": {
      "External id": 2483924,"Sequence number": 24740571, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [33554432, 8192, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 723
    }
  },
  {
    "ph": "f", "id": 77, "pid": 1336759, "tid": 1381173, "ts": 1514656763214.068,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763235.955, "dur": 33.220,
    "args": {
      "External id": 2483925,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763237.484, "dur": 31.503,
    "args": {
      "External id": 2483926,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763240.141, "dur": 7.283,
    "args": {
      "External id": 2483927,"Record function id": 0, "Concrete Inputs": ["", "15", "0", "", "", "0"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[33554432, 8192, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 2048], [], [], [], [], []], "Ev Idx": 726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763244.320, "dur": 2.482,
    "args": {
      "External id": 2483928,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 2048]", "15", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763248.730, "dur": 19.652,
    "args": {
      "External id": 2483929,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [33554432, 8192, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763279.198, "dur": 3.686,
    "args": {
      "External id": 2483930,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763281.544, "dur": 1.231,
    "args": {
      "External id": 2483931,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763286.378, "dur": 3.536,
    "args": {
      "External id": 2483932,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763286.800, "dur": 3.020,
    "args": {
      "External id": 2483933,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763301.677, "dur": 2.200,
    "args": {
      "External id": 2483934,"Record function id": 0, "Concrete Inputs": ["[65536, 2048]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763313.750, "dur": 1.688,
    "args": {
      "External id": 2483935,"Record function id": 0, "Concrete Inputs": ["[132, 2048]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763464.365, "dur": 2.357,
    "args": {
      "External id": 2483936,"Record function id": 0, "Concrete Inputs": ["", "[1, -1, 2048]"], "Input type": ["float", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[132, 2048], []], "Ev Idx": 735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763471.069, "dur": 28.263,
    "args": {
      "External id": 2483937,"Record function id": 0, "Concrete Inputs": ["", "[1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[270336, 2048, 1], [], [], []], "Input Dims": [[1, 132, 2048], [], [], []], "Ev Idx": 736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763478.043, "dur": 0.809,
    "args": {
      "External id": 2483938,"Record function id": 0, "Concrete Inputs": ["", "[1, 1, 2048]", "[2048, 0, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[1, 2048], [], [], []], "Ev Idx": 737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763504.478, "dur": 49.698,
    "args": {
      "External id": 2483939,"Record function id": 0, "Concrete Inputs": ["", "", "15", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[1, 2048], [], [], [], [], []], "Ev Idx": 738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763506.059, "dur": 47.839,
    "args": {
      "External id": 2483940,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], [], []], "Input Dims": [[1, 2048], [], [], [], [], [], []], "Ev Idx": 739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763511.828, "dur": 3.655,
    "args": {
      "External id": 2483941,"Record function id": 0, "Concrete Inputs": ["[1, 2048]", "[2048, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763516.754, "dur": 36.254,
    "args": {
      "External id": 2483942,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "float", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[1, 2048], [1, 2048], []], "Ev Idx": 741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763560.900, "dur": 5.602,
    "args": {
      "External id": 2483943,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1]], "Input Dims": [[1, 2048], [2048]], "Ev Idx": 742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763564.631, "dur": 1.726,
    "args": {
      "External id": 2483944,"Record function id": 0, "Concrete Inputs": ["", "[2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[1, 2048], []], "Ev Idx": 743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763573.202, "dur": 1.929,
    "args": {
      "External id": 2483945,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763573.951, "dur": 1.049,
    "args": {
      "External id": 2483946,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763578.485, "dur": 3.453,
    "args": {
      "External id": 2483947,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763578.844, "dur": 3.001,
    "args": {
      "External id": 2483948,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763605.361, "dur": 21.496,
    "args": {
      "External id": 2483949,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763641.404, "dur": 11.054,
    "args": {
      "External id": 2483950,"Record function id": 0, "Ev Idx": 749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763643.397, "dur": 8.396,
    "args": {
      "External id": 2483951,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763645.447, "dur": 5.512,
    "args": {
      "External id": 2483952,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763648.413, "dur": 2.435,
    "args": {
      "External id": 2483953,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763656.121, "dur": 6.677,
    "args": {
      "External id": 2483954,"Record function id": 0, "Sequence number": 24740570, "Fwd thread id": 1, "Ev Idx": 753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763657.077, "dur": 3.133,
    "args": {
      "External id": 2483955,"Sequence number": 24740570, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 754
    }
  },
  {
    "ph": "f", "id": 78, "pid": 1336759, "tid": 1381173, "ts": 1514656763657.077,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763658.239, "dur": 1.798,
    "args": {
      "External id": 2483956,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763658.655, "dur": 1.233,
    "args": {
      "External id": 2483957,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763666.122, "dur": 125.551,
    "args": {
      "External id": 2483958,"Record function id": 0, "Sequence number": 24740569, "Fwd thread id": 1, "Ev Idx": 757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763666.823, "dur": 119.771,
    "args": {
      "External id": 2483959,"Sequence number": 24740569, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 758
    }
  },
  {
    "ph": "f", "id": 79, "pid": 1336759, "tid": 1381173, "ts": 1514656763666.823,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763670.147, "dur": 9.908,
    "args": {
      "External id": 2483960,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763674.436, "dur": 5.049,
    "args": {
      "External id": 2483961,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[65536, 2048], [], []], "Ev Idx": 760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763678.246, "dur": 1.073,
    "args": {
      "External id": 2483962,"Record function id": 0, "Concrete Inputs": ["", "[2048, 65536]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763681.181, "dur": 52.282,
    "args": {
      "External id": 2483963,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048]], "Ev Idx": 762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763734.647, "dur": 5.494,
    "args": {
      "External id": 2483964,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763735.188, "dur": 4.368,
    "args": {
      "External id": 2483965,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763736.162, "dur": 3.219,
    "args": {
      "External id": 2483966,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763743.562, "dur": 2.950,
    "args": {
      "External id": 2483967,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763744.769, "dur": 1.340,
    "args": {
      "External id": 2483968,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763745.558, "dur": 0.381,
    "args": {
      "External id": 2483969,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763747.135, "dur": 38.477,
    "args": {
      "External id": 2483970,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763796.144, "dur": 10.183,
    "args": {
      "External id": 2483971,"Record function id": 0, "Sequence number": 24740568, "Fwd thread id": 1, "Ev Idx": 770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763797.035, "dur": 7.428,
    "args": {
      "External id": 2483972,"Sequence number": 24740568, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 771
    }
  },
  {
    "ph": "f", "id": 80, "pid": 1336759, "tid": 1381173, "ts": 1514656763797.035,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763800.306, "dur": 4.005,
    "args": {
      "External id": 2483973,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763801.135, "dur": 3.042,
    "args": {
      "External id": 2483974,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763811.914, "dur": 8.872,
    "args": {
      "External id": 2483975,"Record function id": 0, "Sequence number": 24740567, "Fwd thread id": 1, "Ev Idx": 774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763815.270, "dur": 3.631,
    "args": {
      "External id": 2483976,"Sequence number": 24740567, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 775
    }
  },
  {
    "ph": "f", "id": 81, "pid": 1336759, "tid": 1381173, "ts": 1514656763815.270,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763816.215, "dur": 2.485,
    "args": {
      "External id": 2483977,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763816.899, "dur": 1.346,
    "args": {
      "External id": 2483978,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763817.613, "dur": 0.498,
    "args": {
      "External id": 2483979,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763824.633, "dur": 4.461,
    "args": {
      "External id": 2483980,"Record function id": 0, "Ev Idx": 779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763825.739, "dur": 2.865,
    "args": {
      "External id": 2483981,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763826.795, "dur": 1.497,
    "args": {
      "External id": 2483982,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763827.272, "dur": 0.914,
    "args": {
      "External id": 2483983,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763832.134, "dur": 9.641,
    "args": {
      "External id": 2483984,"Record function id": 0, "Sequence number": 24740566, "Fwd thread id": 1, "Ev Idx": 783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763832.994, "dur": 7.340,
    "args": {
      "External id": 2483985,"Sequence number": 24740566, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 784
    }
  },
  {
    "ph": "f", "id": 82, "pid": 1336759, "tid": 1381173, "ts": 1514656763832.994,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763836.241, "dur": 3.931,
    "args": {
      "External id": 2483986,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763838.871, "dur": 1.168,
    "args": {
      "External id": 2483987,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: FlashAttnFuncBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763845.991, "dur": 317.172,
    "args": {
      "External id": 2483988,"Record function id": 0, "Sequence number": 24740565, "Fwd thread id": 1, "Ev Idx": 787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FlashAttnFuncBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763847.122, "dur": 298.447,
    "args": {
      "External id": 2483989,"Sequence number": 24740565, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 788
    }
  },
  {
    "ph": "f", "id": 83, "pid": 1336759, "tid": 1381173, "ts": 1514656763847.122,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763860.577, "dur": 6.583,
    "args": {
      "External id": 2483990,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763862.858, "dur": 3.860,
    "args": {
      "External id": 2483991,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763868.996, "dur": 5.114,
    "args": {
      "External id": 2483992,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763872.247, "dur": 1.630,
    "args": {
      "External id": 2483993,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763875.354, "dur": 4.192,
    "args": {
      "External id": 2483994,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763876.262, "dur": 3.119,
    "args": {
      "External id": 2483995,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656763905.555, "dur": 214.350,
    "args": {
      "External id": 2483996,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764022.113, "dur": 5.300,
    "args": {
      "External id": 2483997,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764030.444, "dur": 4.295,
    "args": {
      "External id": 2483998,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764132.728, "dur": 3.913,
    "args": {
      "External id": 2483999,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764139.515, "dur": 0.796,
    "args": {
      "External id": 2484000,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764142.296, "dur": 0.625,
    "args": {
      "External id": 2484001,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RotaryEmbeddingFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764172.525, "dur": 223.026,
    "args": {
      "External id": 2484002,"Record function id": 0, "Sequence number": 24740564, "Fwd thread id": 1, "Ev Idx": 801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764174.151, "dur": 214.993,
    "args": {
      "External id": 2484003,"Sequence number": 24740564, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 802
    }
  },
  {
    "ph": "f", "id": 84, "pid": 1336759, "tid": 1381173, "ts": 1514656764174.151,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764195.556, "dur": 44.328,
    "args": {
      "External id": 2484004,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764198.163, "dur": 3.205,
    "args": {
      "External id": 2484005,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764202.900, "dur": 36.464,
    "args": {
      "External id": 2484006,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], []], "Ev Idx": 805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764248.749, "dur": 6.620,
    "args": {
      "External id": 2484007,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764252.995, "dur": 1.957,
    "args": {
      "External id": 2484008,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RotaryEmbeddingFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764402.676, "dur": 163.817,
    "args": {
      "External id": 2484009,"Record function id": 0, "Sequence number": 24740563, "Fwd thread id": 1, "Ev Idx": 808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764404.158, "dur": 154.949,
    "args": {
      "External id": 2484010,"Sequence number": 24740563, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 809
    }
  },
  {
    "ph": "f", "id": 85, "pid": 1336759, "tid": 1381173, "ts": 1514656764404.158,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764416.464, "dur": 28.901,
    "args": {
      "External id": 2484011,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764418.407, "dur": 2.672,
    "args": {
      "External id": 2484012,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764422.114, "dur": 22.705,
    "args": {
      "External id": 2484013,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], []], "Ev Idx": 812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764452.152, "dur": 8.142,
    "args": {
      "External id": 2484014,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764457.832, "dur": 2.140,
    "args": {
      "External id": 2484015,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764574.478, "dur": 14.300,
    "args": {
      "External id": 2484016,"Record function id": 0, "Sequence number": 24740562, "Fwd thread id": 1, "Ev Idx": 815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764575.879, "dur": 10.418,
    "args": {
      "External id": 2484017,"Sequence number": 24740562, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 816
    }
  },
  {
    "ph": "f", "id": 86, "pid": 1336759, "tid": 1381173, "ts": 1514656764575.879,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764578.261, "dur": 7.650,
    "args": {
      "External id": 2484018,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764579.525, "dur": 6.182,
    "args": {
      "External id": 2484019,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764592.252, "dur": 10.120,
    "args": {
      "External id": 2484020,"Record function id": 0, "Sequence number": 24740561, "Fwd thread id": 1, "Ev Idx": 819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764595.352, "dur": 5.228,
    "args": {
      "External id": 2484021,"Sequence number": 24740561, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 820
    }
  },
  {
    "ph": "f", "id": 87, "pid": 1336759, "tid": 1381173, "ts": 1514656764595.352,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764596.376, "dur": 4.049,
    "args": {
      "External id": 2484022,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764599.340, "dur": 0.953,
    "args": {
      "External id": 2484023,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764607.738, "dur": 4.770,
    "args": {
      "External id": 2484024,"Record function id": 0, "Sequence number": 24740560, "Fwd thread id": 1, "Ev Idx": 823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764608.362, "dur": 2.672,
    "args": {
      "External id": 2484025,"Sequence number": 24740560, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 824
    }
  },
  {
    "ph": "f", "id": 88, "pid": 1336759, "tid": 1381173, "ts": 1514656764608.362,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764609.246, "dur": 1.638,
    "args": {
      "External id": 2484026,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764609.909, "dur": 0.881,
    "args": {
      "External id": 2484027,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764615.589, "dur": 9.460,
    "args": {
      "External id": 2484028,"Record function id": 0, "Sequence number": 24740559, "Fwd thread id": 1, "Ev Idx": 827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764616.280, "dur": 6.996,
    "args": {
      "External id": 2484029,"Sequence number": 24740559, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 828
    }
  },
  {
    "ph": "f", "id": 89, "pid": 1336759, "tid": 1381173, "ts": 1514656764616.280,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764617.261, "dur": 5.861,
    "args": {
      "External id": 2484030,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764622.104, "dur": 0.930,
    "args": {
      "External id": 2484031,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764628.502, "dur": 140.883,
    "args": {
      "External id": 2484032,"Record function id": 0, "Sequence number": 24740558, "Fwd thread id": 1, "Ev Idx": 831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764629.408, "dur": 132.689,
    "args": {
      "External id": 2484033,"Sequence number": 24740558, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 832
    }
  },
  {
    "ph": "f", "id": 90, "pid": 1336759, "tid": 1381173, "ts": 1514656764629.408,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764632.682, "dur": 5.938,
    "args": {
      "External id": 2484034,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764634.346, "dur": 3.631,
    "args": {
      "External id": 2484035,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[65536, 2048], [], []], "Ev Idx": 834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764636.094, "dur": 1.647,
    "args": {
      "External id": 2484036,"Record function id": 0, "Concrete Inputs": ["", "[2048, 65536]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764640.262, "dur": 63.075,
    "args": {
      "External id": 2484037,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048]], "Ev Idx": 836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764704.557, "dur": 6.441,
    "args": {
      "External id": 2484038,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764707.349, "dur": 2.757,
    "args": {
      "External id": 2484039,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764708.424, "dur": 1.451,
    "args": {
      "External id": 2484040,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764712.736, "dur": 5.239,
    "args": {
      "External id": 2484041,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764713.694, "dur": 3.741,
    "args": {
      "External id": 2484042,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764717.013, "dur": 0.348,
    "args": {
      "External id": 2484043,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764718.600, "dur": 42.657,
    "args": {
      "External id": 2484044,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764773.958, "dur": 8.858,
    "args": {
      "External id": 2484045,"Record function id": 0, "Sequence number": 24740557, "Fwd thread id": 1, "Ev Idx": 844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764774.945, "dur": 6.177,
    "args": {
      "External id": 2484046,"Sequence number": 24740557, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 845
    }
  },
  {
    "ph": "f", "id": 91, "pid": 1336759, "tid": 1381173, "ts": 1514656764774.945,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764778.997, "dur": 1.974,
    "args": {
      "External id": 2484047,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764779.530, "dur": 1.295,
    "args": {
      "External id": 2484048,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764786.348, "dur": 6.108,
    "args": {
      "External id": 2484049,"Record function id": 0, "Sequence number": 24740556, "Fwd thread id": 1, "Ev Idx": 848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764787.096, "dur": 3.608,
    "args": {
      "External id": 2484050,"Sequence number": 24740556, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 849
    }
  },
  {
    "ph": "f", "id": 92, "pid": 1336759, "tid": 1381173, "ts": 1514656764787.096,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764787.904, "dur": 2.589,
    "args": {
      "External id": 2484051,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764788.534, "dur": 1.496,
    "args": {
      "External id": 2484052,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764789.546, "dur": 0.388,
    "args": {
      "External id": 2484053,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764800.421, "dur": 10.626,
    "args": {
      "External id": 2484054,"Record function id": 0, "Ev Idx": 853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764802.003, "dur": 8.257,
    "args": {
      "External id": 2484055,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764804.370, "dur": 5.449,
    "args": {
      "External id": 2484056,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764807.256, "dur": 2.476,
    "args": {
      "External id": 2484057,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764814.016, "dur": 5.120,
    "args": {
      "External id": 2484058,"Record function id": 0, "Sequence number": 24740555, "Fwd thread id": 1, "Ev Idx": 857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764814.956, "dur": 2.406,
    "args": {
      "External id": 2484059,"Sequence number": 24740555, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 858
    }
  },
  {
    "ph": "f", "id": 93, "pid": 1336759, "tid": 1381173, "ts": 1514656764814.956,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764815.851, "dur": 1.365,
    "args": {
      "External id": 2484060,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764816.352, "dur": 0.735,
    "args": {
      "External id": 2484061,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764822.146, "dur": 84.944,
    "args": {
      "External id": 2484062,"Record function id": 0, "Sequence number": 24740554, "Fwd thread id": 1, "Ev Idx": 861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764822.982, "dur": 79.466,
    "args": {
      "External id": 2484063,"Sequence number": 24740554, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 862
    }
  },
  {
    "ph": "f", "id": 94, "pid": 1336759, "tid": 1381173, "ts": 1514656764822.982,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764824.896, "dur": 4.431,
    "args": {
      "External id": 2484064,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764827.642, "dur": 1.230,
    "args": {
      "External id": 2484065,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[65536, 2048], [], []], "Ev Idx": 864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764828.240, "dur": 0.516,
    "args": {
      "External id": 2484066,"Record function id": 0, "Concrete Inputs": ["", "[2048, 65536]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764829.941, "dur": 24.990,
    "args": {
      "External id": 2484067,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048]], "Ev Idx": 866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764855.874, "dur": 5.280,
    "args": {
      "External id": 2484068,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764856.480, "dur": 4.125,
    "args": {
      "External id": 2484069,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764859.552, "dur": 0.910,
    "args": {
      "External id": 2484070,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764864.285, "dur": 4.045,
    "args": {
      "External id": 2484071,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764864.961, "dur": 3.006,
    "args": {
      "External id": 2484072,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764865.727, "dur": 2.150,
    "args": {
      "External id": 2484073,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764868.824, "dur": 32.857,
    "args": {
      "External id": 2484074,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764911.151, "dur": 32.664,
    "args": {
      "External id": 2484075,"Record function id": 0, "Sequence number": 24740553, "Fwd thread id": 1, "Ev Idx": 874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764911.961, "dur": 5.012,
    "args": {
      "External id": 2484076,"Sequence number": 24740553, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 875
    }
  },
  {
    "ph": "f", "id": 95, "pid": 1336759, "tid": 1381173, "ts": 1514656764911.961,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764913.211, "dur": 3.621,
    "args": {
      "External id": 2484077,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764915.639, "dur": 1.093,
    "args": {
      "External id": 2484078,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764919.974, "dur": 19.844,
    "args": {
      "External id": 2484079,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764947.498, "dur": 6.608,
    "args": {
      "External id": 2484080,"Record function id": 0, "Sequence number": 24740552, "Fwd thread id": 1, "Ev Idx": 879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764948.211, "dur": 3.929,
    "args": {
      "External id": 2484081,"Sequence number": 24740552, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 880
    }
  },
  {
    "ph": "f", "id": 96, "pid": 1336759, "tid": 1381173, "ts": 1514656764948.211,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764949.116, "dur": 2.840,
    "args": {
      "External id": 2484082,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764949.894, "dur": 1.583,
    "args": {
      "External id": 2484083,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764950.820, "dur": 0.477,
    "args": {
      "External id": 2484084,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764957.791, "dur": 6.137,
    "args": {
      "External id": 2484085,"Record function id": 0, "Ev Idx": 884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764958.929, "dur": 4.448,
    "args": {
      "External id": 2484086,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764959.789, "dur": 3.237,
    "args": {
      "External id": 2484087,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764961.907, "dur": 1.013,
    "args": {
      "External id": 2484088,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764966.946, "dur": 7.575,
    "args": {
      "External id": 2484089,"Record function id": 0, "Sequence number": 24740551, "Fwd thread id": 1, "Ev Idx": 888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764967.624, "dur": 4.774,
    "args": {
      "External id": 2484090,"Sequence number": 24740551, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 889
    }
  },
  {
    "ph": "f", "id": 97, "pid": 1336759, "tid": 1381173, "ts": 1514656764967.624,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764968.693, "dur": 3.567,
    "args": {
      "External id": 2484091,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764971.314, "dur": 0.809,
    "args": {
      "External id": 2484092,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764980.173, "dur": 141.567,
    "args": {
      "External id": 2484093,"Record function id": 0, "Sequence number": 24740550, "Fwd thread id": 1, "Ev Idx": 892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764980.888, "dur": 132.764,
    "args": {
      "External id": 2484094,"Sequence number": 24740550, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 893
    }
  },
  {
    "ph": "f", "id": 98, "pid": 1336759, "tid": 1381173, "ts": 1514656764980.888,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764982.787, "dur": 44.299,
    "args": {
      "External id": 2484095,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656764983.256, "dur": 43.010,
    "args": {
      "External id": 2484096,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[65536, 2048], [], []], "Ev Idx": 895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765024.995, "dur": 0.820,
    "args": {
      "External id": 2484097,"Record function id": 0, "Concrete Inputs": ["", "[2048, 65536]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765027.801, "dur": 34.450,
    "args": {
      "External id": 2484098,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048]], "Ev Idx": 897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765063.363, "dur": 3.690,
    "args": {
      "External id": 2484099,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765064.006, "dur": 2.484,
    "args": {
      "External id": 2484100,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765065.108, "dur": 1.234,
    "args": {
      "External id": 2484101,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765068.355, "dur": 9.046,
    "args": {
      "External id": 2484102,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765073.370, "dur": 3.608,
    "args": {
      "External id": 2484103,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765076.482, "dur": 0.429,
    "args": {
      "External id": 2484104,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765077.963, "dur": 34.945,
    "args": {
      "External id": 2484105,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765127.494, "dur": 24.952,
    "args": {
      "External id": 2484106,"Record function id": 0, "Sequence number": 24740549, "Fwd thread id": 1, "Ev Idx": 905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765128.685, "dur": 3.504,
    "args": {
      "External id": 2484107,"Sequence number": 24740549, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 906
    }
  },
  {
    "ph": "f", "id": 99, "pid": 1336759, "tid": 1381173, "ts": 1514656765128.685,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765129.834, "dur": 2.216,
    "args": {
      "External id": 2484108,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765130.440, "dur": 1.514,
    "args": {
      "External id": 2484109,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765134.702, "dur": 15.544,
    "args": {
      "External id": 2484110,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765156.028, "dur": 11.562,
    "args": {
      "External id": 2484111,"Record function id": 0, "Sequence number": 24740548, "Fwd thread id": 1, "Ev Idx": 910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765159.522, "dur": 5.937,
    "args": {
      "External id": 2484112,"Sequence number": 24740548, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 911
    }
  },
  {
    "ph": "f", "id": 100, "pid": 1336759, "tid": 1381173, "ts": 1514656765159.522,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765160.382, "dur": 4.870,
    "args": {
      "External id": 2484113,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765160.907, "dur": 3.715,
    "args": {
      "External id": 2484114,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765163.863, "dur": 0.579,
    "args": {
      "External id": 2484115,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765171.630, "dur": 5.456,
    "args": {
      "External id": 2484116,"Record function id": 0, "Ev Idx": 915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765173.103, "dur": 3.550,
    "args": {
      "External id": 2484117,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765174.329, "dur": 2.026,
    "args": {
      "External id": 2484118,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765175.096, "dur": 1.134,
    "args": {
      "External id": 2484119,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: LayerNormFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765180.842, "dur": 337.402,
    "args": {
      "External id": 2484120,"Record function id": 0, "Sequence number": 24740547, "Fwd thread id": 1, "Ev Idx": 919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765181.880, "dur": 309.314,
    "args": {
      "External id": 2484121,"Sequence number": 24740547, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 920
    }
  },
  {
    "ph": "f", "id": 101, "pid": 1336759, "tid": 1381173, "ts": 1514656765181.880,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765215.036, "dur": 4.279,
    "args": {
      "External id": 2484122,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765218.232, "dur": 0.938,
    "args": {
      "External id": 2484123,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765232.780, "dur": 5.850,
    "args": {
      "External id": 2484124,"Record function id": 0, "Concrete Inputs": ["[65536, 2048]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765247.797, "dur": 1.762,
    "args": {
      "External id": 2484125,"Record function id": 0, "Concrete Inputs": ["[132, 2048]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765391.145, "dur": 1.770,
    "args": {
      "External id": 2484126,"Record function id": 0, "Concrete Inputs": ["", "[1, -1, 2048]"], "Input type": ["float", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[132, 2048], []], "Ev Idx": 925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765396.894, "dur": 33.032,
    "args": {
      "External id": 2484127,"Record function id": 0, "Concrete Inputs": ["", "[1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[270336, 2048, 1], [], [], []], "Input Dims": [[1, 132, 2048], [], [], []], "Ev Idx": 926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765406.499, "dur": 0.816,
    "args": {
      "External id": 2484128,"Record function id": 0, "Concrete Inputs": ["", "[1, 1, 2048]", "[2048, 0, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[1, 2048], [], [], []], "Ev Idx": 927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765439.021, "dur": 32.601,
    "args": {
      "External id": 2484129,"Record function id": 0, "Concrete Inputs": ["", "", "15", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[1, 2048], [], [], [], [], []], "Ev Idx": 928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765440.573, "dur": 30.763,
    "args": {
      "External id": 2484130,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], [], []], "Input Dims": [[1, 2048], [], [], [], [], [], []], "Ev Idx": 929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765448.598, "dur": 4.932,
    "args": {
      "External id": 2484131,"Record function id": 0, "Concrete Inputs": ["[1, 2048]", "[2048, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765455.079, "dur": 15.625,
    "args": {
      "External id": 2484132,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "float", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[1, 2048], [1, 2048], []], "Ev Idx": 931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765476.068, "dur": 2.577,
    "args": {
      "External id": 2484133,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1]], "Input Dims": [[1, 2048], [2048]], "Ev Idx": 932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765477.357, "dur": 1.182,
    "args": {
      "External id": 2484134,"Record function id": 0, "Concrete Inputs": ["", "[2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[1, 2048], []], "Ev Idx": 933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765483.441, "dur": 1.575,
    "args": {
      "External id": 2484135,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765484.101, "dur": 0.795,
    "args": {
      "External id": 2484136,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765501.267, "dur": 13.292,
    "args": {
      "External id": 2484137,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765545.214, "dur": 12.008,
    "args": {
      "External id": 2484138,"Record function id": 0, "Ev Idx": 937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765549.630, "dur": 6.555,
    "args": {
      "External id": 2484139,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765551.632, "dur": 3.194,
    "args": {
      "External id": 2484140,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765552.449, "dur": 2.140,
    "args": {
      "External id": 2484141,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: AddBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765562.372, "dur": 5.336,
    "args": {
      "External id": 2484142,"Record function id": 0, "Sequence number": 24740546, "Fwd thread id": 1, "Ev Idx": 941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AddBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765563.520, "dur": 1.254,
    "args": {
      "External id": 2484143,"Sequence number": 24740546, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[33554432, 8192, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 942
    }
  },
  {
    "ph": "f", "id": 102, "pid": 1336759, "tid": 1381173, "ts": 1514656765563.520,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SwiGLULinearFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765571.459, "dur": 465.801,
    "args": {
      "External id": 2484144,"Record function id": 0, "Sequence number": 24740545, "Fwd thread id": 1, "Ev Idx": 943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SwiGLULinearFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765572.534, "dur": 451.830,
    "args": {
      "External id": 2484145,"Sequence number": 24740545, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[33554432, 8192, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 944
    }
  },
  {
    "ph": "f", "id": 103, "pid": 1336759, "tid": 1381173, "ts": 1514656765572.534,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765603.466, "dur": 7.679,
    "args": {
      "External id": 2484146,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[33554432, 8192, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765607.916, "dur": 2.971,
    "args": {
      "External id": 2484147,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]", "[8192, 1]"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList"], "Input Strides": [[33554432, 8192, 1], [], []], "Input Dims": [[16, 4096, 2048], [], []], "Ev Idx": 946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765613.999, "dur": 9.697,
    "args": {
      "External id": 2484148,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765615.560, "dur": 7.518,
    "args": {
      "External id": 2484149,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[5632, 1], [], []], "Input Dims": [[2048, 5632], [], []], "Ev Idx": 948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765619.424, "dur": 3.485,
    "args": {
      "External id": 2484150,"Record function id": 0, "Concrete Inputs": ["", "[5632, 2048]", "[1, 5632]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[5632, 1], [], [], []], "Input Dims": [[2048, 5632], [], [], []], "Ev Idx": 949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765627.389, "dur": 99.396,
    "args": {
      "External id": 2484151,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[8192, 1], [1, 5632], []], "Input Dims": [[65536, 2048], [5632, 2048], []], "Ev Idx": 950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765628.520, "dur": 3.979,
    "args": {
      "External id": 2484152,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 5632]], "Input Dims": [[5632, 2048]], "Ev Idx": 951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765630.780, "dur": 1.306,
    "args": {
      "External id": 2484153,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 5632], [], []], "Input Dims": [[5632, 2048], [], []], "Ev Idx": 952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765631.419, "dur": 0.594,
    "args": {
      "External id": 2484154,"Record function id": 0, "Concrete Inputs": ["", "[2048, 5632]", "[5632, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 5632], [], [], []], "Input Dims": [[5632, 2048], [], [], []], "Ev Idx": 953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765633.823, "dur": 92.496,
    "args": {
      "External id": 2484155,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8192, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632]], "Ev Idx": 954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765635.498, "dur": 89.997,
    "args": {
      "External id": 2484156,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8192, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632]], "Ev Idx": 955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765730.423, "dur": 2.365,
    "args": {
      "External id": 2484157,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [23068672, 5632, 1]], "Input Dims": [[65536, 5632], [16, 4096, 5632]], "Ev Idx": 956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765731.300, "dur": 1.357,
    "args": {
      "External id": 2484158,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 1], []], "Input Dims": [[65536, 5632], []], "Ev Idx": 957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765764.182, "dur": 5.084,
    "args": {
      "External id": 2484159,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 5632]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765772.296, "dur": 1.951,
    "args": {
      "External id": 2484160,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 5632]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765774.919, "dur": 1.575,
    "args": {
      "External id": 2484161,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 5632]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765807.673, "dur": 1.816,
    "args": {
      "External id": 2484162,"Record function id": 0, "Concrete Inputs": ["", "[-1, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765808.227, "dur": 1.112,
    "args": {
      "External id": 2484163,"Record function id": 0, "Concrete Inputs": ["", "[-1, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::einsum", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765831.210, "dur": 137.773,
    "args": {
      "External id": 2484164,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["", "TensorList", ""], "Input Strides": [[], [[8192, 1], [5632, 1]], []], "Input Dims": [[], [[65536, 2048], [65536, 5632]], []], "Ev Idx": 963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765836.199, "dur": 8.832,
    "args": {
      "External id": 2484165,"Record function id": 0, "Concrete Inputs": ["", "2"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765841.161, "dur": 2.765,
    "args": {
      "External id": 2484166,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048, 1]", "[8192, 1, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[8192, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765846.582, "dur": 7.635,
    "args": {
      "External id": 2484167,"Record function id": 0, "Concrete Inputs": ["", "[1, 2, 0]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8192, 1, 1], []], "Input Dims": [[65536, 2048, 1], []], "Ev Idx": 966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765852.990, "dur": 0.497,
    "args": {
      "External id": 2484168,"Record function id": 0, "Concrete Inputs": ["", "[2048, 1, 65536]", "[1, 1, 8192]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[65536, 2048, 1], [], [], []], "Ev Idx": 967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765855.510, "dur": 1.473,
    "args": {
      "External id": 2484169,"Record function id": 0, "Concrete Inputs": ["", "2"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], []], "Input Dims": [[65536, 5632], []], "Ev Idx": 968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765856.240, "dur": 0.387,
    "args": {
      "External id": 2484170,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632, 1]", "[5632, 1, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[5632, 1], [], [], []], "Input Dims": [[65536, 5632], [], [], []], "Ev Idx": 969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765860.203, "dur": 2.064,
    "args": {
      "External id": 2484171,"Record function id": 0, "Concrete Inputs": ["", "[2, 1, 0]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 1, 1], []], "Input Dims": [[65536, 5632, 1], []], "Ev Idx": 970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765861.119, "dur": 0.596,
    "args": {
      "External id": 2484172,"Record function id": 0, "Concrete Inputs": ["", "[1, 5632, 65536]", "[1, 1, 5632]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[5632, 1, 1], [], [], []], "Input Dims": [[65536, 5632, 1], [], [], []], "Ev Idx": 971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765867.425, "dur": 1.810,
    "args": {
      "External id": 2484173,"Record function id": 0, "Concrete Inputs": ["", "[0, 2, 1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1, 1, 8192], []], "Input Dims": [[2048, 1, 65536], []], "Ev Idx": 972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765868.487, "dur": 0.446,
    "args": {
      "External id": 2484174,"Record function id": 0, "Concrete Inputs": ["", "[2048, 65536, 1]", "[1, 8192, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1, 8192], [], [], []], "Input Dims": [[2048, 1, 65536], [], [], []], "Ev Idx": 973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765870.103, "dur": 6.374,
    "args": {
      "External id": 2484175,"Record function id": 0, "Concrete Inputs": ["", "[1, 2048, 65536]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1, 8192, 1], []], "Input Dims": [[2048, 65536, 1], []], "Ev Idx": 974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765874.763, "dur": 1.563,
    "args": {
      "External id": 2484176,"Record function id": 0, "Concrete Inputs": ["", "[1, 2048, 65536]", "[2048, 1, 8192]"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList"], "Input Strides": [[1, 8192, 1], [], []], "Input Dims": [[2048, 65536, 1], [], []], "Ev Idx": 975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765877.240, "dur": 3.417,
    "args": {
      "External id": 2484177,"Record function id": 0, "Concrete Inputs": ["", "[2, 1, 0]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1, 1, 5632], []], "Input Dims": [[1, 5632, 65536], []], "Ev Idx": 976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765879.898, "dur": 0.486,
    "args": {
      "External id": 2484178,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632, 1]", "[5632, 1, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1, 5632], [], [], []], "Input Dims": [[1, 5632, 65536], [], [], []], "Ev Idx": 977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765883.178, "dur": 2.885,
    "args": {
      "External id": 2484179,"Record function id": 0, "Concrete Inputs": ["", "[1, 65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 1, 1], []], "Input Dims": [[65536, 5632, 1], []], "Ev Idx": 978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765883.783, "dur": 2.202,
    "args": {
      "External id": 2484180,"Record function id": 0, "Concrete Inputs": ["", "[1, 65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 1, 1], []], "Input Dims": [[65536, 5632, 1], []], "Ev Idx": 979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765887.270, "dur": 67.927,
    "args": {
      "External id": 2484181,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1, 8192], [369098752, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632]], "Ev Idx": 980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765959.166, "dur": 1.066,
    "args": {
      "External id": 2484182,"Record function id": 0, "Concrete Inputs": ["", "[2048, 1, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 5632, 1], []], "Input Dims": [[1, 2048, 5632], []], "Ev Idx": 981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765960.871, "dur": 4.335,
    "args": {
      "External id": 2484183,"Record function id": 0, "Concrete Inputs": ["", "[0, 2, 1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 5632, 1], []], "Input Dims": [[2048, 1, 5632], []], "Ev Idx": 982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765964.404, "dur": 0.369,
    "args": {
      "External id": 2484184,"Record function id": 0, "Concrete Inputs": ["", "[2048, 5632, 1]", "[5632, 1, 5632]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[5632, 5632, 1], [], [], []], "Input Dims": [[2048, 1, 5632], [], [], []], "Ev Idx": 983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656765967.148, "dur": 0.798,
    "args": {
      "External id": 2484185,"Record function id": 0, "Concrete Inputs": ["", "[2048, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 1, 5632], []], "Input Dims": [[2048, 5632, 1], []], "Ev Idx": 984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766049.286, "dur": 9.260,
    "args": {
      "External id": 2484186,"Record function id": 0, "Ev Idx": 985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766051.262, "dur": 6.666,
    "args": {
      "External id": 2484187,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766053.411, "dur": 3.619,
    "args": {
      "External id": 2484188,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766054.493, "dur": 2.448,
    "args": {
      "External id": 2484189,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766062.170, "dur": 8.973,
    "args": {
      "External id": 2484190,"Record function id": 0, "Sequence number": 24740544, "Fwd thread id": 1, "Ev Idx": 989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766063.095, "dur": 6.211,
    "args": {
      "External id": 2484191,"Sequence number": 24740544, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[23068672, 5632, 1]], "Input Dims": [[16, 4096, 5632]], "Ev Idx": 990
    }
  },
  {
    "ph": "f", "id": 104, "pid": 1336759, "tid": 1381173, "ts": 1514656766063.095,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766067.234, "dur": 1.847,
    "args": {
      "External id": 2484192,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766067.788, "dur": 1.180,
    "args": {
      "External id": 2484193,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766074.581, "dur": 179.590,
    "args": {
      "External id": 2484194,"Record function id": 0, "Sequence number": 24740543, "Fwd thread id": 1, "Ev Idx": 993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766075.386, "dur": 172.966,
    "args": {
      "External id": 2484195,"Sequence number": 24740543, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[65536, 5632]], "Ev Idx": 994
    }
  },
  {
    "ph": "f", "id": 105, "pid": 1336759, "tid": 1381173, "ts": 1514656766075.386,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766078.457, "dur": 5.877,
    "args": {
      "External id": 2484196,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[65536, 5632]], "Ev Idx": 995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766079.462, "dur": 4.251,
    "args": {
      "External id": 2484197,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[5632, 1], [], []], "Input Dims": [[65536, 5632], [], []], "Ev Idx": 996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766082.737, "dur": 0.801,
    "args": {
      "External id": 2484198,"Record function id": 0, "Concrete Inputs": ["", "[5632, 65536]", "[1, 5632]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[5632, 1], [], [], []], "Input Dims": [[65536, 5632], [], [], []], "Ev Idx": 997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766085.529, "dur": 60.069,
    "args": {
      "External id": 2484199,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048]], "Ev Idx": 998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766146.741, "dur": 5.538,
    "args": {
      "External id": 2484200,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766147.526, "dur": 4.033,
    "args": {
      "External id": 2484201,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[5632, 2048], [], []], "Ev Idx": 1000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766150.691, "dur": 0.712,
    "args": {
      "External id": 2484202,"Record function id": 0, "Concrete Inputs": ["", "[2048, 5632]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[5632, 2048], [], [], []], "Ev Idx": 1001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766153.836, "dur": 6.623,
    "args": {
      "External id": 2484203,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 5632]], "Ev Idx": 1002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766154.770, "dur": 5.287,
    "args": {
      "External id": 2484204,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 5632], [], []], "Ev Idx": 1003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766157.870, "dur": 2.088,
    "args": {
      "External id": 2484205,"Record function id": 0, "Concrete Inputs": ["", "[5632, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 5632], [], [], []], "Ev Idx": 1004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766161.031, "dur": 86.520,
    "args": {
      "External id": 2484206,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048]], "Ev Idx": 1005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766258.527, "dur": 7.872,
    "args": {
      "External id": 2484207,"Record function id": 0, "Sequence number": 24740542, "Fwd thread id": 1, "Ev Idx": 1006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766259.705, "dur": 5.487,
    "args": {
      "External id": 2484208,"Sequence number": 24740542, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 1007
    }
  },
  {
    "ph": "f", "id": 106, "pid": 1336759, "tid": 1381173, "ts": 1514656766259.705,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766261.038, "dur": 3.994,
    "args": {
      "External id": 2484209,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766263.551, "dur": 1.356,
    "args": {
      "External id": 2484210,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766269.589, "dur": 8.284,
    "args": {
      "External id": 2484211,"Record function id": 0, "Sequence number": 24740541, "Fwd thread id": 1, "Ev Idx": 1010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766270.327, "dur": 5.900,
    "args": {
      "External id": 2484212,"Sequence number": 24740541, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 5632]], "Ev Idx": 1011
    }
  },
  {
    "ph": "f", "id": 107, "pid": 1336759, "tid": 1381173, "ts": 1514656766270.327,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766271.105, "dur": 4.916,
    "args": {
      "External id": 2484213,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 5632]], "Ev Idx": 1012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766271.918, "dur": 3.646,
    "args": {
      "External id": 2484214,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 5632], [], []], "Ev Idx": 1013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766274.963, "dur": 0.493,
    "args": {
      "External id": 2484215,"Record function id": 0, "Concrete Inputs": ["", "[5632, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 5632], [], [], []], "Ev Idx": 1014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766281.475, "dur": 4.434,
    "args": {
      "External id": 2484216,"Record function id": 0, "Ev Idx": 1015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766282.893, "dur": 2.503,
    "args": {
      "External id": 2484217,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766283.762, "dur": 1.420,
    "args": {
      "External id": 2484218,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766284.183, "dur": 0.908,
    "args": {
      "External id": 2484219,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766291.472, "dur": 6.891,
    "args": {
      "External id": 2484220,"Record function id": 0, "Sequence number": 24740540, "Fwd thread id": 1, "Ev Idx": 1019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766292.344, "dur": 4.487,
    "args": {
      "External id": 2484221,"Sequence number": 24740540, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[23068672, 5632, 1]], "Input Dims": [[16, 4096, 5632]], "Ev Idx": 1020
    }
  },
  {
    "ph": "f", "id": 108, "pid": 1336759, "tid": 1381173, "ts": 1514656766292.344,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766293.358, "dur": 3.336,
    "args": {
      "External id": 2484222,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 1021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766295.823, "dur": 0.749,
    "args": {
      "External id": 2484223,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 1022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766301.480, "dur": 180.406,
    "args": {
      "External id": 2484224,"Record function id": 0, "Sequence number": 24740539, "Fwd thread id": 1, "Ev Idx": 1023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766324.322, "dur": 151.647,
    "args": {
      "External id": 2484225,"Sequence number": 24740539, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[65536, 5632]], "Ev Idx": 1024
    }
  },
  {
    "ph": "f", "id": 109, "pid": 1336759, "tid": 1381173, "ts": 1514656766324.322,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766327.820, "dur": 2.304,
    "args": {
      "External id": 2484226,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[65536, 5632]], "Ev Idx": 1025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766328.451, "dur": 1.247,
    "args": {
      "External id": 2484227,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[5632, 1], [], []], "Input Dims": [[65536, 5632], [], []], "Ev Idx": 1026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766329.116, "dur": 0.424,
    "args": {
      "External id": 2484228,"Record function id": 0, "Concrete Inputs": ["", "[5632, 65536]", "[1, 5632]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[5632, 1], [], [], []], "Input Dims": [[65536, 5632], [], [], []], "Ev Idx": 1027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766330.623, "dur": 49.490,
    "args": {
      "External id": 2484229,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048]], "Ev Idx": 1028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766381.333, "dur": 8.059,
    "args": {
      "External id": 2484230,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766381.902, "dur": 6.948,
    "args": {
      "External id": 2484231,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[5632, 2048], [], []], "Ev Idx": 1030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766384.880, "dur": 3.806,
    "args": {
      "External id": 2484232,"Record function id": 0, "Concrete Inputs": ["", "[2048, 5632]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[5632, 2048], [], [], []], "Ev Idx": 1031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766390.242, "dur": 4.221,
    "args": {
      "External id": 2484233,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 5632]], "Ev Idx": 1032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766390.867, "dur": 3.217,
    "args": {
      "External id": 2484234,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 5632], [], []], "Ev Idx": 1033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766393.718, "dur": 0.299,
    "args": {
      "External id": 2484235,"Record function id": 0, "Concrete Inputs": ["", "[5632, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 5632], [], [], []], "Ev Idx": 1034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766394.929, "dur": 80.230,
    "args": {
      "External id": 2484236,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048]], "Ev Idx": 1035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766486.167, "dur": 30.012,
    "args": {
      "External id": 2484237,"Record function id": 0, "Sequence number": 24740538, "Fwd thread id": 1, "Ev Idx": 1036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766487.027, "dur": 4.743,
    "args": {
      "External id": 2484238,"Sequence number": 24740538, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 1037
    }
  },
  {
    "ph": "f", "id": 110, "pid": 1336759, "tid": 1381173, "ts": 1514656766487.027,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766489.946, "dur": 1.680,
    "args": {
      "External id": 2484239,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766490.464, "dur": 1.035,
    "args": {
      "External id": 2484240,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766494.441, "dur": 18.838,
    "args": {
      "External id": 2484241,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 1040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766519.870, "dur": 25.475,
    "args": {
      "External id": 2484242,"Record function id": 0, "Sequence number": 24740537, "Fwd thread id": 1, "Ev Idx": 1041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766520.853, "dur": 5.592,
    "args": {
      "External id": 2484243,"Sequence number": 24740537, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 5632]], "Ev Idx": 1042
    }
  },
  {
    "ph": "f", "id": 111, "pid": 1336759, "tid": 1381173, "ts": 1514656766520.853,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766521.466, "dur": 4.786,
    "args": {
      "External id": 2484244,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 5632]], "Ev Idx": 1043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766522.310, "dur": 3.446,
    "args": {
      "External id": 2484245,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 5632], [], []], "Ev Idx": 1044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766525.248, "dur": 0.387,
    "args": {
      "External id": 2484246,"Record function id": 0, "Concrete Inputs": ["", "[5632, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 5632], [], [], []], "Ev Idx": 1045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766551.710, "dur": 8.315,
    "args": {
      "External id": 2484247,"Record function id": 0, "Ev Idx": 1046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766553.125, "dur": 6.423,
    "args": {
      "External id": 2484248,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766554.458, "dur": 4.782,
    "args": {
      "External id": 2484249,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766557.401, "dur": 1.676,
    "args": {
      "External id": 2484250,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: LayerNormFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766563.970, "dur": 402.377,
    "args": {
      "External id": 2484251,"Record function id": 0, "Sequence number": 24740536, "Fwd thread id": 1, "Ev Idx": 1050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766565.337, "dur": 369.268,
    "args": {
      "External id": 2484252,"Sequence number": 24740536, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [33554432, 8192, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 1051
    }
  },
  {
    "ph": "f", "id": 112, "pid": 1336759, "tid": 1381173, "ts": 1514656766565.337,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766588.229, "dur": 33.887,
    "args": {
      "External id": 2484253,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766589.630, "dur": 32.291,
    "args": {
      "External id": 2484254,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766592.222, "dur": 6.143,
    "args": {
      "External id": 2484255,"Record function id": 0, "Concrete Inputs": ["", "15", "0", "", "", "0"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[33554432, 8192, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 2048], [], [], [], [], []], "Ev Idx": 1054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766594.489, "dur": 3.380,
    "args": {
      "External id": 2484256,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 2048]", "15", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766599.557, "dur": 21.795,
    "args": {
      "External id": 2484257,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [33554432, 8192, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 1056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766633.847, "dur": 3.717,
    "args": {
      "External id": 2484258,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766634.555, "dur": 2.897,
    "args": {
      "External id": 2484259,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766641.065, "dur": 3.450,
    "args": {
      "External id": 2484260,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766643.700, "dur": 0.713,
    "args": {
      "External id": 2484261,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766655.978, "dur": 2.440,
    "args": {
      "External id": 2484262,"Record function id": 0, "Concrete Inputs": ["[65536, 2048]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766668.460, "dur": 2.441,
    "args": {
      "External id": 2484263,"Record function id": 0, "Concrete Inputs": ["[132, 2048]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766823.050, "dur": 2.439,
    "args": {
      "External id": 2484264,"Record function id": 0, "Concrete Inputs": ["", "[1, -1, 2048]"], "Input type": ["float", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[132, 2048], []], "Ev Idx": 1063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766831.793, "dur": 33.537,
    "args": {
      "External id": 2484265,"Record function id": 0, "Concrete Inputs": ["", "[1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[270336, 2048, 1], [], [], []], "Input Dims": [[1, 132, 2048], [], [], []], "Ev Idx": 1064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766843.855, "dur": 0.873,
    "args": {
      "External id": 2484266,"Record function id": 0, "Concrete Inputs": ["", "[1, 1, 2048]", "[2048, 0, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[1, 2048], [], [], []], "Ev Idx": 1065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766870.673, "dur": 30.317,
    "args": {
      "External id": 2484267,"Record function id": 0, "Concrete Inputs": ["", "", "15", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[1, 2048], [], [], [], [], []], "Ev Idx": 1066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766874.471, "dur": 26.265,
    "args": {
      "External id": 2484268,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], [], []], "Input Dims": [[1, 2048], [], [], [], [], [], []], "Ev Idx": 1067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766880.175, "dur": 4.125,
    "args": {
      "External id": 2484269,"Record function id": 0, "Concrete Inputs": ["[1, 2048]", "[2048, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766885.703, "dur": 14.477,
    "args": {
      "External id": 2484270,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "float", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[1, 2048], [1, 2048], []], "Ev Idx": 1069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766905.532, "dur": 2.161,
    "args": {
      "External id": 2484271,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1]], "Input Dims": [[1, 2048], [2048]], "Ev Idx": 1070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766906.672, "dur": 0.938,
    "args": {
      "External id": 2484272,"Record function id": 0, "Concrete Inputs": ["", "[2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[1, 2048], []], "Ev Idx": 1071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766913.572, "dur": 6.200,
    "args": {
      "External id": 2484273,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766916.290, "dur": 3.345,
    "args": {
      "External id": 2484274,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766921.498, "dur": 4.009,
    "args": {
      "External id": 2484275,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766924.372, "dur": 1.053,
    "args": {
      "External id": 2484276,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766947.810, "dur": 16.926,
    "args": {
      "External id": 2484277,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 1076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766976.248, "dur": 7.180,
    "args": {
      "External id": 2484278,"Record function id": 0, "Ev Idx": 1077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766978.311, "dur": 4.439,
    "args": {
      "External id": 2484279,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 1078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766979.739, "dur": 2.291,
    "args": {
      "External id": 2484280,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 1079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656766980.389, "dur": 1.531,
    "args": {
      "External id": 2484281,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 1080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767028.765, "dur": 11.741,
    "args": {
      "External id": 2484282,"Record function id": 0, "Sequence number": 24740535, "Fwd thread id": 1, "Ev Idx": 1081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767030.252, "dur": 7.198,
    "args": {
      "External id": 2484283,"Sequence number": 24740535, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 1082
    }
  },
  {
    "ph": "f", "id": 113, "pid": 1336759, "tid": 1381173, "ts": 1514656767030.252,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767034.019, "dur": 3.217,
    "args": {
      "External id": 2484284,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767034.650, "dur": 2.377,
    "args": {
      "External id": 2484285,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767043.929, "dur": 137.948,
    "args": {
      "External id": 2484286,"Record function id": 0, "Sequence number": 24740534, "Fwd thread id": 1, "Ev Idx": 1085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767044.680, "dur": 130.346,
    "args": {
      "External id": 2484287,"Sequence number": 24740534, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 1086
    }
  },
  {
    "ph": "f", "id": 114, "pid": 1336759, "tid": 1381173, "ts": 1514656767044.680,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767047.724, "dur": 7.529,
    "args": {
      "External id": 2484288,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 1087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767049.491, "dur": 5.125,
    "args": {
      "External id": 2484289,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[65536, 2048], [], []], "Ev Idx": 1088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767053.409, "dur": 0.991,
    "args": {
      "External id": 2484290,"Record function id": 0, "Concrete Inputs": ["", "[2048, 65536]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 1089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767056.496, "dur": 61.065,
    "args": {
      "External id": 2484291,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048]], "Ev Idx": 1090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767118.945, "dur": 7.188,
    "args": {
      "External id": 2484292,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767119.557, "dur": 5.951,
    "args": {
      "External id": 2484293,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 1092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767122.825, "dur": 2.490,
    "args": {
      "External id": 2484294,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 1093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767127.692, "dur": 4.848,
    "args": {
      "External id": 2484295,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 1094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767128.803, "dur": 3.310,
    "args": {
      "External id": 2484296,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 1095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767131.477, "dur": 0.520,
    "args": {
      "External id": 2484297,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 1096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767133.040, "dur": 41.047,
    "args": {
      "External id": 2484298,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 1097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767186.604, "dur": 10.752,
    "args": {
      "External id": 2484299,"Record function id": 0, "Sequence number": 24740533, "Fwd thread id": 1, "Ev Idx": 1098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767187.870, "dur": 7.337,
    "args": {
      "External id": 2484300,"Sequence number": 24740533, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 1099
    }
  },
  {
    "ph": "f", "id": 115, "pid": 1336759, "tid": 1381173, "ts": 1514656767187.870,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767188.982, "dur": 6.058,
    "args": {
      "External id": 2484301,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767191.834, "dur": 3.073,
    "args": {
      "External id": 2484302,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767202.629, "dur": 9.201,
    "args": {
      "External id": 2484303,"Record function id": 0, "Sequence number": 24740532, "Fwd thread id": 1, "Ev Idx": 1102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767203.659, "dur": 5.905,
    "args": {
      "External id": 2484304,"Sequence number": 24740532, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 1103
    }
  },
  {
    "ph": "f", "id": 116, "pid": 1336759, "tid": 1381173, "ts": 1514656767203.659,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767204.565, "dur": 4.794,
    "args": {
      "External id": 2484305,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 1104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767205.352, "dur": 3.464,
    "args": {
      "External id": 2484306,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 1105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767208.165, "dur": 0.497,
    "args": {
      "External id": 2484307,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 1106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767216.329, "dur": 5.725,
    "args": {
      "External id": 2484308,"Record function id": 0, "Ev Idx": 1107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767217.710, "dur": 3.795,
    "args": {
      "External id": 2484309,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767219.030, "dur": 2.235,
    "args": {
      "External id": 2484310,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767219.801, "dur": 1.345,
    "args": {
      "External id": 2484311,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767226.790, "dur": 6.178,
    "args": {
      "External id": 2484312,"Record function id": 0, "Sequence number": 24740531, "Fwd thread id": 1, "Ev Idx": 1111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767228.045, "dur": 3.137,
    "args": {
      "External id": 2484313,"Sequence number": 24740531, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 1112
    }
  },
  {
    "ph": "f", "id": 117, "pid": 1336759, "tid": 1381173, "ts": 1514656767228.045,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767229.064, "dur": 1.974,
    "args": {
      "External id": 2484314,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767229.624, "dur": 1.277,
    "args": {
      "External id": 2484315,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: FlashAttnFuncBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767238.779, "dur": 280.784,
    "args": {
      "External id": 2484316,"Record function id": 0, "Sequence number": 24740530, "Fwd thread id": 1, "Ev Idx": 1115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FlashAttnFuncBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767240.159, "dur": 263.054,
    "args": {
      "External id": 2484317,"Sequence number": 24740530, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 1116
    }
  },
  {
    "ph": "f", "id": 118, "pid": 1336759, "tid": 1381173, "ts": 1514656767240.159,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767253.641, "dur": 6.743,
    "args": {
      "External id": 2484318,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 1117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767256.170, "dur": 3.702,
    "args": {
      "External id": 2484319,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767262.219, "dur": 6.684,
    "args": {
      "External id": 2484320,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 1119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767265.278, "dur": 3.425,
    "args": {
      "External id": 2484321,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767277.593, "dur": 4.597,
    "args": {
      "External id": 2484322,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 1121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767280.498, "dur": 1.525,
    "args": {
      "External id": 2484323,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767309.874, "dur": 168.430,
    "args": {
      "External id": 2484324,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 1123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767388.060, "dur": 5.246,
    "args": {
      "External id": 2484325,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767394.813, "dur": 3.947,
    "args": {
      "External id": 2484326,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767491.033, "dur": 3.431,
    "args": {
      "External id": 2484327,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 1126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767497.217, "dur": 0.670,
    "args": {
      "External id": 2484328,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 1127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767499.852, "dur": 0.757,
    "args": {
      "External id": 2484329,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 1128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RotaryEmbeddingFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767526.383, "dur": 245.899,
    "args": {
      "External id": 2484330,"Record function id": 0, "Sequence number": 24740529, "Fwd thread id": 1, "Ev Idx": 1129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767547.573, "dur": 217.808,
    "args": {
      "External id": 2484331,"Sequence number": 24740529, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 1130
    }
  },
  {
    "ph": "f", "id": 119, "pid": 1336759, "tid": 1381173, "ts": 1514656767547.573,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767568.114, "dur": 46.455,
    "args": {
      "External id": 2484332,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 1131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767570.748, "dur": 4.202,
    "args": {
      "External id": 2484333,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767576.543, "dur": 37.351,
    "args": {
      "External id": 2484334,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], []], "Ev Idx": 1133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767623.815, "dur": 6.212,
    "args": {
      "External id": 2484335,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 1134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767627.627, "dur": 2.084,
    "args": {
      "External id": 2484336,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RotaryEmbeddingFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767780.626, "dur": 149.300,
    "args": {
      "External id": 2484337,"Record function id": 0, "Sequence number": 24740528, "Fwd thread id": 1, "Ev Idx": 1136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767782.211, "dur": 142.966,
    "args": {
      "External id": 2484338,"Sequence number": 24740528, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 1137
    }
  },
  {
    "ph": "f", "id": 120, "pid": 1336759, "tid": 1381173, "ts": 1514656767782.211,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767793.524, "dur": 39.602,
    "args": {
      "External id": 2484339,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 1138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767795.387, "dur": 2.563,
    "args": {
      "External id": 2484340,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767803.355, "dur": 29.249,
    "args": {
      "External id": 2484341,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], []], "Ev Idx": 1140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767839.860, "dur": 5.288,
    "args": {
      "External id": 2484342,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 1141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767842.820, "dur": 2.034,
    "args": {
      "External id": 2484343,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767935.861, "dur": 17.488,
    "args": {
      "External id": 2484344,"Record function id": 0, "Sequence number": 24740527, "Fwd thread id": 1, "Ev Idx": 1143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767937.240, "dur": 13.939,
    "args": {
      "External id": 2484345,"Sequence number": 24740527, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 1144
    }
  },
  {
    "ph": "f", "id": 121, "pid": 1336759, "tid": 1381173, "ts": 1514656767937.240,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767939.169, "dur": 11.598,
    "args": {
      "External id": 2484346,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 1145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767943.276, "dur": 7.266,
    "args": {
      "External id": 2484347,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 1146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767956.989, "dur": 7.624,
    "args": {
      "External id": 2484348,"Record function id": 0, "Sequence number": 24740526, "Fwd thread id": 1, "Ev Idx": 1147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767957.921, "dur": 4.903,
    "args": {
      "External id": 2484349,"Sequence number": 24740526, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 1148
    }
  },
  {
    "ph": "f", "id": 122, "pid": 1336759, "tid": 1381173, "ts": 1514656767957.921,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767958.882, "dur": 3.810,
    "args": {
      "External id": 2484350,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 1149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767961.644, "dur": 0.938,
    "args": {
      "External id": 2484351,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 1150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767967.746, "dur": 4.914,
    "args": {
      "External id": 2484352,"Record function id": 0, "Sequence number": 24740525, "Fwd thread id": 1, "Ev Idx": 1151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767968.422, "dur": 2.573,
    "args": {
      "External id": 2484353,"Sequence number": 24740525, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 1152
    }
  },
  {
    "ph": "f", "id": 123, "pid": 1336759, "tid": 1381173, "ts": 1514656767968.422,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767969.321, "dur": 1.532,
    "args": {
      "External id": 2484354,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 1153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767969.904, "dur": 0.859,
    "args": {
      "External id": 2484355,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 1154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767976.262, "dur": 46.303,
    "args": {
      "External id": 2484356,"Record function id": 0, "Sequence number": 24740524, "Fwd thread id": 1, "Ev Idx": 1155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767977.028, "dur": 6.177,
    "args": {
      "External id": 2484357,"Sequence number": 24740524, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 1156
    }
  },
  {
    "ph": "f", "id": 124, "pid": 1336759, "tid": 1381173, "ts": 1514656767977.028,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767977.961, "dur": 5.111,
    "args": {
      "External id": 2484358,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656767982.127, "dur": 0.862,
    "args": {
      "External id": 2484359,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768028.894, "dur": 149.078,
    "args": {
      "External id": 2484360,"Record function id": 0, "Sequence number": 24740523, "Fwd thread id": 1, "Ev Idx": 1159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768029.947, "dur": 141.612,
    "args": {
      "External id": 2484361,"Sequence number": 24740523, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 1160
    }
  },
  {
    "ph": "f", "id": 125, "pid": 1336759, "tid": 1381173, "ts": 1514656768029.947,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768033.803, "dur": 7.048,
    "args": {
      "External id": 2484362,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 1161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768035.901, "dur": 4.157,
    "args": {
      "External id": 2484363,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[65536, 2048], [], []], "Ev Idx": 1162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768037.955, "dur": 1.860,
    "args": {
      "External id": 2484364,"Record function id": 0, "Concrete Inputs": ["", "[2048, 65536]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 1163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768045.023, "dur": 67.985,
    "args": {
      "External id": 2484365,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048]], "Ev Idx": 1164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768114.161, "dur": 3.863,
    "args": {
      "External id": 2484366,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768114.809, "dur": 2.485,
    "args": {
      "External id": 2484367,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 1166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768115.914, "dur": 1.251,
    "args": {
      "External id": 2484368,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 1167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768119.707, "dur": 4.802,
    "args": {
      "External id": 2484369,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 1168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768120.349, "dur": 3.631,
    "args": {
      "External id": 2484370,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 1169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768123.275, "dur": 0.609,
    "args": {
      "External id": 2484371,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 1170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768127.000, "dur": 43.643,
    "args": {
      "External id": 2484372,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 1171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768184.801, "dur": 6.430,
    "args": {
      "External id": 2484373,"Record function id": 0, "Sequence number": 24740522, "Fwd thread id": 1, "Ev Idx": 1172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768185.790, "dur": 3.794,
    "args": {
      "External id": 2484374,"Sequence number": 24740522, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 1173
    }
  },
  {
    "ph": "f", "id": 126, "pid": 1336759, "tid": 1381173, "ts": 1514656768185.790,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768187.103, "dur": 2.320,
    "args": {
      "External id": 2484375,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768187.705, "dur": 1.588,
    "args": {
      "External id": 2484376,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768194.446, "dur": 8.408,
    "args": {
      "External id": 2484377,"Record function id": 0, "Sequence number": 24740521, "Fwd thread id": 1, "Ev Idx": 1176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768195.407, "dur": 5.472,
    "args": {
      "External id": 2484378,"Sequence number": 24740521, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 1177
    }
  },
  {
    "ph": "f", "id": 127, "pid": 1336759, "tid": 1381173, "ts": 1514656768195.407,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768196.136, "dur": 4.540,
    "args": {
      "External id": 2484379,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 1178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768196.855, "dur": 3.375,
    "args": {
      "External id": 2484380,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 1179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768199.763, "dur": 0.372,
    "args": {
      "External id": 2484381,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 1180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768210.632, "dur": 9.035,
    "args": {
      "External id": 2484382,"Record function id": 0, "Ev Idx": 1181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768212.069, "dur": 6.816,
    "args": {
      "External id": 2484383,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768214.436, "dur": 4.107,
    "args": {
      "External id": 2484384,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768215.916, "dur": 2.510,
    "args": {
      "External id": 2484385,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768222.782, "dur": 5.350,
    "args": {
      "External id": 2484386,"Record function id": 0, "Sequence number": 24740520, "Fwd thread id": 1, "Ev Idx": 1185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768224.103, "dur": 2.552,
    "args": {
      "External id": 2484387,"Sequence number": 24740520, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 1186
    }
  },
  {
    "ph": "f", "id": 128, "pid": 1336759, "tid": 1381173, "ts": 1514656768224.103,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768224.879, "dur": 1.617,
    "args": {
      "External id": 2484388,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768225.659, "dur": 0.716,
    "args": {
      "External id": 2484389,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768231.247, "dur": 88.657,
    "args": {
      "External id": 2484390,"Record function id": 0, "Sequence number": 24740519, "Fwd thread id": 1, "Ev Idx": 1189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768231.934, "dur": 82.362,
    "args": {
      "External id": 2484391,"Sequence number": 24740519, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 1190
    }
  },
  {
    "ph": "f", "id": 129, "pid": 1336759, "tid": 1381173, "ts": 1514656768231.934,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768236.283, "dur": 2.852,
    "args": {
      "External id": 2484392,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 1191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768237.042, "dur": 1.682,
    "args": {
      "External id": 2484393,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[65536, 2048], [], []], "Ev Idx": 1192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768237.855, "dur": 0.755,
    "args": {
      "External id": 2484394,"Record function id": 0, "Concrete Inputs": ["", "[2048, 65536]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 1193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768239.837, "dur": 25.215,
    "args": {
      "External id": 2484395,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048]], "Ev Idx": 1194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768266.152, "dur": 6.754,
    "args": {
      "External id": 2484396,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768266.799, "dur": 5.611,
    "args": {
      "External id": 2484397,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 1196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768271.724, "dur": 0.574,
    "args": {
      "External id": 2484398,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 1197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768273.990, "dur": 4.650,
    "args": {
      "External id": 2484399,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 1198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768275.122, "dur": 3.133,
    "args": {
      "External id": 2484400,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 1199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768275.841, "dur": 2.277,
    "args": {
      "External id": 2484401,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 1200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768279.132, "dur": 34.363,
    "args": {
      "External id": 2484402,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 1201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768323.972, "dur": 33.693,
    "args": {
      "External id": 2484403,"Record function id": 0, "Sequence number": 24740518, "Fwd thread id": 1, "Ev Idx": 1202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768324.852, "dur": 7.231,
    "args": {
      "External id": 2484404,"Sequence number": 24740518, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 1203
    }
  },
  {
    "ph": "f", "id": 130, "pid": 1336759, "tid": 1381173, "ts": 1514656768324.852,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768327.947, "dur": 3.998,
    "args": {
      "External id": 2484405,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768330.524, "dur": 1.302,
    "args": {
      "External id": 2484406,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768334.783, "dur": 20.679,
    "args": {
      "External id": 2484407,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 1206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768363.543, "dur": 6.658,
    "args": {
      "External id": 2484408,"Record function id": 0, "Sequence number": 24740517, "Fwd thread id": 1, "Ev Idx": 1207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768364.475, "dur": 3.944,
    "args": {
      "External id": 2484409,"Sequence number": 24740517, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 1208
    }
  },
  {
    "ph": "f", "id": 131, "pid": 1336759, "tid": 1381173, "ts": 1514656768364.475,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768365.495, "dur": 2.736,
    "args": {
      "External id": 2484410,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 1209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768366.167, "dur": 1.471,
    "args": {
      "External id": 2484411,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 1210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768367.074, "dur": 0.407,
    "args": {
      "External id": 2484412,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 1211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768374.147, "dur": 6.065,
    "args": {
      "External id": 2484413,"Record function id": 0, "Ev Idx": 1212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768375.248, "dur": 4.488,
    "args": {
      "External id": 2484414,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768376.036, "dur": 3.422,
    "args": {
      "External id": 2484415,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768378.349, "dur": 1.032,
    "args": {
      "External id": 2484416,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768383.354, "dur": 9.530,
    "args": {
      "External id": 2484417,"Record function id": 0, "Sequence number": 24740516, "Fwd thread id": 1, "Ev Idx": 1216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768384.284, "dur": 7.050,
    "args": {
      "External id": 2484418,"Sequence number": 24740516, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 1217
    }
  },
  {
    "ph": "f", "id": 132, "pid": 1336759, "tid": 1381173, "ts": 1514656768384.284,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768385.036, "dur": 6.158,
    "args": {
      "External id": 2484419,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768390.319, "dur": 0.774,
    "args": {
      "External id": 2484420,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768395.938, "dur": 87.406,
    "args": {
      "External id": 2484421,"Record function id": 0, "Sequence number": 24740515, "Fwd thread id": 1, "Ev Idx": 1220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768396.735, "dur": 79.959,
    "args": {
      "External id": 2484422,"Sequence number": 24740515, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 1221
    }
  },
  {
    "ph": "f", "id": 133, "pid": 1336759, "tid": 1381173, "ts": 1514656768396.735,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768400.292, "dur": 2.194,
    "args": {
      "External id": 2484423,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 1222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768400.718, "dur": 1.345,
    "args": {
      "External id": 2484424,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[65536, 2048], [], []], "Ev Idx": 1223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768401.378, "dur": 0.593,
    "args": {
      "External id": 2484425,"Record function id": 0, "Concrete Inputs": ["", "[2048, 65536]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 1224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768403.054, "dur": 26.860,
    "args": {
      "External id": 2484426,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048]], "Ev Idx": 1225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768430.820, "dur": 6.590,
    "args": {
      "External id": 2484427,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768431.413, "dur": 5.385,
    "args": {
      "External id": 2484428,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 1227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768434.533, "dur": 2.072,
    "args": {
      "External id": 2484429,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 1228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768438.490, "dur": 5.063,
    "args": {
      "External id": 2484430,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 1229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768439.615, "dur": 3.574,
    "args": {
      "External id": 2484431,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 1230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768442.719, "dur": 0.373,
    "args": {
      "External id": 2484432,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 1231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768444.171, "dur": 31.719,
    "args": {
      "External id": 2484433,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 1232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768487.385, "dur": 26.514,
    "args": {
      "External id": 2484434,"Record function id": 0, "Sequence number": 24740514, "Fwd thread id": 1, "Ev Idx": 1233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768488.468, "dur": 3.468,
    "args": {
      "External id": 2484435,"Sequence number": 24740514, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 1234
    }
  },
  {
    "ph": "f", "id": 134, "pid": 1336759, "tid": 1381173, "ts": 1514656768488.468,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768489.847, "dur": 1.942,
    "args": {
      "External id": 2484436,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768490.466, "dur": 1.180,
    "args": {
      "External id": 2484437,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768494.659, "dur": 17.477,
    "args": {
      "External id": 2484438,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 1237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768519.680, "dur": 7.137,
    "args": {
      "External id": 2484439,"Record function id": 0, "Sequence number": 24740513, "Fwd thread id": 1, "Ev Idx": 1238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768520.728, "dur": 4.256,
    "args": {
      "External id": 2484440,"Sequence number": 24740513, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 1239
    }
  },
  {
    "ph": "f", "id": 135, "pid": 1336759, "tid": 1381173, "ts": 1514656768520.728,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768521.653, "dur": 3.107,
    "args": {
      "External id": 2484441,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 1240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768522.420, "dur": 1.821,
    "args": {
      "External id": 2484442,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 1241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768523.409, "dur": 0.728,
    "args": {
      "External id": 2484443,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 1242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768556.833, "dur": 9.535,
    "args": {
      "External id": 2484444,"Record function id": 0, "Ev Idx": 1243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768561.644, "dur": 4.028,
    "args": {
      "External id": 2484445,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768562.928, "dur": 2.376,
    "args": {
      "External id": 2484446,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768563.726, "dur": 1.507,
    "args": {
      "External id": 2484447,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: LayerNormFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768572.559, "dur": 331.658,
    "args": {
      "External id": 2484448,"Record function id": 0, "Sequence number": 24740512, "Fwd thread id": 1, "Ev Idx": 1247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768573.802, "dur": 303.228,
    "args": {
      "External id": 2484449,"Sequence number": 24740512, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 1248
    }
  },
  {
    "ph": "f", "id": 136, "pid": 1336759, "tid": 1381173, "ts": 1514656768573.802,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768607.151, "dur": 2.204,
    "args": {
      "External id": 2484450,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768607.685, "dur": 1.538,
    "args": {
      "External id": 2484451,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768622.119, "dur": 6.803,
    "args": {
      "External id": 2484452,"Record function id": 0, "Concrete Inputs": ["[65536, 2048]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768637.460, "dur": 1.676,
    "args": {
      "External id": 2484453,"Record function id": 0, "Concrete Inputs": ["[132, 2048]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768779.394, "dur": 1.912,
    "args": {
      "External id": 2484454,"Record function id": 0, "Concrete Inputs": ["", "[1, -1, 2048]"], "Input type": ["float", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[132, 2048], []], "Ev Idx": 1253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768785.575, "dur": 31.266,
    "args": {
      "External id": 2484455,"Record function id": 0, "Concrete Inputs": ["", "[1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[270336, 2048, 1], [], [], []], "Input Dims": [[1, 132, 2048], [], [], []], "Ev Idx": 1254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768793.715, "dur": 0.958,
    "args": {
      "External id": 2484456,"Record function id": 0, "Concrete Inputs": ["", "[1, 1, 2048]", "[2048, 0, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[1, 2048], [], [], []], "Ev Idx": 1255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768822.024, "dur": 32.524,
    "args": {
      "External id": 2484457,"Record function id": 0, "Concrete Inputs": ["", "", "15", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[1, 2048], [], [], [], [], []], "Ev Idx": 1256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768825.953, "dur": 28.347,
    "args": {
      "External id": 2484458,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], [], []], "Input Dims": [[1, 2048], [], [], [], [], [], []], "Ev Idx": 1257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768831.155, "dur": 5.344,
    "args": {
      "External id": 2484459,"Record function id": 0, "Concrete Inputs": ["[1, 2048]", "[2048, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768837.964, "dur": 15.884,
    "args": {
      "External id": 2484460,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "float", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[1, 2048], [1, 2048], []], "Ev Idx": 1259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768859.041, "dur": 2.165,
    "args": {
      "External id": 2484461,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1]], "Input Dims": [[1, 2048], [2048]], "Ev Idx": 1260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768860.097, "dur": 0.981,
    "args": {
      "External id": 2484462,"Record function id": 0, "Concrete Inputs": ["", "[2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[1, 2048], []], "Ev Idx": 1261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768866.535, "dur": 4.051,
    "args": {
      "External id": 2484463,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768869.475, "dur": 1.023,
    "args": {
      "External id": 2484464,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768885.828, "dur": 14.674,
    "args": {
      "External id": 2484465,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 1264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768912.671, "dur": 7.117,
    "args": {
      "External id": 2484466,"Record function id": 0, "Ev Idx": 1265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768914.615, "dur": 4.558,
    "args": {
      "External id": 2484467,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 1266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768916.142, "dur": 2.033,
    "args": {
      "External id": 2484468,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 1267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768916.810, "dur": 1.276,
    "args": {
      "External id": 2484469,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 1268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: AddBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768926.603, "dur": 4.748,
    "args": {
      "External id": 2484470,"Record function id": 0, "Sequence number": 24740511, "Fwd thread id": 1, "Ev Idx": 1269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "AddBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768927.788, "dur": 1.059,
    "args": {
      "External id": 2484471,"Sequence number": 24740511, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[33554432, 8192, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 1270
    }
  },
  {
    "ph": "f", "id": 137, "pid": 1336759, "tid": 1381173, "ts": 1514656768927.788,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: SwiGLULinearFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768934.683, "dur": 449.422,
    "args": {
      "External id": 2484472,"Record function id": 0, "Sequence number": 24740510, "Fwd thread id": 1, "Ev Idx": 1271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SwiGLULinearFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768937.503, "dur": 435.443,
    "args": {
      "External id": 2484473,"Sequence number": 24740510, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[33554432, 8192, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 1272
    }
  },
  {
    "ph": "f", "id": 138, "pid": 1336759, "tid": 1381173, "ts": 1514656768937.503,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768967.434, "dur": 7.716,
    "args": {
      "External id": 2484474,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[33554432, 8192, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768971.941, "dur": 2.969,
    "args": {
      "External id": 2484475,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]", "[8192, 1]"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList"], "Input Strides": [[33554432, 8192, 1], [], []], "Input Dims": [[16, 4096, 2048], [], []], "Ev Idx": 1274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768978.103, "dur": 47.156,
    "args": {
      "External id": 2484476,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 1275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768979.441, "dur": 4.345,
    "args": {
      "External id": 2484477,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[5632, 1], [], []], "Input Dims": [[2048, 5632], [], []], "Ev Idx": 1276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656768980.690, "dur": 2.947,
    "args": {
      "External id": 2484478,"Record function id": 0, "Concrete Inputs": ["", "[5632, 2048]", "[1, 5632]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[5632, 1], [], [], []], "Input Dims": [[2048, 5632], [], [], []], "Ev Idx": 1277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769030.736, "dur": 92.167,
    "args": {
      "External id": 2484479,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[8192, 1], [1, 5632], []], "Input Dims": [[65536, 2048], [5632, 2048], []], "Ev Idx": 1278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769033.636, "dur": 5.380,
    "args": {
      "External id": 2484480,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 5632]], "Input Dims": [[5632, 2048]], "Ev Idx": 1279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769034.658, "dur": 3.663,
    "args": {
      "External id": 2484481,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 5632], [], []], "Input Dims": [[5632, 2048], [], []], "Ev Idx": 1280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769037.508, "dur": 0.692,
    "args": {
      "External id": 2484482,"Record function id": 0, "Concrete Inputs": ["", "[2048, 5632]", "[5632, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 5632], [], [], []], "Input Dims": [[5632, 2048], [], [], []], "Ev Idx": 1281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769040.315, "dur": 81.916,
    "args": {
      "External id": 2484483,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8192, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632]], "Ev Idx": 1282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769041.915, "dur": 79.439,
    "args": {
      "External id": 2484484,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8192, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632]], "Ev Idx": 1283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769126.531, "dur": 2.705,
    "args": {
      "External id": 2484485,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [23068672, 5632, 1]], "Input Dims": [[65536, 5632], [16, 4096, 5632]], "Ev Idx": 1284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769127.522, "dur": 1.614,
    "args": {
      "External id": 2484486,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 1], []], "Input Dims": [[65536, 5632], []], "Ev Idx": 1285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769161.343, "dur": 5.896,
    "args": {
      "External id": 2484487,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 5632]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769170.362, "dur": 1.538,
    "args": {
      "External id": 2484488,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 5632]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769172.637, "dur": 1.730,
    "args": {
      "External id": 2484489,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 5632]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769205.984, "dur": 2.269,
    "args": {
      "External id": 2484490,"Record function id": 0, "Concrete Inputs": ["", "[-1, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 1289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769206.642, "dur": 1.462,
    "args": {
      "External id": 2484491,"Record function id": 0, "Concrete Inputs": ["", "[-1, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 1290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::einsum", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769229.882, "dur": 123.106,
    "args": {
      "External id": 2484492,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["", "TensorList", ""], "Input Strides": [[], [[8192, 1], [5632, 1]], []], "Input Dims": [[], [[65536, 2048], [65536, 5632]], []], "Ev Idx": 1291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769235.088, "dur": 10.624,
    "args": {
      "External id": 2484493,"Record function id": 0, "Concrete Inputs": ["", "2"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769240.181, "dur": 4.449,
    "args": {
      "External id": 2484494,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048, 1]", "[8192, 1, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[8192, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 1293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769247.225, "dur": 6.001,
    "args": {
      "External id": 2484495,"Record function id": 0, "Concrete Inputs": ["", "[1, 2, 0]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8192, 1, 1], []], "Input Dims": [[65536, 2048, 1], []], "Ev Idx": 1294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769251.632, "dur": 0.633,
    "args": {
      "External id": 2484496,"Record function id": 0, "Concrete Inputs": ["", "[2048, 1, 65536]", "[1, 1, 8192]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[65536, 2048, 1], [], [], []], "Ev Idx": 1295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769254.939, "dur": 1.983,
    "args": {
      "External id": 2484497,"Record function id": 0, "Concrete Inputs": ["", "2"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], []], "Input Dims": [[65536, 5632], []], "Ev Idx": 1296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769255.962, "dur": 0.645,
    "args": {
      "External id": 2484498,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632, 1]", "[5632, 1, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[5632, 1], [], [], []], "Input Dims": [[65536, 5632], [], [], []], "Ev Idx": 1297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769259.286, "dur": 2.085,
    "args": {
      "External id": 2484499,"Record function id": 0, "Concrete Inputs": ["", "[2, 1, 0]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 1, 1], []], "Input Dims": [[65536, 5632, 1], []], "Ev Idx": 1298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769260.119, "dur": 0.786,
    "args": {
      "External id": 2484500,"Record function id": 0, "Concrete Inputs": ["", "[1, 5632, 65536]", "[1, 1, 5632]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[5632, 1, 1], [], [], []], "Input Dims": [[65536, 5632, 1], [], [], []], "Ev Idx": 1299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769264.724, "dur": 3.807,
    "args": {
      "External id": 2484501,"Record function id": 0, "Concrete Inputs": ["", "[0, 2, 1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1, 1, 8192], []], "Input Dims": [[2048, 1, 65536], []], "Ev Idx": 1300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769267.615, "dur": 0.596,
    "args": {
      "External id": 2484502,"Record function id": 0, "Concrete Inputs": ["", "[2048, 65536, 1]", "[1, 8192, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1, 8192], [], [], []], "Input Dims": [[2048, 1, 65536], [], [], []], "Ev Idx": 1301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769269.316, "dur": 7.585,
    "args": {
      "External id": 2484503,"Record function id": 0, "Concrete Inputs": ["", "[1, 2048, 65536]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1, 8192, 1], []], "Input Dims": [[2048, 65536, 1], []], "Ev Idx": 1302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_reshape_alias", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769274.515, "dur": 2.206,
    "args": {
      "External id": 2484504,"Record function id": 0, "Concrete Inputs": ["", "[1, 2048, 65536]", "[2048, 1, 8192]"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList"], "Input Strides": [[1, 8192, 1], [], []], "Input Dims": [[2048, 65536, 1], [], []], "Ev Idx": 1303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769279.790, "dur": 3.527,
    "args": {
      "External id": 2484505,"Record function id": 0, "Concrete Inputs": ["", "[2, 1, 0]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1, 1, 5632], []], "Input Dims": [[1, 5632, 65536], []], "Ev Idx": 1304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769282.718, "dur": 0.337,
    "args": {
      "External id": 2484506,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632, 1]", "[5632, 1, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1, 5632], [], [], []], "Input Dims": [[1, 5632, 65536], [], [], []], "Ev Idx": 1305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769283.846, "dur": 3.169,
    "args": {
      "External id": 2484507,"Record function id": 0, "Concrete Inputs": ["", "[1, 65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 1, 1], []], "Input Dims": [[65536, 5632, 1], []], "Ev Idx": 1306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769284.328, "dur": 2.607,
    "args": {
      "External id": 2484508,"Record function id": 0, "Concrete Inputs": ["", "[1, 65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 1, 1], []], "Input Dims": [[65536, 5632, 1], []], "Ev Idx": 1307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769288.287, "dur": 52.055,
    "args": {
      "External id": 2484509,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1, 8192], [369098752, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632]], "Ev Idx": 1308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769343.761, "dur": 1.283,
    "args": {
      "External id": 2484510,"Record function id": 0, "Concrete Inputs": ["", "[2048, 1, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[11534336, 5632, 1], []], "Input Dims": [[1, 2048, 5632], []], "Ev Idx": 1309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::permute", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769345.545, "dur": 1.857,
    "args": {
      "External id": 2484511,"Record function id": 0, "Concrete Inputs": ["", "[0, 2, 1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 5632, 1], []], "Input Dims": [[2048, 1, 5632], []], "Ev Idx": 1310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769346.496, "dur": 0.528,
    "args": {
      "External id": 2484512,"Record function id": 0, "Concrete Inputs": ["", "[2048, 5632, 1]", "[5632, 1, 5632]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[5632, 5632, 1], [], [], []], "Input Dims": [[2048, 1, 5632], [], [], []], "Ev Idx": 1311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769351.127, "dur": 0.760,
    "args": {
      "External id": 2484513,"Record function id": 0, "Concrete Inputs": ["", "[2048, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 1, 5632], []], "Input Dims": [[2048, 5632, 1], []], "Ev Idx": 1312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769395.215, "dur": 9.351,
    "args": {
      "External id": 2484514,"Record function id": 0, "Ev Idx": 1313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769397.435, "dur": 6.459,
    "args": {
      "External id": 2484515,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 1314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769399.621, "dur": 3.339,
    "args": {
      "External id": 2484516,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 1315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769400.634, "dur": 2.209,
    "args": {
      "External id": 2484517,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 1316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769408.182, "dur": 5.652,
    "args": {
      "External id": 2484518,"Record function id": 0, "Sequence number": 24740509, "Fwd thread id": 1, "Ev Idx": 1317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769409.364, "dur": 2.784,
    "args": {
      "External id": 2484519,"Sequence number": 24740509, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[23068672, 5632, 1]], "Input Dims": [[16, 4096, 5632]], "Ev Idx": 1318
    }
  },
  {
    "ph": "f", "id": 139, "pid": 1336759, "tid": 1381173, "ts": 1514656769409.364,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769410.671, "dur": 1.279,
    "args": {
      "External id": 2484520,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 1319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769411.136, "dur": 0.691,
    "args": {
      "External id": 2484521,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 1320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769417.071, "dur": 140.271,
    "args": {
      "External id": 2484522,"Record function id": 0, "Sequence number": 24740508, "Fwd thread id": 1, "Ev Idx": 1321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769418.239, "dur": 131.667,
    "args": {
      "External id": 2484523,"Sequence number": 24740508, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[65536, 5632]], "Ev Idx": 1322
    }
  },
  {
    "ph": "f", "id": 140, "pid": 1336759, "tid": 1381173, "ts": 1514656769418.239,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769421.156, "dur": 6.286,
    "args": {
      "External id": 2484524,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[65536, 5632]], "Ev Idx": 1323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769422.507, "dur": 4.413,
    "args": {
      "External id": 2484525,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[5632, 1], [], []], "Input Dims": [[65536, 5632], [], []], "Ev Idx": 1324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769425.818, "dur": 0.974,
    "args": {
      "External id": 2484526,"Record function id": 0, "Concrete Inputs": ["", "[5632, 65536]", "[1, 5632]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[5632, 1], [], [], []], "Input Dims": [[65536, 5632], [], [], []], "Ev Idx": 1325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769428.468, "dur": 39.641,
    "args": {
      "External id": 2484527,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048]], "Ev Idx": 1326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769469.325, "dur": 5.230,
    "args": {
      "External id": 2484528,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769469.929, "dur": 4.000,
    "args": {
      "External id": 2484529,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[5632, 2048], [], []], "Ev Idx": 1328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769472.773, "dur": 1.000,
    "args": {
      "External id": 2484530,"Record function id": 0, "Concrete Inputs": ["", "[2048, 5632]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[5632, 2048], [], [], []], "Ev Idx": 1329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769476.142, "dur": 6.785,
    "args": {
      "External id": 2484531,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 5632]], "Ev Idx": 1330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769478.981, "dur": 3.532,
    "args": {
      "External id": 2484532,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 5632], [], []], "Ev Idx": 1331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769479.799, "dur": 2.596,
    "args": {
      "External id": 2484533,"Record function id": 0, "Concrete Inputs": ["", "[5632, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 5632], [], [], []], "Ev Idx": 1332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769483.640, "dur": 64.904,
    "args": {
      "External id": 2484534,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048]], "Ev Idx": 1333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769564.084, "dur": 6.075,
    "args": {
      "External id": 2484535,"Record function id": 0, "Sequence number": 24740507, "Fwd thread id": 1, "Ev Idx": 1334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769565.137, "dur": 3.905,
    "args": {
      "External id": 2484536,"Sequence number": 24740507, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 1335
    }
  },
  {
    "ph": "f", "id": 141, "pid": 1336759, "tid": 1381173, "ts": 1514656769565.137,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769566.604, "dur": 2.288,
    "args": {
      "External id": 2484537,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769567.109, "dur": 1.667,
    "args": {
      "External id": 2484538,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769573.558, "dur": 10.646,
    "args": {
      "External id": 2484539,"Record function id": 0, "Sequence number": 24740506, "Fwd thread id": 1, "Ev Idx": 1338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769574.344, "dur": 7.863,
    "args": {
      "External id": 2484540,"Sequence number": 24740506, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 5632]], "Ev Idx": 1339
    }
  },
  {
    "ph": "f", "id": 142, "pid": 1336759, "tid": 1381173, "ts": 1514656769574.344,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769576.882, "dur": 5.123,
    "args": {
      "External id": 2484541,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 5632]], "Ev Idx": 1340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769577.629, "dur": 3.852,
    "args": {
      "External id": 2484542,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 5632], [], []], "Ev Idx": 1341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769580.613, "dur": 0.698,
    "args": {
      "External id": 2484543,"Record function id": 0, "Concrete Inputs": ["", "[5632, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 5632], [], [], []], "Ev Idx": 1342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769588.297, "dur": 5.418,
    "args": {
      "External id": 2484544,"Record function id": 0, "Ev Idx": 1343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769589.577, "dur": 3.607,
    "args": {
      "External id": 2484545,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769590.668, "dur": 2.234,
    "args": {
      "External id": 2484546,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769591.295, "dur": 1.524,
    "args": {
      "External id": 2484547,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769596.686, "dur": 7.224,
    "args": {
      "External id": 2484548,"Record function id": 0, "Sequence number": 24740505, "Fwd thread id": 1, "Ev Idx": 1347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769597.741, "dur": 4.524,
    "args": {
      "External id": 2484549,"Sequence number": 24740505, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[23068672, 5632, 1]], "Input Dims": [[16, 4096, 5632]], "Ev Idx": 1348
    }
  },
  {
    "ph": "f", "id": 143, "pid": 1336759, "tid": 1381173, "ts": 1514656769597.741,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769598.814, "dur": 3.311,
    "args": {
      "External id": 2484550,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 1349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769601.422, "dur": 0.582,
    "args": {
      "External id": 2484551,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 1350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769606.903, "dur": 111.513,
    "args": {
      "External id": 2484552,"Record function id": 0, "Sequence number": 24740504, "Fwd thread id": 1, "Ev Idx": 1351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769607.621, "dur": 104.039,
    "args": {
      "External id": 2484553,"Sequence number": 24740504, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[65536, 5632]], "Ev Idx": 1352
    }
  },
  {
    "ph": "f", "id": 144, "pid": 1336759, "tid": 1381173, "ts": 1514656769607.621,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769609.582, "dur": 3.972,
    "args": {
      "External id": 2484554,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[65536, 5632]], "Ev Idx": 1353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769609.948, "dur": 3.193,
    "args": {
      "External id": 2484555,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[5632, 1], [], []], "Input Dims": [[65536, 5632], [], []], "Ev Idx": 1354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769612.587, "dur": 0.420,
    "args": {
      "External id": 2484556,"Record function id": 0, "Concrete Inputs": ["", "[5632, 65536]", "[1, 5632]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[5632, 1], [], [], []], "Input Dims": [[65536, 5632], [], [], []], "Ev Idx": 1355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769614.268, "dur": 36.897,
    "args": {
      "External id": 2484557,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048]], "Ev Idx": 1356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769653.979, "dur": 5.042,
    "args": {
      "External id": 2484558,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769654.783, "dur": 3.709,
    "args": {
      "External id": 2484559,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[5632, 2048], [], []], "Ev Idx": 1358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769655.662, "dur": 2.682,
    "args": {
      "External id": 2484560,"Record function id": 0, "Concrete Inputs": ["", "[2048, 5632]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[5632, 2048], [], [], []], "Ev Idx": 1359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769660.331, "dur": 2.601,
    "args": {
      "External id": 2484561,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 5632]], "Ev Idx": 1360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769661.233, "dur": 1.345,
    "args": {
      "External id": 2484562,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 5632], [], []], "Ev Idx": 1361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769661.987, "dur": 0.523,
    "args": {
      "External id": 2484563,"Record function id": 0, "Concrete Inputs": ["", "[5632, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 5632], [], [], []], "Ev Idx": 1362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769665.967, "dur": 45.015,
    "args": {
      "External id": 2484564,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048]], "Ev Idx": 1363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769724.725, "dur": 33.416,
    "args": {
      "External id": 2484565,"Record function id": 0, "Sequence number": 24740503, "Fwd thread id": 1, "Ev Idx": 1364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769725.785, "dur": 5.620,
    "args": {
      "External id": 2484566,"Sequence number": 24740503, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 1365
    }
  },
  {
    "ph": "f", "id": 145, "pid": 1336759, "tid": 1381173, "ts": 1514656769725.785,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769729.318, "dur": 1.947,
    "args": {
      "External id": 2484567,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769730.097, "dur": 1.053,
    "args": {
      "External id": 2484568,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769734.420, "dur": 20.580,
    "args": {
      "External id": 2484569,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 1368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769761.826, "dur": 9.355,
    "args": {
      "External id": 2484570,"Record function id": 0, "Sequence number": 24740502, "Fwd thread id": 1, "Ev Idx": 1369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769763.217, "dur": 6.365,
    "args": {
      "External id": 2484571,"Sequence number": 24740502, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 5632]], "Ev Idx": 1370
    }
  },
  {
    "ph": "f", "id": 146, "pid": 1336759, "tid": 1381173, "ts": 1514656769763.217,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769763.869, "dur": 5.499,
    "args": {
      "External id": 2484572,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 5632]], "Ev Idx": 1371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769764.829, "dur": 4.051,
    "args": {
      "External id": 2484573,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 5632], [], []], "Ev Idx": 1372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769768.225, "dur": 0.560,
    "args": {
      "External id": 2484574,"Record function id": 0, "Concrete Inputs": ["", "[5632, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 5632], [], [], []], "Ev Idx": 1373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769775.078, "dur": 7.021,
    "args": {
      "External id": 2484575,"Record function id": 0, "Ev Idx": 1374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769778.529, "dur": 3.057,
    "args": {
      "External id": 2484576,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769779.558, "dur": 1.664,
    "args": {
      "External id": 2484577,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769780.078, "dur": 1.038,
    "args": {
      "External id": 2484578,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: LayerNormFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769785.972, "dur": 459.527,
    "args": {
      "External id": 2484579,"Record function id": 0, "Sequence number": 24740501, "Fwd thread id": 1, "Ev Idx": 1378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769790.974, "dur": 417.812,
    "args": {
      "External id": 2484580,"Sequence number": 24740501, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [33554432, 8192, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 1379
    }
  },
  {
    "ph": "f", "id": 147, "pid": 1336759, "tid": 1381173, "ts": 1514656769790.974,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769813.452, "dur": 31.475,
    "args": {
      "External id": 2484581,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769814.933, "dur": 29.796,
    "args": {
      "External id": 2484582,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[33554432, 8192, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769817.669, "dur": 5.413,
    "args": {
      "External id": 2484583,"Record function id": 0, "Concrete Inputs": ["", "15", "0", "", "", "0"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[33554432, 8192, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 2048], [], [], [], [], []], "Ev Idx": 1382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769819.903, "dur": 2.694,
    "args": {
      "External id": 2484584,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 2048]", "15", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769824.449, "dur": 19.767,
    "args": {
      "External id": 2484585,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [33554432, 8192, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 1384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769857.107, "dur": 6.207,
    "args": {
      "External id": 2484586,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769859.698, "dur": 3.489,
    "args": {
      "External id": 2484587,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769866.747, "dur": 1.754,
    "args": {
      "External id": 2484588,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769867.656, "dur": 0.712,
    "args": {
      "External id": 2484589,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769881.984, "dur": 2.602,
    "args": {
      "External id": 2484590,"Record function id": 0, "Concrete Inputs": ["[65536, 2048]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656769894.949, "dur": 2.886,
    "args": {
      "External id": 2484591,"Record function id": 0, "Concrete Inputs": ["[132, 2048]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770094.276, "dur": 5.143,
    "args": {
      "External id": 2484592,"Record function id": 0, "Concrete Inputs": ["", "[1, -1, 2048]"], "Input type": ["float", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[132, 2048], []], "Ev Idx": 1391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770103.601, "dur": 34.781,
    "args": {
      "External id": 2484593,"Record function id": 0, "Concrete Inputs": ["", "[1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[270336, 2048, 1], [], [], []], "Input Dims": [[1, 132, 2048], [], [], []], "Ev Idx": 1392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770114.580, "dur": 0.952,
    "args": {
      "External id": 2484594,"Record function id": 0, "Concrete Inputs": ["", "[1, 1, 2048]", "[2048, 0, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[1, 2048], [], [], []], "Ev Idx": 1393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770143.942, "dur": 29.066,
    "args": {
      "External id": 2484595,"Record function id": 0, "Concrete Inputs": ["", "", "15", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[1, 2048], [], [], [], [], []], "Ev Idx": 1394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770145.806, "dur": 26.889,
    "args": {
      "External id": 2484596,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], [], []], "Input Dims": [[1, 2048], [], [], [], [], [], []], "Ev Idx": 1395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770149.642, "dur": 3.690,
    "args": {
      "External id": 2484597,"Record function id": 0, "Concrete Inputs": ["[1, 2048]", "[2048, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770156.807, "dur": 15.417,
    "args": {
      "External id": 2484598,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "float", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[1, 2048], [1, 2048], []], "Ev Idx": 1397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770177.333, "dur": 4.185,
    "args": {
      "External id": 2484599,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1]], "Input Dims": [[1, 2048], [2048]], "Ev Idx": 1398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770180.366, "dur": 1.020,
    "args": {
      "External id": 2484600,"Record function id": 0, "Concrete Inputs": ["", "[2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[1, 2048], []], "Ev Idx": 1399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770187.499, "dur": 4.211,
    "args": {
      "External id": 2484601,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770188.287, "dur": 3.304,
    "args": {
      "External id": 2484602,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770193.524, "dur": 4.011,
    "args": {
      "External id": 2484603,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770196.405, "dur": 1.047,
    "args": {
      "External id": 2484604,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770227.544, "dur": 16.515,
    "args": {
      "External id": 2484605,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 1404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770258.473, "dur": 8.102,
    "args": {
      "External id": 2484606,"Record function id": 0, "Ev Idx": 1405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770260.463, "dur": 5.425,
    "args": {
      "External id": 2484607,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 1406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770262.236, "dur": 2.806,
    "args": {
      "External id": 2484608,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 1407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770263.280, "dur": 1.653,
    "args": {
      "External id": 2484609,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 1408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770270.197, "dur": 8.786,
    "args": {
      "External id": 2484610,"Record function id": 0, "Sequence number": 24740500, "Fwd thread id": 1, "Ev Idx": 1409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770270.939, "dur": 5.636,
    "args": {
      "External id": 2484611,"Sequence number": 24740500, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 1410
    }
  },
  {
    "ph": "f", "id": 148, "pid": 1336759, "tid": 1381173, "ts": 1514656770270.939,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770272.515, "dur": 3.827,
    "args": {
      "External id": 2484612,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770275.291, "dur": 0.955,
    "args": {
      "External id": 2484613,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770282.339, "dur": 123.924,
    "args": {
      "External id": 2484614,"Record function id": 0, "Sequence number": 24740499, "Fwd thread id": 1, "Ev Idx": 1413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770283.315, "dur": 117.301,
    "args": {
      "External id": 2484615,"Sequence number": 24740499, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 1414
    }
  },
  {
    "ph": "f", "id": 149, "pid": 1336759, "tid": 1381173, "ts": 1514656770283.315,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770286.419, "dur": 4.171,
    "args": {
      "External id": 2484616,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 1415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770287.821, "dur": 2.123,
    "args": {
      "External id": 2484617,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[65536, 2048], [], []], "Ev Idx": 1416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770289.020, "dur": 0.721,
    "args": {
      "External id": 2484618,"Record function id": 0, "Concrete Inputs": ["", "[2048, 65536]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 1417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770291.754, "dur": 52.827,
    "args": {
      "External id": 2484619,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048]], "Ev Idx": 1418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770347.304, "dur": 7.114,
    "args": {
      "External id": 2484620,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770348.086, "dur": 5.702,
    "args": {
      "External id": 2484621,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 1420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770350.817, "dur": 2.829,
    "args": {
      "External id": 2484622,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 1421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770355.865, "dur": 2.673,
    "args": {
      "External id": 2484623,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 1422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770356.658, "dur": 1.387,
    "args": {
      "External id": 2484624,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 1423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770357.553, "dur": 0.371,
    "args": {
      "External id": 2484625,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 1424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770361.025, "dur": 38.700,
    "args": {
      "External id": 2484626,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 1425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770411.168, "dur": 10.728,
    "args": {
      "External id": 2484627,"Record function id": 0, "Sequence number": 24740498, "Fwd thread id": 1, "Ev Idx": 1426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770412.266, "dur": 7.642,
    "args": {
      "External id": 2484628,"Sequence number": 24740498, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 1427
    }
  },
  {
    "ph": "f", "id": 150, "pid": 1336759, "tid": 1381173, "ts": 1514656770412.266,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770413.777, "dur": 5.973,
    "args": {
      "External id": 2484629,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770416.496, "dur": 3.149,
    "args": {
      "External id": 2484630,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770425.199, "dur": 8.179,
    "args": {
      "External id": 2484631,"Record function id": 0, "Sequence number": 24740497, "Fwd thread id": 1, "Ev Idx": 1430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770426.074, "dur": 5.822,
    "args": {
      "External id": 2484632,"Sequence number": 24740497, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 1431
    }
  },
  {
    "ph": "f", "id": 151, "pid": 1336759, "tid": 1381173, "ts": 1514656770426.074,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770426.827, "dur": 4.865,
    "args": {
      "External id": 2484633,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 1432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770427.533, "dur": 3.665,
    "args": {
      "External id": 2484634,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 1433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770430.601, "dur": 0.478,
    "args": {
      "External id": 2484635,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 1434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770437.127, "dur": 4.566,
    "args": {
      "External id": 2484636,"Record function id": 0, "Ev Idx": 1435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770438.381, "dur": 2.769,
    "args": {
      "External id": 2484637,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770439.189, "dur": 1.719,
    "args": {
      "External id": 2484638,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770439.576, "dur": 1.247,
    "args": {
      "External id": 2484639,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770444.838, "dur": 5.421,
    "args": {
      "External id": 2484640,"Record function id": 0, "Sequence number": 24740496, "Fwd thread id": 1, "Ev Idx": 1439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770445.608, "dur": 2.702,
    "args": {
      "External id": 2484641,"Sequence number": 24740496, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 1440
    }
  },
  {
    "ph": "f", "id": 152, "pid": 1336759, "tid": 1381173, "ts": 1514656770445.608,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770446.434, "dur": 1.722,
    "args": {
      "External id": 2484642,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770446.933, "dur": 1.094,
    "args": {
      "External id": 2484643,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: FlashAttnFuncBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770454.299, "dur": 302.517,
    "args": {
      "External id": 2484644,"Record function id": 0, "Sequence number": 24740495, "Fwd thread id": 1, "Ev Idx": 1443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FlashAttnFuncBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770456.787, "dur": 279.524,
    "args": {
      "External id": 2484645,"Sequence number": 24740495, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 1444
    }
  },
  {
    "ph": "f", "id": 153, "pid": 1336759, "tid": 1381173, "ts": 1514656770456.787,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770469.313, "dur": 8.438,
    "args": {
      "External id": 2484646,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 1445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770473.629, "dur": 3.562,
    "args": {
      "External id": 2484647,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770479.466, "dur": 4.783,
    "args": {
      "External id": 2484648,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 1447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770482.045, "dur": 1.968,
    "args": {
      "External id": 2484649,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770485.392, "dur": 2.466,
    "args": {
      "External id": 2484650,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 1449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770485.986, "dur": 1.692,
    "args": {
      "External id": 2484651,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770511.194, "dur": 198.735,
    "args": {
      "External id": 2484652,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 1451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770608.128, "dur": 6.251,
    "args": {
      "External id": 2484653,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770617.991, "dur": 8.042,
    "args": {
      "External id": 2484654,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770723.460, "dur": 3.815,
    "args": {
      "External id": 2484655,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 1454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770730.368, "dur": 0.744,
    "args": {
      "External id": 2484656,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 1455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770733.054, "dur": 0.789,
    "args": {
      "External id": 2484657,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 1456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RotaryEmbeddingFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770766.175, "dur": 269.855,
    "args": {
      "External id": 2484658,"Record function id": 0, "Sequence number": 24740494, "Fwd thread id": 1, "Ev Idx": 1457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770767.782, "dur": 259.323,
    "args": {
      "External id": 2484659,"Sequence number": 24740494, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 1458
    }
  },
  {
    "ph": "f", "id": 154, "pid": 1336759, "tid": 1381173, "ts": 1514656770767.782,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770786.044, "dur": 53.801,
    "args": {
      "External id": 2484660,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 1459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770790.436, "dur": 3.335,
    "args": {
      "External id": 2484661,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770795.169, "dur": 43.974,
    "args": {
      "External id": 2484662,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], []], "Ev Idx": 1461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770848.855, "dur": 4.353,
    "args": {
      "External id": 2484663,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 1462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656770850.125, "dur": 2.773,
    "args": {
      "External id": 2484664,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RotaryEmbeddingFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771045.840, "dur": 170.193,
    "args": {
      "External id": 2484665,"Record function id": 0, "Sequence number": 24740493, "Fwd thread id": 1, "Ev Idx": 1464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771049.927, "dur": 159.740,
    "args": {
      "External id": 2484666,"Sequence number": 24740493, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 1465
    }
  },
  {
    "ph": "f", "id": 155, "pid": 1336759, "tid": 1381173, "ts": 1514656771049.927,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771063.063, "dur": 39.862,
    "args": {
      "External id": 2484667,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 1466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771065.743, "dur": 3.383,
    "args": {
      "External id": 2484668,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771070.205, "dur": 32.021,
    "args": {
      "External id": 2484669,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], []], "Ev Idx": 1468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771110.858, "dur": 3.898,
    "args": {
      "External id": 2484670,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 1469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771111.939, "dur": 2.475,
    "args": {
      "External id": 2484671,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771222.007, "dur": 15.299,
    "args": {
      "External id": 2484672,"Record function id": 0, "Sequence number": 24740492, "Fwd thread id": 1, "Ev Idx": 1471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771223.251, "dur": 11.510,
    "args": {
      "External id": 2484673,"Sequence number": 24740492, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 1472
    }
  },
  {
    "ph": "f", "id": 156, "pid": 1336759, "tid": 1381173, "ts": 1514656771223.251,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771227.054, "dur": 7.452,
    "args": {
      "External id": 2484674,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 1473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771228.260, "dur": 6.010,
    "args": {
      "External id": 2484675,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 1474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771240.975, "dur": 5.532,
    "args": {
      "External id": 2484676,"Record function id": 0, "Sequence number": 24740491, "Fwd thread id": 1, "Ev Idx": 1475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771241.897, "dur": 2.995,
    "args": {
      "External id": 2484677,"Sequence number": 24740491, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 1476
    }
  },
  {
    "ph": "f", "id": 157, "pid": 1336759, "tid": 1381173, "ts": 1514656771241.897,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771242.831, "dur": 1.875,
    "args": {
      "External id": 2484678,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 1477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771243.604, "dur": 0.966,
    "args": {
      "External id": 2484679,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 1478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771249.405, "dur": 8.463,
    "args": {
      "External id": 2484680,"Record function id": 0, "Sequence number": 24740490, "Fwd thread id": 1, "Ev Idx": 1479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771250.067, "dur": 6.426,
    "args": {
      "External id": 2484681,"Sequence number": 24740490, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 1480
    }
  },
  {
    "ph": "f", "id": 158, "pid": 1336759, "tid": 1381173, "ts": 1514656771250.067,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771254.820, "dur": 1.525,
    "args": {
      "External id": 2484682,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 1481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771255.402, "dur": 0.845,
    "args": {
      "External id": 2484683,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 1482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771261.358, "dur": 5.037,
    "args": {
      "External id": 2484684,"Record function id": 0, "Sequence number": 24740489, "Fwd thread id": 1, "Ev Idx": 1483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771262.168, "dur": 2.470,
    "args": {
      "External id": 2484685,"Sequence number": 24740489, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 1484
    }
  },
  {
    "ph": "f", "id": 159, "pid": 1336759, "tid": 1381173, "ts": 1514656771262.168,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771263.019, "dur": 1.479,
    "args": {
      "External id": 2484686,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771263.613, "dur": 0.792,
    "args": {
      "External id": 2484687,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771269.768, "dur": 148.907,
    "args": {
      "External id": 2484688,"Record function id": 0, "Sequence number": 24740488, "Fwd thread id": 1, "Ev Idx": 1487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771270.662, "dur": 141.250,
    "args": {
      "External id": 2484689,"Sequence number": 24740488, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 1488
    }
  },
  {
    "ph": "f", "id": 160, "pid": 1336759, "tid": 1381173, "ts": 1514656771270.662,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771274.468, "dur": 10.922,
    "args": {
      "External id": 2484690,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 1489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771278.608, "dur": 6.171,
    "args": {
      "External id": 2484691,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[65536, 2048], [], []], "Ev Idx": 1490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771282.605, "dur": 1.947,
    "args": {
      "External id": 2484692,"Record function id": 0, "Concrete Inputs": ["", "[2048, 65536]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 1491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771286.830, "dur": 65.122,
    "args": {
      "External id": 2484693,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048]], "Ev Idx": 1492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771353.274, "dur": 4.401,
    "args": {
      "External id": 2484694,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771354.055, "dur": 2.905,
    "args": {
      "External id": 2484695,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 1494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771355.219, "dur": 1.593,
    "args": {
      "External id": 2484696,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 1495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771359.444, "dur": 4.671,
    "args": {
      "External id": 2484697,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 1496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771362.564, "dur": 1.185,
    "args": {
      "External id": 2484698,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 1497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771363.375, "dur": 0.292,
    "args": {
      "External id": 2484699,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 1498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771364.832, "dur": 42.940,
    "args": {
      "External id": 2484700,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 1499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771423.414, "dur": 7.503,
    "args": {
      "External id": 2484701,"Record function id": 0, "Sequence number": 24740487, "Fwd thread id": 1, "Ev Idx": 1500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771424.346, "dur": 4.960,
    "args": {
      "External id": 2484702,"Sequence number": 24740487, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 1501
    }
  },
  {
    "ph": "f", "id": 161, "pid": 1336759, "tid": 1381173, "ts": 1514656771424.346,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771427.341, "dur": 1.824,
    "args": {
      "External id": 2484703,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771428.014, "dur": 1.056,
    "args": {
      "External id": 2484704,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771434.069, "dur": 7.617,
    "args": {
      "External id": 2484705,"Record function id": 0, "Sequence number": 24740486, "Fwd thread id": 1, "Ev Idx": 1504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771434.989, "dur": 4.975,
    "args": {
      "External id": 2484706,"Sequence number": 24740486, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 1505
    }
  },
  {
    "ph": "f", "id": 162, "pid": 1336759, "tid": 1381173, "ts": 1514656771434.989,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771437.338, "dur": 2.423,
    "args": {
      "External id": 2484707,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 1506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771437.934, "dur": 1.358,
    "args": {
      "External id": 2484708,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 1507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771438.703, "dur": 0.448,
    "args": {
      "External id": 2484709,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 1508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771447.512, "dur": 8.800,
    "args": {
      "External id": 2484710,"Record function id": 0, "Ev Idx": 1509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771449.204, "dur": 6.260,
    "args": {
      "External id": 2484711,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771451.573, "dur": 3.537,
    "args": {
      "External id": 2484712,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771452.633, "dur": 2.362,
    "args": {
      "External id": 2484713,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771459.740, "dur": 9.391,
    "args": {
      "External id": 2484714,"Record function id": 0, "Sequence number": 24740485, "Fwd thread id": 1, "Ev Idx": 1513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771462.694, "dur": 4.521,
    "args": {
      "External id": 2484715,"Sequence number": 24740485, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 1514
    }
  },
  {
    "ph": "f", "id": 163, "pid": 1336759, "tid": 1381173, "ts": 1514656771462.694,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771463.711, "dur": 3.354,
    "args": {
      "External id": 2484716,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771466.263, "dur": 0.686,
    "args": {
      "External id": 2484717,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771472.256, "dur": 107.741,
    "args": {
      "External id": 2484718,"Record function id": 0, "Sequence number": 24740484, "Fwd thread id": 1, "Ev Idx": 1517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771473.257, "dur": 100.511,
    "args": {
      "External id": 2484719,"Sequence number": 24740484, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 1518
    }
  },
  {
    "ph": "f", "id": 164, "pid": 1336759, "tid": 1381173, "ts": 1514656771473.257,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771474.953, "dur": 2.456,
    "args": {
      "External id": 2484720,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 1519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771475.643, "dur": 1.354,
    "args": {
      "External id": 2484721,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[65536, 2048], [], []], "Ev Idx": 1520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771476.464, "dur": 0.421,
    "args": {
      "External id": 2484722,"Record function id": 0, "Concrete Inputs": ["", "[2048, 65536]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 1521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771478.124, "dur": 24.915,
    "args": {
      "External id": 2484723,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048]], "Ev Idx": 1522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771506.284, "dur": 3.623,
    "args": {
      "External id": 2484724,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771506.967, "dur": 2.427,
    "args": {
      "External id": 2484725,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 1524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771507.854, "dur": 1.417,
    "args": {
      "External id": 2484726,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 1525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771510.939, "dur": 6.888,
    "args": {
      "External id": 2484727,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 1526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771512.201, "dur": 5.270,
    "args": {
      "External id": 2484728,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 1527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771515.193, "dur": 2.189,
    "args": {
      "External id": 2484729,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 1528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771520.457, "dur": 52.003,
    "args": {
      "External id": 2484730,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 1529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771588.969, "dur": 36.910,
    "args": {
      "External id": 2484731,"Record function id": 0, "Sequence number": 24740483, "Fwd thread id": 1, "Ev Idx": 1530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771590.230, "dur": 8.887,
    "args": {
      "External id": 2484732,"Sequence number": 24740483, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 1531
    }
  },
  {
    "ph": "f", "id": 165, "pid": 1336759, "tid": 1381173, "ts": 1514656771590.230,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771591.553, "dur": 7.416,
    "args": {
      "External id": 2484733,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771597.049, "dur": 1.795,
    "args": {
      "External id": 2484734,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771602.486, "dur": 21.047,
    "args": {
      "External id": 2484735,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 1534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771629.542, "dur": 31.734,
    "args": {
      "External id": 2484736,"Record function id": 0, "Sequence number": 24740482, "Fwd thread id": 1, "Ev Idx": 1535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771651.411, "dur": 8.155,
    "args": {
      "External id": 2484737,"Sequence number": 24740482, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 1536
    }
  },
  {
    "ph": "f", "id": 166, "pid": 1336759, "tid": 1381173, "ts": 1514656771651.411,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771654.254, "dur": 5.089,
    "args": {
      "External id": 2484738,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 1537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771654.911, "dur": 3.893,
    "args": {
      "External id": 2484739,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 1538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771658.185, "dur": 0.477,
    "args": {
      "External id": 2484740,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 1539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771665.426, "dur": 5.643,
    "args": {
      "External id": 2484741,"Record function id": 0, "Ev Idx": 1540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771666.540, "dur": 3.995,
    "args": {
      "External id": 2484742,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771667.771, "dur": 2.375,
    "args": {
      "External id": 2484743,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771668.387, "dur": 1.607,
    "args": {
      "External id": 2484744,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771673.921, "dur": 7.023,
    "args": {
      "External id": 2484745,"Record function id": 0, "Sequence number": 24740481, "Fwd thread id": 1, "Ev Idx": 1544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "UnsafeViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771674.726, "dur": 4.329,
    "args": {
      "External id": 2484746,"Sequence number": 24740481, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 1545
    }
  },
  {
    "ph": "f", "id": 167, "pid": 1336759, "tid": 1381173, "ts": 1514656771674.726,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771675.490, "dur": 3.419,
    "args": {
      "External id": 2484747,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771677.926, "dur": 0.864,
    "args": {
      "External id": 2484748,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771683.912, "dur": 89.176,
    "args": {
      "External id": 2484749,"Record function id": 0, "Sequence number": 24740480, "Fwd thread id": 1, "Ev Idx": 1548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "MmBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771684.580, "dur": 81.931,
    "args": {
      "External id": 2484750,"Sequence number": 24740480, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 1549
    }
  },
  {
    "ph": "f", "id": 168, "pid": 1336759, "tid": 1381173, "ts": 1514656771684.580,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771686.597, "dur": 3.989,
    "args": {
      "External id": 2484751,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 1550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771687.164, "dur": 3.008,
    "args": {
      "External id": 2484752,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[65536, 2048], [], []], "Ev Idx": 1551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771689.590, "dur": 0.470,
    "args": {
      "External id": 2484753,"Record function id": 0, "Concrete Inputs": ["", "[2048, 65536]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 1552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771691.346, "dur": 29.333,
    "args": {
      "External id": 2484754,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048]], "Ev Idx": 1553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771724.302, "dur": 2.555,
    "args": {
      "External id": 2484755,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771724.851, "dur": 1.508,
    "args": {
      "External id": 2484756,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 1555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771725.682, "dur": 0.562,
    "args": {
      "External id": 2484757,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 1556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771727.850, "dur": 2.191,
    "args": {
      "External id": 2484758,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 1557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771728.634, "dur": 1.054,
    "args": {
      "External id": 2484759,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 1558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771729.288, "dur": 0.280,
    "args": {
      "External id": 2484760,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 1559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771731.958, "dur": 33.900,
    "args": {
      "External id": 2484761,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 1560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771777.362, "dur": 25.015,
    "args": {
      "External id": 2484762,"Record function id": 0, "Sequence number": 24740479, "Fwd thread id": 1, "Ev Idx": 1561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "ViewBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771778.195, "dur": 4.867,
    "args": {
      "External id": 2484763,"Sequence number": 24740479, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 1562
    }
  },
  {
    "ph": "f", "id": 169, "pid": 1336759, "tid": 1381173, "ts": 1514656771778.195,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771781.140, "dur": 1.776,
    "args": {
      "External id": 2484764,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771781.779, "dur": 1.043,
    "args": {
      "External id": 2484765,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771785.454, "dur": 14.766,
    "args": {
      "External id": 2484766,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 1565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771805.717, "dur": 7.682,
    "args": {
      "External id": 2484767,"Record function id": 0, "Sequence number": 24740478, "Fwd thread id": 1, "Ev Idx": 1566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TBackward0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771806.594, "dur": 5.399,
    "args": {
      "External id": 2484768,"Sequence number": 24740478, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 1567
    }
  },
  {
    "ph": "f", "id": 170, "pid": 1336759, "tid": 1381173, "ts": 1514656771806.594,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771807.246, "dur": 4.540,
    "args": {
      "External id": 2484769,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1, 2048]], "Input Dims": [[2048, 2048]], "Ev Idx": 1568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771807.900, "dur": 3.342,
    "args": {
      "External id": 2484770,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[1, 2048], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 1569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771810.699, "dur": 0.426,
    "args": {
      "External id": 2484771,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[2048, 1]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 2048], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 1570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771817.027, "dur": 4.928,
    "args": {
      "External id": 2484772,"Record function id": 0, "Ev Idx": 1571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771818.259, "dur": 3.228,
    "args": {
      "External id": 2484773,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771819.129, "dur": 2.051,
    "args": {
      "External id": 2484774,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771819.554, "dur": 1.516,
    "args": {
      "External id": 2484775,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: LayerNormFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771825.573, "dur": 388.180,
    "args": {
      "External id": 2484776,"Record function id": 0, "Sequence number": 24740477, "Fwd thread id": 1, "Ev Idx": 1575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771827.975, "dur": 353.026,
    "args": {
      "External id": 2484777,"Sequence number": 24740477, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 1576
    }
  },
  {
    "ph": "f", "id": 171, "pid": 1336759, "tid": 1381173, "ts": 1514656771827.975,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771858.685, "dur": 1.404,
    "args": {
      "External id": 2484778,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771859.122, "dur": 0.820,
    "args": {
      "External id": 2484779,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 1578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771875.309, "dur": 6.485,
    "args": {
      "External id": 2484780,"Record function id": 0, "Concrete Inputs": ["[65536, 2048]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656771890.656, "dur": 1.637,
    "args": {
      "External id": 2484781,"Record function id": 0, "Concrete Inputs": ["[132, 2048]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772075.017, "dur": 3.352,
    "args": {
      "External id": 2484782,"Record function id": 0, "Concrete Inputs": ["", "[1, -1, 2048]"], "Input type": ["float", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[132, 2048], []], "Ev Idx": 1581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772082.604, "dur": 38.460,
    "args": {
      "External id": 2484783,"Record function id": 0, "Concrete Inputs": ["", "[1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[270336, 2048, 1], [], [], []], "Input Dims": [[1, 132, 2048], [], [], []], "Ev Idx": 1582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772096.336, "dur": 1.030,
    "args": {
      "External id": 2484784,"Record function id": 0, "Concrete Inputs": ["", "[1, 1, 2048]", "[2048, 0, 1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[1, 2048], [], [], []], "Ev Idx": 1583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772127.082, "dur": 30.264,
    "args": {
      "External id": 2484785,"Record function id": 0, "Concrete Inputs": ["", "", "15", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[1, 2048], [], [], [], [], []], "Ev Idx": 1584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772128.778, "dur": 28.373,
    "args": {
      "External id": 2484786,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], [], []], "Input Dims": [[1, 2048], [], [], [], [], [], []], "Ev Idx": 1585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772132.626, "dur": 4.891,
    "args": {
      "External id": 2484787,"Record function id": 0, "Concrete Inputs": ["[1, 2048]", "[2048, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772141.070, "dur": 15.469,
    "args": {
      "External id": 2484788,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "float", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[1, 2048], [1, 2048], []], "Ev Idx": 1587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772161.597, "dur": 4.601,
    "args": {
      "External id": 2484789,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1]], "Input Dims": [[1, 2048], [2048]], "Ev Idx": 1588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772164.817, "dur": 1.252,
    "args": {
      "External id": 2484790,"Record function id": 0, "Concrete Inputs": ["", "[2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[1, 2048], []], "Ev Idx": 1589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772172.202, "dur": 1.870,
    "args": {
      "External id": 2484791,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772173.093, "dur": 0.868,
    "args": {
      "External id": 2484792,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 1591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772194.507, "dur": 14.938,
    "args": {
      "External id": 2484793,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 1592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772225.886, "dur": 11.736,
    "args": {
      "External id": 2484794,"Record function id": 0, "Ev Idx": 1593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772228.049, "dur": 8.883,
    "args": {
      "External id": 2484795,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 1594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772230.092, "dur": 5.936,
    "args": {
      "External id": 2484796,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 1595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772233.790, "dur": 2.122,
    "args": {
      "External id": 2484797,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 1596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772242.373, "dur": 2559.837,
    "args": {
      "External id": 2484798,"Record function id": 0, "Ev Idx": 1597
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.27)", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772274.797, "dur": 904.517,
    "args": {
      "External id": 2484799,"Record function id": 0, "Ev Idx": 1598
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.26", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772301.691, "dur": 869.814,
    "args": {
      "External id": 2484800,"Record function id": 0, "Ev Idx": 1599
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.26)", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772313.451, "dur": 843.621,
    "args": {
      "External id": 2484801,"Record function id": 0, "Ev Idx": 1600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772384.245, "dur": 9.069,
    "args": {
      "External id": 2484802,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772408.005, "dur": 38.836,
    "args": {
      "External id": 2484803,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 1602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772413.983, "dur": 3.228,
    "args": {
      "External id": 2484804,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772418.330, "dur": 0.546,
    "args": {
      "External id": 2484805,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772423.208, "dur": 0.378,
    "args": {
      "External id": 2484806,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772424.470, "dur": 0.672,
    "args": {
      "External id": 2484807,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772426.049, "dur": 0.492,
    "args": {
      "External id": 2484808,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772430.719, "dur": 0.546,
    "args": {
      "External id": 2484809,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772431.736, "dur": 0.376,
    "args": {
      "External id": 2484810,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772432.783, "dur": 1.450,
    "args": {
      "External id": 2484811,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772436.807, "dur": 1.995,
    "args": {
      "External id": 2484812,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772460.428, "dur": 28.972,
    "args": {
      "External id": 2484813,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 1612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772523.358, "dur": 132.706,
    "args": {
      "External id": 2484814,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 1613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772554.963, "dur": 7.572,
    "args": {
      "External id": 2484815,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772567.531, "dur": 9.031,
    "args": {
      "External id": 2484816,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 1615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772571.140, "dur": 5.015,
    "args": {
      "External id": 2484817,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 1616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772574.022, "dur": 0.734,
    "args": {
      "External id": 2484818,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 1617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772583.525, "dur": 29.875,
    "args": {
      "External id": 2484819,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 1618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772586.900, "dur": 0.458,
    "args": {
      "External id": 2484820,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772588.380, "dur": 1.940,
    "args": {
      "External id": 2484821,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772591.177, "dur": 0.419,
    "args": {
      "External id": 2484822,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772594.095, "dur": 0.407,
    "args": {
      "External id": 2484823,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772597.016, "dur": 0.571,
    "args": {
      "External id": 2484824,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772598.209, "dur": 2.332,
    "args": {
      "External id": 2484825,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772602.841, "dur": 0.591,
    "args": {
      "External id": 2484826,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772605.861, "dur": 0.400,
    "args": {
      "External id": 2484827,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772608.985, "dur": 0.295,
    "args": {
      "External id": 2484828,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772624.416, "dur": 23.307,
    "args": {
      "External id": 2484829,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 1628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772710.165, "dur": 342.801,
    "args": {
      "External id": 2484830,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 1629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772747.387, "dur": 300.450,
    "args": {
      "External id": 2484831,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 1630, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514656772757.241, "dur": 284.239,
    "args": {
      "External id": 2484832,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 1631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514656773077.951, "dur": 2.389,
    "args": {
      "External id": 2484833,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 1632, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656773187.445, "dur": 1595.216,
    "args": {
      "External id": 2484834,"Sequence number": 24740476, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 1633
    }
  },
  {
    "ph": "f", "id": 172, "pid": 1336759, "tid": 1381173, "ts": 1514656773187.445,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656773302.656, "dur": 107.681,
    "args": {
      "External id": 2484835,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [5632, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 1634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656773454.363, "dur": 35.264,
    "args": {
      "External id": 2484836,"kernel_hash": "cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/m5/cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [65536, 5632], [16, 4096, 5632], [16, 4096, 5632], [16, 4096, 5632], []], "Ev Idx": 1635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656773507.173, "dur": 70.014,
    "args": {
      "External id": 2484837,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 2048], [0, 5632, 1], [11534336, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632], [1, 2048, 5632]], "Ev Idx": 1636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656773590.842, "dur": 34.083,
    "args": {
      "External id": 2484838,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 1637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656773631.457, "dur": 48.391,
    "args": {
      "External id": 2484839,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 1638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656773688.845, "dur": 27.557,
    "args": {
      "External id": 2484840,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 1639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656773723.592, "dur": 42.077,
    "args": {
      "External id": 2484841,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 1640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514656773792.818, "dur": 23.476,
    "args": {
      "External id": 2484842,"kernel_hash": "c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/47/c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 1641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656773835.367, "dur": 28.313,
    "args": {
      "External id": 2484843,"kernel_hash": "cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/xr/cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 1642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514656773887.887, "dur": 18.981,
    "args": {
      "External id": 2484844,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 1643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514656773922.922, "dur": 17.520,
    "args": {
      "External id": 2484845,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 1644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656773949.015, "dur": 26.713,
    "args": {
      "External id": 2484846,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 1645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656773979.118, "dur": 75.300,
    "args": {
      "External id": 2484847,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 1646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774086.455, "dur": 169.578,
    "args": {
      "External id": 2484848,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 1647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774166.471, "dur": 7.626,
    "args": {
      "External id": 2484849,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774175.980, "dur": 2.383,
    "args": {
      "External id": 2484850,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774290.318, "dur": 24.825,
    "args": {
      "External id": 2484851,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 1650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774332.362, "dur": 14.083,
    "args": {
      "External id": 2484852,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 1651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774354.428, "dur": 32.831,
    "args": {
      "External id": 2484853,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 1652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774392.560, "dur": 34.718,
    "args": {
      "External id": 2484854,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 1653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774434.461, "dur": 19.966,
    "args": {
      "External id": 2484855,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 1654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774460.679, "dur": 29.303,
    "args": {
      "External id": 2484856,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 1655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774495.737, "dur": 19.166,
    "args": {
      "External id": 2484857,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 1656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774521.136, "dur": 47.975,
    "args": {
      "External id": 2484858,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 1657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774594.668, "dur": 21.748,
    "args": {
      "External id": 2484859,"kernel_hash": "cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/g4/cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [65536, 2048], []], "Ev Idx": 1658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774635.169, "dur": 22.062,
    "args": {
      "External id": 2484860,"kernel_hash": "c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/4n/c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 1659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774676.652, "dur": 15.520,
    "args": {
      "External id": 2484861,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 1660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774705.960, "dur": 13.818,
    "args": {
      "External id": 2484862,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 1661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774737.680, "dur": 17.336,
    "args": {
      "External id": 2484863,"kernel_hash": "cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u7/cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [8388608, 2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [16, 4096, 2048], []], "Ev Idx": 1662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774824.548, "dur": 14.020,
    "args": {
      "External id": 2484864,"Record function id": 0, "Ev Idx": 1663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774827.559, "dur": 10.072,
    "args": {
      "External id": 2484865,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 1664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774831.615, "dur": 5.177,
    "args": {
      "External id": 2484866,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 1665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774833.084, "dur": 3.624,
    "args": {
      "External id": 2484867,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 1666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774842.615, "dur": 6.079,
    "args": {
      "External id": 2484868,"Record function id": 0, "Ev Idx": 1667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774845.828, "dur": 2.424,
    "args": {
      "External id": 2484869,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774846.539, "dur": 1.299,
    "args": {
      "External id": 2484870,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774847.164, "dur": 0.605,
    "args": {
      "External id": 2484871,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774852.015, "dur": 4.072,
    "args": {
      "External id": 2484872,"Record function id": 0, "Ev Idx": 1671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774853.084, "dur": 2.579,
    "args": {
      "External id": 2484873,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 1672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774853.630, "dur": 1.606,
    "args": {
      "External id": 2484874,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 1673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774854.450, "dur": 0.696,
    "args": {
      "External id": 2484875,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 1674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774859.232, "dur": 6.269,
    "args": {
      "External id": 2484876,"Record function id": 0, "Ev Idx": 1675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774860.138, "dur": 4.950,
    "args": {
      "External id": 2484877,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 1676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774860.945, "dur": 3.598,
    "args": {
      "External id": 2484878,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 1677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774861.409, "dur": 3.054,
    "args": {
      "External id": 2484879,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 1678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774868.582, "dur": 3.690,
    "args": {
      "External id": 2484880,"Record function id": 0, "Ev Idx": 1679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774869.517, "dur": 2.344,
    "args": {
      "External id": 2484881,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774870.162, "dur": 1.077,
    "args": {
      "External id": 2484882,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774870.404, "dur": 0.768,
    "args": {
      "External id": 2484883,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774875.246, "dur": 3.934,
    "args": {
      "External id": 2484884,"Record function id": 0, "Ev Idx": 1683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774876.626, "dur": 2.144,
    "args": {
      "External id": 2484885,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774877.087, "dur": 1.232,
    "args": {
      "External id": 2484886,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774877.703, "dur": 0.549,
    "args": {
      "External id": 2484887,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774882.270, "dur": 3.225,
    "args": {
      "External id": 2484888,"Record function id": 0, "Ev Idx": 1687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774883.181, "dur": 1.915,
    "args": {
      "External id": 2484889,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774883.709, "dur": 0.926,
    "args": {
      "External id": 2484890,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774884.037, "dur": 0.529,
    "args": {
      "External id": 2484891,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774888.444, "dur": 5.297,
    "args": {
      "External id": 2484892,"Record function id": 0, "Ev Idx": 1691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774889.588, "dur": 3.764,
    "args": {
      "External id": 2484893,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774890.167, "dur": 2.692,
    "args": {
      "External id": 2484894,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774892.118, "dur": 0.673,
    "args": {
      "External id": 2484895,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774897.952, "dur": 3.504,
    "args": {
      "External id": 2484896,"Record function id": 0, "Ev Idx": 1695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774898.904, "dur": 2.124,
    "args": {
      "External id": 2484897,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774899.485, "dur": 1.100,
    "args": {
      "External id": 2484898,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774899.958, "dur": 0.561,
    "args": {
      "External id": 2484899,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774905.319, "dur": 178059.117,
    "args": {
      "External id": 2484900,"Record function id": 0, "Sequence number": 24740475, "Fwd thread id": 1, "Ev Idx": 1699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774906.634, "dur": 178049.412,
    "args": {
      "External id": 2484901,"Sequence number": 24740475, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 1700
    }
  },
  {
    "ph": "f", "id": 173, "pid": 1336759, "tid": 1381173, "ts": 1514656774906.634,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.27)", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774933.927, "dur": 38.660,
    "args": {
      "External id": 2484902,"Record function id": 0, "Ev Idx": 1701
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.27)", "pid": 1336759, "tid": 1381173,
    "ts": 1514656774980.224, "dur": 108.893,
    "args": {
      "External id": 2484903,"Record function id": 0, "Ev Idx": 1702
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.27)", "pid": 1336759, "tid": 1381173,
    "ts": 1514656775098.153, "dur": 177849.045,
    "args": {
      "External id": 2484904,"Record function id": 0, "Ev Idx": 1703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656775153.957, "dur": 8.611,
    "args": {
      "External id": 2484905,"Record function id": 0, "Concrete Inputs": ["[51384320]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656775173.719, "dur": 4.902,
    "args": {
      "External id": 2484906,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 1705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514656775195.168, "dur": 176904.033,
    "args": {
      "External id": 2484907,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 1706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514656775209.546, "dur": 176878.137,
    "args": {
      "External id": 2484908,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 1707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656775293.176, "dur": 9.322,
    "args": {
      "External id": 2484909,"Record function id": 0, "Concrete Inputs": ["[12602]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514656775311.738, "dur": 176731.190,
    "args": {
      "External id": 2484910,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], [], []], "Ev Idx": 1709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514656775314.291, "dur": 176727.531,
    "args": {
      "External id": 2484911,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], []], "Ev Idx": 1710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656775317.590, "dur": 8.017,
    "args": {
      "External id": 2484912,"Record function id": 0, "Concrete Inputs": ["[12602]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656775327.206, "dur": 176709.886,
    "args": {
      "External id": 2484913,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[12602], [12602], []], "Ev Idx": 1712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656952202.559, "dur": 10.757,
    "args": {
      "External id": 2484914,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[51384320], [], [], [], [], []], "Ev Idx": 1713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656952205.927, "dur": 6.951,
    "args": {
      "External id": 2484915,"Record function id": 0, "Concrete Inputs": ["[6423040]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656952251.238, "dur": 388.748,
    "args": {
      "External id": 2484916,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[6423040], [51384320], [], [], [], []], "Ev Idx": 1715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514656952285.809, "dur": 349.039,
    "args": {
      "External id": 2484917,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 6423040, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[51384320], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 1716, "In msg nelems": 51384320
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514656952301.691, "dur": 327.209,
    "args": {
      "External id": 2484918,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[51384320]], "Ev Idx": 1717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514656952664.737, "dur": 2.300,
    "args": {
      "External id": 2484919,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 1718, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656952726.752, "dur": 6.753,
    "args": {
      "External id": 2484920,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656952786.527, "dur": 1.534,
    "args": {
      "External id": 2484921,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "256"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656952804.845, "dur": 1.772,
    "args": {
      "External id": 2484922,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656952822.161, "dur": 0.850,
    "args": {
      "External id": 2484923,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1048832"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656952833.343, "dur": 1.114,
    "args": {
      "External id": 2484924,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1573120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656952843.229, "dur": 0.763,
    "args": {
      "External id": 2484925,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656952855.624, "dur": 1.244,
    "args": {
      "External id": 2484926,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "2097664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656952869.831, "dur": 2.931,
    "args": {
      "External id": 2484927,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "3539456"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656952883.137, "dur": 0.832,
    "args": {
      "External id": 2484928,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "4981248"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656952978.623, "dur": 2707.327,
    "args": {
      "External id": 2484929,"Record function id": 0, "Ev Idx": 1728
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.26)", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953032.317, "dur": 1044.003,
    "args": {
      "External id": 2484930,"Record function id": 0, "Ev Idx": 1729
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.26)", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953048.286, "dur": 316.818,
    "args": {
      "External id": 2484931,"Record function id": 0, "Ev Idx": 1730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953127.556, "dur": 4.841,
    "args": {
      "External id": 2484932,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 1731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953135.405, "dur": 0.989,
    "args": {
      "External id": 2484933,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 1732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953137.961, "dur": 0.934,
    "args": {
      "External id": 2484934,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 1733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953140.350, "dur": 2.929,
    "args": {
      "External id": 2484935,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 1734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953146.455, "dur": 0.894,
    "args": {
      "External id": 2484936,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 1735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953148.747, "dur": 0.878,
    "args": {
      "External id": 2484937,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 1736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953151.195, "dur": 2.326,
    "args": {
      "External id": 2484938,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 1737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953155.127, "dur": 0.781,
    "args": {
      "External id": 2484939,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 1738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953160.021, "dur": 0.917,
    "args": {
      "External id": 2484940,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 1739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953162.187, "dur": 0.796,
    "args": {
      "External id": 2484941,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 1740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953180.912, "dur": 156.476,
    "args": {
      "External id": 2484942,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 1741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953198.575, "dur": 134.466,
    "args": {
      "External id": 2484943,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 1742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953223.382, "dur": 12.522,
    "args": {
      "External id": 2484944,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953238.867, "dur": 67.520,
    "args": {
      "External id": 2484945,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 1744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953243.519, "dur": 62.525,
    "args": {
      "External id": 2484946,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 1745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953246.863, "dur": 8.562,
    "args": {
      "External id": 2484947,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953257.102, "dur": 48.157,
    "args": {
      "External id": 2484948,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 1747
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.25", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953438.899, "dur": 629.400,
    "args": {
      "External id": 2484949,"Record function id": 0, "Ev Idx": 1748
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.25)", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953453.520, "dur": 601.379,
    "args": {
      "External id": 2484950,"Record function id": 0, "Ev Idx": 1749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953506.415, "dur": 4.536,
    "args": {
      "External id": 2484951,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953525.349, "dur": 53.600,
    "args": {
      "External id": 2484952,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 1751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953547.807, "dur": 3.726,
    "args": {
      "External id": 2484953,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953552.988, "dur": 0.481,
    "args": {
      "External id": 2484954,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953554.315, "dur": 0.330,
    "args": {
      "External id": 2484955,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953559.608, "dur": 0.272,
    "args": {
      "External id": 2484956,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953560.565, "dur": 0.443,
    "args": {
      "External id": 2484957,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953561.573, "dur": 0.415,
    "args": {
      "External id": 2484958,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953565.720, "dur": 2.429,
    "args": {
      "External id": 2484959,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953568.849, "dur": 0.614,
    "args": {
      "External id": 2484960,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953572.162, "dur": 0.980,
    "args": {
      "External id": 2484961,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953600.589, "dur": 33.477,
    "args": {
      "External id": 2484962,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 1761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953667.502, "dur": 106.088,
    "args": {
      "External id": 2484963,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 1762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953677.890, "dur": 5.858,
    "args": {
      "External id": 2484964,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953688.484, "dur": 11.088,
    "args": {
      "External id": 2484965,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 1764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953692.517, "dur": 6.616,
    "args": {
      "External id": 2484966,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 1765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953697.217, "dur": 0.560,
    "args": {
      "External id": 2484967,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 1766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953706.667, "dur": 29.426,
    "args": {
      "External id": 2484968,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 1767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953708.283, "dur": 0.444,
    "args": {
      "External id": 2484969,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953711.889, "dur": 1.962,
    "args": {
      "External id": 2484970,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953714.523, "dur": 0.432,
    "args": {
      "External id": 2484971,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953715.847, "dur": 2.561,
    "args": {
      "External id": 2484972,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953723.536, "dur": 0.252,
    "args": {
      "External id": 2484973,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953724.493, "dur": 0.332,
    "args": {
      "External id": 2484974,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953726.919, "dur": 0.431,
    "args": {
      "External id": 2484975,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953729.874, "dur": 0.532,
    "args": {
      "External id": 2484976,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953730.904, "dur": 0.694,
    "args": {
      "External id": 2484977,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953747.680, "dur": 18.043,
    "args": {
      "External id": 2484978,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 1777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953819.150, "dur": 123.576,
    "args": {
      "External id": 2484979,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 1778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953850.303, "dur": 88.912,
    "args": {
      "External id": 2484980,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 1779, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953862.131, "dur": 72.923,
    "args": {
      "External id": 2484981,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 1780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514656953960.959, "dur": 2.037,
    "args": {
      "External id": 2484982,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 1781, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656954083.586, "dur": 1582.007,
    "args": {
      "External id": 2484983,"Sequence number": 24740474, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 1782
    }
  },
  {
    "ph": "f", "id": 174, "pid": 1336759, "tid": 1381173, "ts": 1514656954083.586,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656954191.106, "dur": 103.315,
    "args": {
      "External id": 2484984,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [5632, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 1783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656954337.120, "dur": 35.061,
    "args": {
      "External id": 2484985,"kernel_hash": "cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/m5/cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [65536, 5632], [16, 4096, 5632], [16, 4096, 5632], [16, 4096, 5632], []], "Ev Idx": 1784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656954387.938, "dur": 48.334,
    "args": {
      "External id": 2484986,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 2048], [0, 5632, 1], [11534336, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632], [1, 2048, 5632]], "Ev Idx": 1785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656954445.379, "dur": 31.836,
    "args": {
      "External id": 2484987,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 1786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656954486.376, "dur": 58.477,
    "args": {
      "External id": 2484988,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 1787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656954555.234, "dur": 30.885,
    "args": {
      "External id": 2484989,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 1788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656954593.930, "dur": 41.177,
    "args": {
      "External id": 2484990,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 1789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514656954660.706, "dur": 23.015,
    "args": {
      "External id": 2484991,"kernel_hash": "c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/47/c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 1790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656954706.088, "dur": 27.882,
    "args": {
      "External id": 2484992,"kernel_hash": "cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/xr/cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 1791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514656954755.663, "dur": 17.976,
    "args": {
      "External id": 2484993,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 1792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514656954790.561, "dur": 15.062,
    "args": {
      "External id": 2484994,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 1793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656954813.579, "dur": 26.700,
    "args": {
      "External id": 2484995,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 1794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656954843.670, "dur": 31.413,
    "args": {
      "External id": 2484996,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 1795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656954903.964, "dur": 203.171,
    "args": {
      "External id": 2484997,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 1796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656954978.980, "dur": 40.887,
    "args": {
      "External id": 2484998,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955023.054, "dur": 2.344,
    "args": {
      "External id": 2484999,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955146.643, "dur": 24.833,
    "args": {
      "External id": 2485000,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 1799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955185.312, "dur": 14.353,
    "args": {
      "External id": 2485001,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 1800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955207.370, "dur": 38.542,
    "args": {
      "External id": 2485002,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 1801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955251.525, "dur": 33.171,
    "args": {
      "External id": 2485003,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 1802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955295.842, "dur": 20.198,
    "args": {
      "External id": 2485004,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 1803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955319.996, "dur": 29.223,
    "args": {
      "External id": 2485005,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 1804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955354.444, "dur": 18.375,
    "args": {
      "External id": 2485006,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 1805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955379.168, "dur": 45.258,
    "args": {
      "External id": 2485007,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 1806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955451.836, "dur": 20.361,
    "args": {
      "External id": 2485008,"kernel_hash": "cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/g4/cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [65536, 2048], []], "Ev Idx": 1807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955490.886, "dur": 24.299,
    "args": {
      "External id": 2485009,"kernel_hash": "c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/4n/c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 1808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955545.454, "dur": 22.688,
    "args": {
      "External id": 2485010,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 1809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955591.064, "dur": 13.390,
    "args": {
      "External id": 2485011,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 1810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955619.147, "dur": 19.138,
    "args": {
      "External id": 2485012,"kernel_hash": "cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u7/cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [8388608, 2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [16, 4096, 2048], []], "Ev Idx": 1811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955707.588, "dur": 14.300,
    "args": {
      "External id": 2485013,"Record function id": 0, "Ev Idx": 1812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955710.666, "dur": 10.316,
    "args": {
      "External id": 2485014,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 1813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955714.742, "dur": 5.329,
    "args": {
      "External id": 2485015,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 1814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955716.244, "dur": 3.698,
    "args": {
      "External id": 2485016,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 1815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955725.593, "dur": 4.165,
    "args": {
      "External id": 2485017,"Record function id": 0, "Ev Idx": 1816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955726.784, "dur": 2.564,
    "args": {
      "External id": 2485018,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955727.555, "dur": 1.336,
    "args": {
      "External id": 2485019,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955728.046, "dur": 0.787,
    "args": {
      "External id": 2485020,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955732.936, "dur": 6.018,
    "args": {
      "External id": 2485021,"Record function id": 0, "Ev Idx": 1820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955733.988, "dur": 4.565,
    "args": {
      "External id": 2485022,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 1821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955734.638, "dur": 3.489,
    "args": {
      "External id": 2485023,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 1822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955735.380, "dur": 2.626,
    "args": {
      "External id": 2485024,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 1823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955742.120, "dur": 3.927,
    "args": {
      "External id": 2485025,"Record function id": 0, "Ev Idx": 1824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955743.218, "dur": 2.429,
    "args": {
      "External id": 2485026,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 1825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955743.784, "dur": 1.414,
    "args": {
      "External id": 2485027,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 1826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955744.172, "dur": 0.960,
    "args": {
      "External id": 2485028,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 1827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955749.222, "dur": 3.671,
    "args": {
      "External id": 2485029,"Record function id": 0, "Ev Idx": 1828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955750.195, "dur": 2.272,
    "args": {
      "External id": 2485030,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955750.648, "dur": 1.365,
    "args": {
      "External id": 2485031,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955751.191, "dur": 0.755,
    "args": {
      "External id": 2485032,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955755.972, "dur": 5.825,
    "args": {
      "External id": 2485033,"Record function id": 0, "Ev Idx": 1832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955758.840, "dur": 2.506,
    "args": {
      "External id": 2485034,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955759.658, "dur": 1.264,
    "args": {
      "External id": 2485035,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955760.178, "dur": 0.677,
    "args": {
      "External id": 2485036,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955764.960, "dur": 3.481,
    "args": {
      "External id": 2485037,"Record function id": 0, "Ev Idx": 1836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955766.024, "dur": 2.002,
    "args": {
      "External id": 2485038,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955766.692, "dur": 0.881,
    "args": {
      "External id": 2485039,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955766.991, "dur": 0.516,
    "args": {
      "External id": 2485040,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955771.493, "dur": 5.382,
    "args": {
      "External id": 2485041,"Record function id": 0, "Ev Idx": 1840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955772.346, "dur": 4.136,
    "args": {
      "External id": 2485042,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955772.852, "dur": 3.165,
    "args": {
      "External id": 2485043,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955775.353, "dur": 0.599,
    "args": {
      "External id": 2485044,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955779.938, "dur": 3.074,
    "args": {
      "External id": 2485045,"Record function id": 0, "Ev Idx": 1844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955780.764, "dur": 1.863,
    "args": {
      "External id": 2485046,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955781.206, "dur": 0.998,
    "args": {
      "External id": 2485047,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955781.504, "dur": 0.635,
    "args": {
      "External id": 2485048,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955786.585, "dur": 38770.019,
    "args": {
      "External id": 2485049,"Record function id": 0, "Sequence number": 24740473, "Fwd thread id": 1, "Ev Idx": 1848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955788.035, "dur": 38759.355,
    "args": {
      "External id": 2485050,"Sequence number": 24740473, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 1849
    }
  },
  {
    "ph": "f", "id": 175, "pid": 1336759, "tid": 1381173, "ts": 1514656955788.035,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.26)", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955815.842, "dur": 35.687,
    "args": {
      "External id": 2485051,"Record function id": 0, "Ev Idx": 1850
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.26)", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955858.747, "dur": 56.238,
    "args": {
      "External id": 2485052,"Record function id": 0, "Ev Idx": 1851
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.26)", "pid": 1336759, "tid": 1381173,
    "ts": 1514656955922.778, "dur": 38602.296,
    "args": {
      "External id": 2485053,"Record function id": 0, "Ev Idx": 1852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656956039.868, "dur": 7.828,
    "args": {
      "External id": 2485054,"Record function id": 0, "Concrete Inputs": ["[51384320]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656956057.873, "dur": 6.843,
    "args": {
      "External id": 2485055,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 1854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514656956083.824, "dur": 37556.409,
    "args": {
      "External id": 2485056,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 1855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514656956097.153, "dur": 37534.186,
    "args": {
      "External id": 2485057,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 1856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656956151.739, "dur": 13.854,
    "args": {
      "External id": 2485058,"Record function id": 0, "Concrete Inputs": ["[12602]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514656956173.645, "dur": 37422.518,
    "args": {
      "External id": 2485059,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], [], []], "Ev Idx": 1858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514656956175.999, "dur": 37419.552,
    "args": {
      "External id": 2485060,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], []], "Ev Idx": 1859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656956181.265, "dur": 4.903,
    "args": {
      "External id": 2485061,"Record function id": 0, "Concrete Inputs": ["[12602]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656956187.657, "dur": 37403.137,
    "args": {
      "External id": 2485062,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[12602], [12602], []], "Ev Idx": 1861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656993726.899, "dur": 8.853,
    "args": {
      "External id": 2485063,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[51384320], [], [], [], [], []], "Ev Idx": 1862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656993729.870, "dur": 5.496,
    "args": {
      "External id": 2485064,"Record function id": 0, "Concrete Inputs": ["[6423040]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656993763.790, "dur": 468.723,
    "args": {
      "External id": 2485065,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[6423040], [51384320], [], [], [], []], "Ev Idx": 1864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514656993796.727, "dur": 430.289,
    "args": {
      "External id": 2485066,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 6423040, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[51384320], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 1865, "In msg nelems": 51384320
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514656993810.561, "dur": 409.388,
    "args": {
      "External id": 2485067,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[51384320]], "Ev Idx": 1866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514656994257.532, "dur": 2.610,
    "args": {
      "External id": 2485068,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 1867, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656994321.138, "dur": 6.948,
    "args": {
      "External id": 2485069,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656994371.075, "dur": 1.413,
    "args": {
      "External id": 2485070,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "256"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656994387.058, "dur": 4.068,
    "args": {
      "External id": 2485071,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656994404.500, "dur": 1.051,
    "args": {
      "External id": 2485072,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1048832"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656994415.335, "dur": 1.080,
    "args": {
      "External id": 2485073,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1573120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656994427.819, "dur": 0.920,
    "args": {
      "External id": 2485074,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656994437.477, "dur": 3.500,
    "args": {
      "External id": 2485075,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "2097664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656994450.813, "dur": 1.972,
    "args": {
      "External id": 2485076,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "3539456"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656994464.437, "dur": 1.137,
    "args": {
      "External id": 2485077,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "4981248"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656994572.846, "dur": 2677.820,
    "args": {
      "External id": 2485078,"Record function id": 0, "Ev Idx": 1877
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.25)", "pid": 1336759, "tid": 1381173,
    "ts": 1514656994592.734, "dur": 1033.173,
    "args": {
      "External id": 2485079,"Record function id": 0, "Ev Idx": 1878
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.25)", "pid": 1336759, "tid": 1381173,
    "ts": 1514656994609.798, "dur": 310.971,
    "args": {
      "External id": 2485080,"Record function id": 0, "Ev Idx": 1879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656994685.193, "dur": 4.382,
    "args": {
      "External id": 2485081,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 1880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656994692.796, "dur": 1.003,
    "args": {
      "External id": 2485082,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 1881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656994695.715, "dur": 2.931,
    "args": {
      "External id": 2485083,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 1882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656994699.970, "dur": 0.900,
    "args": {
      "External id": 2485084,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 1883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656994704.386, "dur": 0.806,
    "args": {
      "External id": 2485085,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 1884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656994706.590, "dur": 0.893,
    "args": {
      "External id": 2485086,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 1885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656994711.142, "dur": 1.764,
    "args": {
      "External id": 2485087,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 1886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656994714.510, "dur": 0.712,
    "args": {
      "External id": 2485088,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 1887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656994718.693, "dur": 0.916,
    "args": {
      "External id": 2485089,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 1888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656994720.716, "dur": 0.822,
    "args": {
      "External id": 2485090,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 1889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514656994739.434, "dur": 152.107,
    "args": {
      "External id": 2485091,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 1890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514656994757.841, "dur": 129.065,
    "args": {
      "External id": 2485092,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 1891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656994778.441, "dur": 15.425,
    "args": {
      "External id": 2485093,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514656994796.770, "dur": 63.871,
    "args": {
      "External id": 2485094,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 1893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514656994801.338, "dur": 59.011,
    "args": {
      "External id": 2485095,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 1894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656994804.711, "dur": 4.734,
    "args": {
      "External id": 2485096,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656994811.041, "dur": 48.510,
    "args": {
      "External id": 2485097,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 1896
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.24", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995031.615, "dur": 586.781,
    "args": {
      "External id": 2485098,"Record function id": 0, "Ev Idx": 1897
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.24)", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995048.044, "dur": 557.553,
    "args": {
      "External id": 2485099,"Record function id": 0, "Ev Idx": 1898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995105.619, "dur": 6.091,
    "args": {
      "External id": 2485100,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995127.217, "dur": 36.129,
    "args": {
      "External id": 2485101,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 1900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995133.538, "dur": 1.573,
    "args": {
      "External id": 2485102,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995138.146, "dur": 0.606,
    "args": {
      "External id": 2485103,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995139.468, "dur": 0.467,
    "args": {
      "External id": 2485104,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995142.956, "dur": 2.059,
    "args": {
      "External id": 2485105,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995145.902, "dur": 0.386,
    "args": {
      "External id": 2485106,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995148.755, "dur": 3.015,
    "args": {
      "External id": 2485107,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995153.793, "dur": 0.337,
    "args": {
      "External id": 2485108,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995154.585, "dur": 0.394,
    "args": {
      "External id": 2485109,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995157.219, "dur": 0.609,
    "args": {
      "External id": 2485110,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995172.250, "dur": 32.470,
    "args": {
      "External id": 2485111,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 1910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995234.551, "dur": 111.543,
    "args": {
      "External id": 2485112,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 1911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995244.034, "dur": 3.101,
    "args": {
      "External id": 2485113,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995251.843, "dur": 12.517,
    "args": {
      "External id": 2485114,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 1913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995255.805, "dur": 8.098,
    "args": {
      "External id": 2485115,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 1914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995260.625, "dur": 1.818,
    "args": {
      "External id": 2485116,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 1915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995270.920, "dur": 37.605,
    "args": {
      "External id": 2485117,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 1916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995272.217, "dur": 0.328,
    "args": {
      "External id": 2485118,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995275.221, "dur": 0.524,
    "args": {
      "External id": 2485119,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995286.497, "dur": 2.845,
    "args": {
      "External id": 2485120,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995293.528, "dur": 0.501,
    "args": {
      "External id": 2485121,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995294.785, "dur": 0.507,
    "args": {
      "External id": 2485122,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995295.728, "dur": 2.199,
    "args": {
      "External id": 2485123,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995299.921, "dur": 0.362,
    "args": {
      "External id": 2485124,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995300.769, "dur": 0.332,
    "args": {
      "External id": 2485125,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995303.605, "dur": 0.234,
    "args": {
      "External id": 2485126,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 1925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995318.888, "dur": 19.620,
    "args": {
      "External id": 2485127,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 1926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995386.853, "dur": 125.550,
    "args": {
      "External id": 2485128,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 1927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995417.422, "dur": 91.599,
    "args": {
      "External id": 2485129,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 1928, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995426.528, "dur": 78.091,
    "args": {
      "External id": 2485130,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 1929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995548.350, "dur": 2.788,
    "args": {
      "External id": 2485131,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 1930, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995633.601, "dur": 1597.742,
    "args": {
      "External id": 2485132,"Sequence number": 24740472, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 1931
    }
  },
  {
    "ph": "f", "id": 176, "pid": 1336759, "tid": 1381173, "ts": 1514656995633.601,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995738.658, "dur": 102.405,
    "args": {
      "External id": 2485133,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [5632, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 1932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995880.306, "dur": 34.935,
    "args": {
      "External id": 2485134,"kernel_hash": "cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/m5/cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [65536, 5632], [16, 4096, 5632], [16, 4096, 5632], [16, 4096, 5632], []], "Ev Idx": 1933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656995930.010, "dur": 47.368,
    "args": {
      "External id": 2485135,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 2048], [0, 5632, 1], [11534336, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632], [1, 2048, 5632]], "Ev Idx": 1934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656996028.307, "dur": 37.779,
    "args": {
      "External id": 2485136,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 1935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656996072.878, "dur": 45.030,
    "args": {
      "External id": 2485137,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 1936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656996124.760, "dur": 27.023,
    "args": {
      "External id": 2485138,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 1937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656996165.852, "dur": 41.768,
    "args": {
      "External id": 2485139,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 1938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514656996238.271, "dur": 25.183,
    "args": {
      "External id": 2485140,"kernel_hash": "c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/47/c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 1939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514656996284.509, "dur": 28.135,
    "args": {
      "External id": 2485141,"kernel_hash": "cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/xr/cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 1940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514656996334.330, "dur": 17.968,
    "args": {
      "External id": 2485142,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 1941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514656996371.017, "dur": 14.708,
    "args": {
      "External id": 2485143,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 1942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656996394.206, "dur": 28.560,
    "args": {
      "External id": 2485144,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 1943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656996425.845, "dur": 32.326,
    "args": {
      "External id": 2485145,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 1944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656996485.329, "dur": 186.818,
    "args": {
      "External id": 2485146,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 1945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656996579.638, "dur": 8.835,
    "args": {
      "External id": 2485147,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656996590.525, "dur": 2.222,
    "args": {
      "External id": 2485148,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 1947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514656996704.777, "dur": 22.098,
    "args": {
      "External id": 2485149,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 1948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514656996739.704, "dur": 13.093,
    "args": {
      "External id": 2485150,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 1949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656996760.369, "dur": 38.209,
    "args": {
      "External id": 2485151,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 1950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656996804.074, "dur": 34.663,
    "args": {
      "External id": 2485152,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 1951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656996847.445, "dur": 20.179,
    "args": {
      "External id": 2485153,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 1952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656996871.677, "dur": 29.138,
    "args": {
      "External id": 2485154,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 1953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656996906.416, "dur": 18.776,
    "args": {
      "External id": 2485155,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 1954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514656996931.797, "dur": 28.701,
    "args": {
      "External id": 2485156,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 1955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997029.436, "dur": 31.698,
    "args": {
      "External id": 2485157,"kernel_hash": "cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/g4/cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [65536, 2048], []], "Ev Idx": 1956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997085.958, "dur": 22.346,
    "args": {
      "External id": 2485158,"kernel_hash": "c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/4n/c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 1957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997122.720, "dur": 19.264,
    "args": {
      "External id": 2485159,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 1958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997159.202, "dur": 13.154,
    "args": {
      "External id": 2485160,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 1959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997186.046, "dur": 18.568,
    "args": {
      "External id": 2485161,"kernel_hash": "cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u7/cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [8388608, 2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [16, 4096, 2048], []], "Ev Idx": 1960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997272.978, "dur": 14.595,
    "args": {
      "External id": 2485162,"Record function id": 0, "Ev Idx": 1961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997276.090, "dur": 10.453,
    "args": {
      "External id": 2485163,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 1962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997280.054, "dur": 5.668,
    "args": {
      "External id": 2485164,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 1963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997281.616, "dur": 3.989,
    "args": {
      "External id": 2485165,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 1964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997291.402, "dur": 8.492,
    "args": {
      "External id": 2485166,"Record function id": 0, "Ev Idx": 1965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997293.869, "dur": 5.514,
    "args": {
      "External id": 2485167,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997294.799, "dur": 4.139,
    "args": {
      "External id": 2485168,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997295.460, "dur": 3.389,
    "args": {
      "External id": 2485169,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997303.213, "dur": 3.995,
    "args": {
      "External id": 2485170,"Record function id": 0, "Ev Idx": 1969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997304.362, "dur": 2.443,
    "args": {
      "External id": 2485171,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 1970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997304.873, "dur": 1.513,
    "args": {
      "External id": 2485172,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 1971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997305.503, "dur": 0.748,
    "args": {
      "External id": 2485173,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 1972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997310.386, "dur": 3.496,
    "args": {
      "External id": 2485174,"Record function id": 0, "Ev Idx": 1973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997311.329, "dur": 2.140,
    "args": {
      "External id": 2485175,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 1974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997311.938, "dur": 1.106,
    "args": {
      "External id": 2485176,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 1975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997312.319, "dur": 0.660,
    "args": {
      "External id": 2485177,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 1976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997316.886, "dur": 3.676,
    "args": {
      "External id": 2485178,"Record function id": 0, "Ev Idx": 1977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997317.940, "dur": 2.233,
    "args": {
      "External id": 2485179,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997318.641, "dur": 1.063,
    "args": {
      "External id": 2485180,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997318.894, "dur": 0.747,
    "args": {
      "External id": 2485181,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997323.565, "dur": 3.653,
    "args": {
      "External id": 2485182,"Record function id": 0, "Ev Idx": 1981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997324.641, "dur": 2.193,
    "args": {
      "External id": 2485183,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997325.152, "dur": 1.262,
    "args": {
      "External id": 2485184,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997325.629, "dur": 0.719,
    "args": {
      "External id": 2485185,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997330.378, "dur": 5.349,
    "args": {
      "External id": 2485186,"Record function id": 0, "Ev Idx": 1985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997331.549, "dur": 3.783,
    "args": {
      "External id": 2485187,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997332.203, "dur": 2.725,
    "args": {
      "External id": 2485188,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997334.280, "dur": 0.581,
    "args": {
      "External id": 2485189,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997338.794, "dur": 3.333,
    "args": {
      "External id": 2485190,"Record function id": 0, "Ev Idx": 1989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997339.752, "dur": 1.982,
    "args": {
      "External id": 2485191,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997340.248, "dur": 1.073,
    "args": {
      "External id": 2485192,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997340.529, "dur": 0.723,
    "args": {
      "External id": 2485193,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 1992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997345.182, "dur": 5.338,
    "args": {
      "External id": 2485194,"Record function id": 0, "Ev Idx": 1993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997346.129, "dur": 3.998,
    "args": {
      "External id": 2485195,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997346.697, "dur": 3.028,
    "args": {
      "External id": 2485196,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997347.131, "dur": 2.532,
    "args": {
      "External id": 2485197,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 1996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997353.987, "dur": 36707.507,
    "args": {
      "External id": 2485198,"Record function id": 0, "Sequence number": 24740471, "Fwd thread id": 1, "Ev Idx": 1997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997355.123, "dur": 36696.500,
    "args": {
      "External id": 2485199,"Sequence number": 24740471, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 1998
    }
  },
  {
    "ph": "f", "id": 177, "pid": 1336759, "tid": 1381173, "ts": 1514656997355.123,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.25)", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997385.559, "dur": 37.549,
    "args": {
      "External id": 2485200,"Record function id": 0, "Ev Idx": 1999
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.25)", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997430.824, "dur": 58.252,
    "args": {
      "External id": 2485201,"Record function id": 0, "Ev Idx": 2000
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.25)", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997498.634, "dur": 36543.842,
    "args": {
      "External id": 2485202,"Record function id": 0, "Ev Idx": 2001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997603.135, "dur": 6.948,
    "args": {
      "External id": 2485203,"Record function id": 0, "Concrete Inputs": ["[51384320]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997622.032, "dur": 5.093,
    "args": {
      "External id": 2485204,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 2003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997640.814, "dur": 35569.915,
    "args": {
      "External id": 2485205,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 2004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997653.259, "dur": 35547.590,
    "args": {
      "External id": 2485206,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 2005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997700.450, "dur": 13.514,
    "args": {
      "External id": 2485207,"Record function id": 0, "Concrete Inputs": ["[12602]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997721.288, "dur": 35438.180,
    "args": {
      "External id": 2485208,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], [], []], "Ev Idx": 2007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997723.811, "dur": 35434.782,
    "args": {
      "External id": 2485209,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], []], "Ev Idx": 2008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997727.529, "dur": 5.394,
    "args": {
      "External id": 2485210,"Record function id": 0, "Concrete Inputs": ["[12602]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514656997734.736, "dur": 35419.740,
    "args": {
      "External id": 2485211,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[12602], [12602], []], "Ev Idx": 2010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657033303.803, "dur": 9.521,
    "args": {
      "External id": 2485212,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[51384320], [], [], [], [], []], "Ev Idx": 2011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657033306.715, "dur": 6.249,
    "args": {
      "External id": 2485213,"Record function id": 0, "Concrete Inputs": ["[6423040]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657033340.028, "dur": 370.549,
    "args": {
      "External id": 2485214,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[6423040], [51384320], [], [], [], []], "Ev Idx": 2013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657033372.887, "dur": 332.619,
    "args": {
      "External id": 2485215,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 6423040, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[51384320], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 2014, "In msg nelems": 51384320
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657033384.892, "dur": 314.842,
    "args": {
      "External id": 2485216,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[51384320]], "Ev Idx": 2015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657033734.221, "dur": 2.631,
    "args": {
      "External id": 2485217,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 2016, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657033798.268, "dur": 6.845,
    "args": {
      "External id": 2485218,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657033851.664, "dur": 3.503,
    "args": {
      "External id": 2485219,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "256"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657033869.824, "dur": 0.908,
    "args": {
      "External id": 2485220,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657033882.076, "dur": 1.017,
    "args": {
      "External id": 2485221,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1048832"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657033892.920, "dur": 0.810,
    "args": {
      "External id": 2485222,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1573120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657033902.434, "dur": 2.903,
    "args": {
      "External id": 2485223,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657033921.347, "dur": 0.967,
    "args": {
      "External id": 2485224,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "2097664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657033931.813, "dur": 2.433,
    "args": {
      "External id": 2485225,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "3539456"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657033942.152, "dur": 0.978,
    "args": {
      "External id": 2485226,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "4981248"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034077.826, "dur": 2652.944,
    "args": {
      "External id": 2485227,"Record function id": 0, "Ev Idx": 2026
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.24)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034099.997, "dur": 1041.243,
    "args": {
      "External id": 2485228,"Record function id": 0, "Ev Idx": 2027
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.24)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034114.462, "dur": 317.992,
    "args": {
      "External id": 2485229,"Record function id": 0, "Ev Idx": 2028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034194.787, "dur": 6.063,
    "args": {
      "External id": 2485230,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 2029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034203.769, "dur": 0.924,
    "args": {
      "External id": 2485231,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 2030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034208.553, "dur": 0.672,
    "args": {
      "External id": 2485232,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 2031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034210.656, "dur": 0.975,
    "args": {
      "External id": 2485233,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 2032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034212.898, "dur": 0.908,
    "args": {
      "External id": 2485234,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 2033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034215.091, "dur": 0.822,
    "args": {
      "External id": 2485235,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 2034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034219.352, "dur": 2.054,
    "args": {
      "External id": 2485236,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 2035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034222.617, "dur": 0.861,
    "args": {
      "External id": 2485237,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 2036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034224.874, "dur": 6.183,
    "args": {
      "External id": 2485238,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 2037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034234.705, "dur": 0.917,
    "args": {
      "External id": 2485239,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 2038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034255.540, "dur": 149.110,
    "args": {
      "External id": 2485240,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 2039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034271.790, "dur": 128.362,
    "args": {
      "External id": 2485241,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 2040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034291.567, "dur": 12.876,
    "args": {
      "External id": 2485242,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034307.389, "dur": 66.231,
    "args": {
      "External id": 2485243,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 2042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034310.002, "dur": 63.281,
    "args": {
      "External id": 2485244,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 2043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034314.268, "dur": 7.102,
    "args": {
      "External id": 2485245,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034323.126, "dur": 49.606,
    "args": {
      "External id": 2485246,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 2045
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.23", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034504.059, "dur": 629.984,
    "args": {
      "External id": 2485247,"Record function id": 0, "Ev Idx": 2046
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.23)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034518.218, "dur": 603.619,
    "args": {
      "External id": 2485248,"Record function id": 0, "Ev Idx": 2047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034613.720, "dur": 6.173,
    "args": {
      "External id": 2485249,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034635.863, "dur": 33.163,
    "args": {
      "External id": 2485250,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 2049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034640.593, "dur": 2.776,
    "args": {
      "External id": 2485251,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034645.465, "dur": 0.377,
    "args": {
      "External id": 2485252,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034647.183, "dur": 0.333,
    "args": {
      "External id": 2485253,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034650.145, "dur": 2.503,
    "args": {
      "External id": 2485254,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034653.835, "dur": 0.291,
    "args": {
      "External id": 2485255,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034655.211, "dur": 0.417,
    "args": {
      "External id": 2485256,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034658.098, "dur": 0.279,
    "args": {
      "External id": 2485257,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034660.171, "dur": 0.253,
    "args": {
      "External id": 2485258,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034661.629, "dur": 1.282,
    "args": {
      "External id": 2485259,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034678.214, "dur": 33.608,
    "args": {
      "External id": 2485260,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 2059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034742.770, "dur": 102.020,
    "args": {
      "External id": 2485261,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 2060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034752.526, "dur": 3.174,
    "args": {
      "External id": 2485262,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034760.587, "dur": 9.847,
    "args": {
      "External id": 2485263,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 2062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034764.586, "dur": 5.431,
    "args": {
      "External id": 2485264,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 2063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034768.297, "dur": 0.575,
    "args": {
      "External id": 2485265,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 2064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034776.810, "dur": 28.222,
    "args": {
      "External id": 2485266,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 2065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034778.568, "dur": 2.452,
    "args": {
      "External id": 2485267,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034782.552, "dur": 0.406,
    "args": {
      "External id": 2485268,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034784.634, "dur": 1.584,
    "args": {
      "External id": 2485269,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034787.531, "dur": 0.297,
    "args": {
      "External id": 2485270,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034789.304, "dur": 0.328,
    "args": {
      "External id": 2485271,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034792.282, "dur": 0.314,
    "args": {
      "External id": 2485272,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034793.831, "dur": 0.297,
    "args": {
      "External id": 2485273,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034795.518, "dur": 0.314,
    "args": {
      "External id": 2485274,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034798.127, "dur": 2.034,
    "args": {
      "External id": 2485275,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034819.229, "dur": 17.963,
    "args": {
      "External id": 2485276,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 2075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034889.149, "dur": 155.391,
    "args": {
      "External id": 2485277,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 2076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034917.950, "dur": 122.767,
    "args": {
      "External id": 2485278,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 2077, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657034926.761, "dur": 109.462,
    "args": {
      "External id": 2485279,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 2078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657035064.822, "dur": 2.035,
    "args": {
      "External id": 2485280,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 2079, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657035148.127, "dur": 1562.548,
    "args": {
      "External id": 2485281,"Sequence number": 24740470, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 2080
    }
  },
  {
    "ph": "f", "id": 178, "pid": 1336759, "tid": 1381173, "ts": 1514657035148.127,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657035252.622, "dur": 103.630,
    "args": {
      "External id": 2485282,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [5632, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 2081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657035396.507, "dur": 35.250,
    "args": {
      "External id": 2485283,"kernel_hash": "cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/m5/cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [65536, 5632], [16, 4096, 5632], [16, 4096, 5632], [16, 4096, 5632], []], "Ev Idx": 2082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657035449.283, "dur": 47.016,
    "args": {
      "External id": 2485284,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 2048], [0, 5632, 1], [11534336, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632], [1, 2048, 5632]], "Ev Idx": 2083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657035505.501, "dur": 48.718,
    "args": {
      "External id": 2485285,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 2084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657035563.556, "dur": 47.289,
    "args": {
      "External id": 2485286,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 2085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657035617.308, "dur": 26.663,
    "args": {
      "External id": 2485287,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 2086
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657035652.949, "dur": 41.082,
    "args": {
      "External id": 2485288,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 2087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657035719.835, "dur": 22.741,
    "args": {
      "External id": 2485289,"kernel_hash": "c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/47/c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 2088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657035764.073, "dur": 27.799,
    "args": {
      "External id": 2485290,"kernel_hash": "cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/xr/cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657035810.032, "dur": 22.834,
    "args": {
      "External id": 2485291,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 2090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657035849.160, "dur": 13.759,
    "args": {
      "External id": 2485292,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 2091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657035872.750, "dur": 29.628,
    "args": {
      "External id": 2485293,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 2092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657035905.276, "dur": 32.310,
    "args": {
      "External id": 2485294,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 2093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657035968.522, "dur": 203.268,
    "args": {
      "External id": 2485295,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 2094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036081.068, "dur": 6.300,
    "args": {
      "External id": 2485296,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036089.337, "dur": 2.329,
    "args": {
      "External id": 2485297,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036202.761, "dur": 23.947,
    "args": {
      "External id": 2485298,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036237.997, "dur": 13.955,
    "args": {
      "External id": 2485299,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036259.592, "dur": 38.321,
    "args": {
      "External id": 2485300,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 2099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036304.933, "dur": 34.969,
    "args": {
      "External id": 2485301,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 2100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036347.257, "dur": 19.741,
    "args": {
      "External id": 2485302,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 2101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036370.961, "dur": 29.399,
    "args": {
      "External id": 2485303,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 2102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036405.854, "dur": 18.764,
    "args": {
      "External id": 2485304,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 2103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036433.411, "dur": 42.776,
    "args": {
      "External id": 2485305,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 2104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036500.854, "dur": 20.743,
    "args": {
      "External id": 2485306,"kernel_hash": "cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/g4/cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [65536, 2048], []], "Ev Idx": 2105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036560.371, "dur": 27.941,
    "args": {
      "External id": 2485307,"kernel_hash": "c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/4n/c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036603.636, "dur": 15.699,
    "args": {
      "External id": 2485308,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 2107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036640.880, "dur": 12.988,
    "args": {
      "External id": 2485309,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 2108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036667.311, "dur": 16.839,
    "args": {
      "External id": 2485310,"kernel_hash": "cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u7/cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [8388608, 2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [16, 4096, 2048], []], "Ev Idx": 2109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036752.637, "dur": 16.392,
    "args": {
      "External id": 2485311,"Record function id": 0, "Ev Idx": 2110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036756.023, "dur": 12.117,
    "args": {
      "External id": 2485312,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036760.205, "dur": 6.954,
    "args": {
      "External id": 2485313,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036761.664, "dur": 5.370,
    "args": {
      "External id": 2485314,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036772.907, "dur": 5.000,
    "args": {
      "External id": 2485315,"Record function id": 0, "Ev Idx": 2114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036774.477, "dur": 2.998,
    "args": {
      "External id": 2485316,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036775.367, "dur": 1.663,
    "args": {
      "External id": 2485317,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036776.128, "dur": 0.817,
    "args": {
      "External id": 2485318,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036781.189, "dur": 4.100,
    "args": {
      "External id": 2485319,"Record function id": 0, "Ev Idx": 2118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036782.346, "dur": 2.527,
    "args": {
      "External id": 2485320,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036783.157, "dur": 1.283,
    "args": {
      "External id": 2485321,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036783.771, "dur": 0.590,
    "args": {
      "External id": 2485322,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036788.533, "dur": 3.844,
    "args": {
      "External id": 2485323,"Record function id": 0, "Ev Idx": 2122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036789.806, "dur": 2.183,
    "args": {
      "External id": 2485324,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 2123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036790.401, "dur": 1.166,
    "args": {
      "External id": 2485325,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 2124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036790.948, "dur": 0.552,
    "args": {
      "External id": 2485326,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 2125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036795.413, "dur": 3.411,
    "args": {
      "External id": 2485327,"Record function id": 0, "Ev Idx": 2126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036796.519, "dur": 1.880,
    "args": {
      "External id": 2485328,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036797.043, "dur": 0.946,
    "args": {
      "External id": 2485329,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036797.299, "dur": 0.627,
    "args": {
      "External id": 2485330,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036801.794, "dur": 3.428,
    "args": {
      "External id": 2485331,"Record function id": 0, "Ev Idx": 2130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036802.948, "dur": 1.891,
    "args": {
      "External id": 2485332,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036803.439, "dur": 0.987,
    "args": {
      "External id": 2485333,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036803.796, "dur": 0.558,
    "args": {
      "External id": 2485334,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036808.354, "dur": 5.546,
    "args": {
      "External id": 2485335,"Record function id": 0, "Ev Idx": 2134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036809.676, "dur": 3.821,
    "args": {
      "External id": 2485336,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036810.172, "dur": 2.940,
    "args": {
      "External id": 2485337,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036812.391, "dur": 0.653,
    "args": {
      "External id": 2485338,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036816.947, "dur": 5.741,
    "args": {
      "External id": 2485339,"Record function id": 0, "Ev Idx": 2138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036818.288, "dur": 3.996,
    "args": {
      "External id": 2485340,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036818.801, "dur": 3.091,
    "args": {
      "External id": 2485341,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036819.075, "dur": 2.757,
    "args": {
      "External id": 2485342,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036825.718, "dur": 3.370,
    "args": {
      "External id": 2485343,"Record function id": 0, "Ev Idx": 2142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036826.894, "dur": 1.771,
    "args": {
      "External id": 2485344,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036827.349, "dur": 0.932,
    "args": {
      "External id": 2485345,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036827.601, "dur": 0.616,
    "args": {
      "External id": 2485346,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036832.501, "dur": 37750.831,
    "args": {
      "External id": 2485347,"Record function id": 0, "Sequence number": 24740469, "Fwd thread id": 1, "Ev Idx": 2146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036834.056, "dur": 37740.758,
    "args": {
      "External id": 2485348,"Sequence number": 24740469, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 2147
    }
  },
  {
    "ph": "f", "id": 179, "pid": 1336759, "tid": 1381173, "ts": 1514657036834.056,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.24)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036861.138, "dur": 38.327,
    "args": {
      "External id": 2485349,"Record function id": 0, "Ev Idx": 2148
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.24)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036907.132, "dur": 58.197,
    "args": {
      "External id": 2485350,"Record function id": 0, "Ev Idx": 2149
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.24)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657036973.249, "dur": 37593.380,
    "args": {
      "External id": 2485351,"Record function id": 0, "Ev Idx": 2150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657037096.951, "dur": 7.344,
    "args": {
      "External id": 2485352,"Record function id": 0, "Concrete Inputs": ["[51384320]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657037114.140, "dur": 4.951,
    "args": {
      "External id": 2485353,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 2152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657037133.760, "dur": 36507.210,
    "args": {
      "External id": 2485354,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 2153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657037147.537, "dur": 36484.350,
    "args": {
      "External id": 2485355,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 2154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657037198.166, "dur": 12.737,
    "args": {
      "External id": 2485356,"Record function id": 0, "Concrete Inputs": ["[12602]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657037217.480, "dur": 36377.730,
    "args": {
      "External id": 2485357,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], [], []], "Ev Idx": 2156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657037221.179, "dur": 36373.343,
    "args": {
      "External id": 2485358,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], []], "Ev Idx": 2157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657037225.098, "dur": 4.684,
    "args": {
      "External id": 2485359,"Record function id": 0, "Concrete Inputs": ["[12602]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657037231.417, "dur": 36359.686,
    "args": {
      "External id": 2485360,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[12602], [12602], []], "Ev Idx": 2159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657073729.728, "dur": 8.565,
    "args": {
      "External id": 2485361,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[51384320], [], [], [], [], []], "Ev Idx": 2160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657073732.488, "dur": 5.377,
    "args": {
      "External id": 2485362,"Record function id": 0, "Concrete Inputs": ["[6423040]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657073770.058, "dur": 461.971,
    "args": {
      "External id": 2485363,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[6423040], [51384320], [], [], [], []], "Ev Idx": 2162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657073800.512, "dur": 426.276,
    "args": {
      "External id": 2485364,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 6423040, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[51384320], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 2163, "In msg nelems": 51384320
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657073814.334, "dur": 406.026,
    "args": {
      "External id": 2485365,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[51384320]], "Ev Idx": 2164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657074259.300, "dur": 2.292,
    "args": {
      "External id": 2485366,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 2165, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657074330.089, "dur": 6.715,
    "args": {
      "External id": 2485367,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657074385.753, "dur": 1.366,
    "args": {
      "External id": 2485368,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "256"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657074403.195, "dur": 1.036,
    "args": {
      "External id": 2485369,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657074417.249, "dur": 0.801,
    "args": {
      "External id": 2485370,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1048832"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657074431.391, "dur": 0.865,
    "args": {
      "External id": 2485371,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1573120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657074448.863, "dur": 0.819,
    "args": {
      "External id": 2485372,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657074462.013, "dur": 0.928,
    "args": {
      "External id": 2485373,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "2097664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657074474.615, "dur": 2.640,
    "args": {
      "External id": 2485374,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "3539456"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657074488.932, "dur": 0.787,
    "args": {
      "External id": 2485375,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "4981248"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657074599.808, "dur": 2646.934,
    "args": {
      "External id": 2485376,"Record function id": 0, "Ev Idx": 2175
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.23)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657074617.705, "dur": 1016.554,
    "args": {
      "External id": 2485377,"Record function id": 0, "Ev Idx": 2176
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.23)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657074631.099, "dur": 306.324,
    "args": {
      "External id": 2485378,"Record function id": 0, "Ev Idx": 2177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657074705.672, "dur": 4.529,
    "args": {
      "External id": 2485379,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 2178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657074713.672, "dur": 0.874,
    "args": {
      "External id": 2485380,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 2179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657074716.554, "dur": 0.818,
    "args": {
      "External id": 2485381,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 2180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657074719.433, "dur": 1.051,
    "args": {
      "External id": 2485382,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 2181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657074723.775, "dur": 0.960,
    "args": {
      "External id": 2485383,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 2182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657074726.203, "dur": 0.951,
    "args": {
      "External id": 2485384,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 2183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657074728.684, "dur": 2.080,
    "args": {
      "External id": 2485385,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 2184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657074732.574, "dur": 2.864,
    "args": {
      "External id": 2485386,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 2185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657074739.017, "dur": 1.042,
    "args": {
      "External id": 2485387,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 2186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657074741.622, "dur": 0.828,
    "args": {
      "External id": 2485388,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 2187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657074759.682, "dur": 146.283,
    "args": {
      "External id": 2485389,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 2188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657074775.527, "dur": 125.702,
    "args": {
      "External id": 2485390,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 2189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657074795.310, "dur": 13.523,
    "args": {
      "External id": 2485391,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657074811.935, "dur": 64.217,
    "args": {
      "External id": 2485392,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 2191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657074814.518, "dur": 61.301,
    "args": {
      "External id": 2485393,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 2192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657074818.619, "dur": 4.951,
    "args": {
      "External id": 2485394,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657074827.198, "dur": 47.906,
    "args": {
      "External id": 2485395,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 2194
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.22", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075054.750, "dur": 572.365,
    "args": {
      "External id": 2485396,"Record function id": 0, "Ev Idx": 2195
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.22)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075071.983, "dur": 542.934,
    "args": {
      "External id": 2485397,"Record function id": 0, "Ev Idx": 2196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075127.612, "dur": 5.815,
    "args": {
      "External id": 2485398,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075149.431, "dur": 34.168,
    "args": {
      "External id": 2485399,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 2198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075154.535, "dur": 1.672,
    "args": {
      "External id": 2485400,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075159.348, "dur": 0.299,
    "args": {
      "External id": 2485401,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075162.436, "dur": 2.298,
    "args": {
      "External id": 2485402,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075166.229, "dur": 1.105,
    "args": {
      "External id": 2485403,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075168.850, "dur": 0.363,
    "args": {
      "External id": 2485404,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075170.674, "dur": 0.450,
    "args": {
      "External id": 2485405,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075173.998, "dur": 0.262,
    "args": {
      "External id": 2485406,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075175.934, "dur": 0.333,
    "args": {
      "External id": 2485407,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075177.651, "dur": 0.315,
    "args": {
      "External id": 2485408,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075193.183, "dur": 30.910,
    "args": {
      "External id": 2485409,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 2208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075254.592, "dur": 98.785,
    "args": {
      "External id": 2485410,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 2209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075263.942, "dur": 3.367,
    "args": {
      "External id": 2485411,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075272.059, "dur": 13.211,
    "args": {
      "External id": 2485412,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 2211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075276.200, "dur": 8.639,
    "args": {
      "External id": 2485413,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 2212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075279.564, "dur": 4.061,
    "args": {
      "External id": 2485414,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 2213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075291.833, "dur": 26.681,
    "args": {
      "External id": 2485415,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 2214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075293.635, "dur": 0.280,
    "args": {
      "External id": 2485416,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075295.327, "dur": 0.364,
    "args": {
      "External id": 2485417,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075297.044, "dur": 0.265,
    "args": {
      "External id": 2485418,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075299.640, "dur": 0.381,
    "args": {
      "External id": 2485419,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075301.462, "dur": 0.371,
    "args": {
      "External id": 2485420,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075303.052, "dur": 1.546,
    "args": {
      "External id": 2485421,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075305.901, "dur": 0.384,
    "args": {
      "External id": 2485422,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075307.904, "dur": 2.148,
    "args": {
      "External id": 2485423,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075313.004, "dur": 0.320,
    "args": {
      "External id": 2485424,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075327.938, "dur": 17.784,
    "args": {
      "External id": 2485425,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 2224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075396.416, "dur": 121.826,
    "args": {
      "External id": 2485426,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 2225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075425.452, "dur": 89.264,
    "args": {
      "External id": 2485427,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 2226, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075438.186, "dur": 72.021,
    "args": {
      "External id": 2485428,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 2227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075555.962, "dur": 2.980,
    "args": {
      "External id": 2485429,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 2228, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075641.575, "dur": 1584.586,
    "args": {
      "External id": 2485430,"Sequence number": 24740468, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 2229
    }
  },
  {
    "ph": "f", "id": 180, "pid": 1336759, "tid": 1381173, "ts": 1514657075641.575,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075743.957, "dur": 102.401,
    "args": {
      "External id": 2485431,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [5632, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 2230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075884.664, "dur": 35.147,
    "args": {
      "External id": 2485432,"kernel_hash": "cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/m5/cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [65536, 5632], [16, 4096, 5632], [16, 4096, 5632], [16, 4096, 5632], []], "Ev Idx": 2231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657075939.936, "dur": 90.095,
    "args": {
      "External id": 2485433,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 2048], [0, 5632, 1], [11534336, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632], [1, 2048, 5632]], "Ev Idx": 2232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657076043.670, "dur": 36.783,
    "args": {
      "External id": 2485434,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 2233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657076086.852, "dur": 44.525,
    "args": {
      "External id": 2485435,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 2234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657076139.790, "dur": 27.322,
    "args": {
      "External id": 2485436,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 2235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657076175.608, "dur": 41.547,
    "args": {
      "External id": 2485437,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 2236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657076243.707, "dur": 23.228,
    "args": {
      "External id": 2485438,"kernel_hash": "c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/47/c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 2237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657076287.084, "dur": 26.894,
    "args": {
      "External id": 2485439,"kernel_hash": "cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/xr/cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657076333.544, "dur": 16.782,
    "args": {
      "External id": 2485440,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 2239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657076365.319, "dur": 13.773,
    "args": {
      "External id": 2485441,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 2240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657076390.719, "dur": 26.864,
    "args": {
      "External id": 2485442,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 2241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657076420.717, "dur": 31.943,
    "args": {
      "External id": 2485443,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 2242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657076477.503, "dur": 179.757,
    "args": {
      "External id": 2485444,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 2243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657076566.612, "dur": 6.911,
    "args": {
      "External id": 2485445,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657076575.642, "dur": 2.396,
    "args": {
      "External id": 2485446,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657076689.664, "dur": 21.911,
    "args": {
      "External id": 2485447,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657076723.458, "dur": 13.976,
    "args": {
      "External id": 2485448,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657076746.243, "dur": 35.856,
    "args": {
      "External id": 2485449,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 2248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657076787.366, "dur": 33.313,
    "args": {
      "External id": 2485450,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 2249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657076836.785, "dur": 20.458,
    "args": {
      "External id": 2485451,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 2250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657076861.392, "dur": 30.389,
    "args": {
      "External id": 2485452,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 2251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657076899.742, "dur": 19.281,
    "args": {
      "External id": 2485453,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 2252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657076926.566, "dur": 28.878,
    "args": {
      "External id": 2485454,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 2253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 1336759, "tid": 1381173,
    "ts": 1514657076975.283, "dur": 68.303,
    "args": {
      "External id": 2485455,"kernel_hash": "cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/g4/cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [65536, 2048], []], "Ev Idx": 2254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077074.297, "dur": 24.249,
    "args": {
      "External id": 2485456,"kernel_hash": "c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/4n/c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077115.402, "dur": 15.046,
    "args": {
      "External id": 2485457,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 2256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077149.088, "dur": 16.385,
    "args": {
      "External id": 2485458,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 2257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077178.444, "dur": 15.273,
    "args": {
      "External id": 2485459,"kernel_hash": "cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u7/cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [8388608, 2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [16, 4096, 2048], []], "Ev Idx": 2258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077269.755, "dur": 14.327,
    "args": {
      "External id": 2485460,"Record function id": 0, "Ev Idx": 2259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077273.010, "dur": 10.092,
    "args": {
      "External id": 2485461,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077277.062, "dur": 5.093,
    "args": {
      "External id": 2485462,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077278.450, "dur": 3.611,
    "args": {
      "External id": 2485463,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077288.114, "dur": 4.893,
    "args": {
      "External id": 2485464,"Record function id": 0, "Ev Idx": 2263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077289.773, "dur": 2.821,
    "args": {
      "External id": 2485465,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077290.534, "dur": 1.597,
    "args": {
      "External id": 2485466,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077290.929, "dur": 1.118,
    "args": {
      "External id": 2485467,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077296.280, "dur": 3.950,
    "args": {
      "External id": 2485468,"Record function id": 0, "Ev Idx": 2267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077297.647, "dur": 2.196,
    "args": {
      "External id": 2485469,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077298.146, "dur": 1.265,
    "args": {
      "External id": 2485470,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077298.807, "dur": 0.506,
    "args": {
      "External id": 2485471,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077303.366, "dur": 4.040,
    "args": {
      "External id": 2485472,"Record function id": 0, "Ev Idx": 2271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077304.833, "dur": 2.192,
    "args": {
      "External id": 2485473,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 2272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077305.533, "dur": 1.062,
    "args": {
      "External id": 2485474,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 2273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077305.941, "dur": 0.575,
    "args": {
      "External id": 2485475,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 2274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077310.391, "dur": 3.655,
    "args": {
      "External id": 2485476,"Record function id": 0, "Ev Idx": 2275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077311.574, "dur": 2.084,
    "args": {
      "External id": 2485477,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077312.289, "dur": 0.951,
    "args": {
      "External id": 2485478,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077312.631, "dur": 0.542,
    "args": {
      "External id": 2485479,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077317.404, "dur": 4.047,
    "args": {
      "External id": 2485480,"Record function id": 0, "Ev Idx": 2279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077318.635, "dur": 2.408,
    "args": {
      "External id": 2485481,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077319.383, "dur": 1.202,
    "args": {
      "External id": 2485482,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077319.928, "dur": 0.591,
    "args": {
      "External id": 2485483,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077324.713, "dur": 5.320,
    "args": {
      "External id": 2485484,"Record function id": 0, "Ev Idx": 2283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077325.732, "dur": 3.905,
    "args": {
      "External id": 2485485,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077326.287, "dur": 2.920,
    "args": {
      "External id": 2485486,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077326.534, "dur": 2.605,
    "args": {
      "External id": 2485487,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077333.080, "dur": 6.136,
    "args": {
      "External id": 2485488,"Record function id": 0, "Ev Idx": 2287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077334.322, "dur": 4.460,
    "args": {
      "External id": 2485489,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077334.769, "dur": 3.596,
    "args": {
      "External id": 2485490,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077337.200, "dur": 1.102,
    "args": {
      "External id": 2485491,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077342.241, "dur": 3.311,
    "args": {
      "External id": 2485492,"Record function id": 0, "Ev Idx": 2291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077343.351, "dur": 1.814,
    "args": {
      "External id": 2485493,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077343.813, "dur": 0.937,
    "args": {
      "External id": 2485494,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077344.180, "dur": 0.503,
    "args": {
      "External id": 2485495,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077348.955, "dur": 36112.419,
    "args": {
      "External id": 2485496,"Record function id": 0, "Sequence number": 24740467, "Fwd thread id": 1, "Ev Idx": 2295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077350.182, "dur": 36103.225,
    "args": {
      "External id": 2485497,"Sequence number": 24740467, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 2296
    }
  },
  {
    "ph": "f", "id": 181, "pid": 1336759, "tid": 1381173, "ts": 1514657077350.182,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.23)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077378.596, "dur": 38.158,
    "args": {
      "External id": 2485498,"Record function id": 0, "Ev Idx": 2297
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.23)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077424.330, "dur": 58.101,
    "args": {
      "External id": 2485499,"Record function id": 0, "Ev Idx": 2298
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.23)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077488.110, "dur": 35957.335,
    "args": {
      "External id": 2485500,"Record function id": 0, "Ev Idx": 2299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077589.379, "dur": 7.492,
    "args": {
      "External id": 2485501,"Record function id": 0, "Concrete Inputs": ["[51384320]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077607.427, "dur": 5.267,
    "args": {
      "External id": 2485502,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 2301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077629.194, "dur": 35014.015,
    "args": {
      "External id": 2485503,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 2302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077641.983, "dur": 34992.225,
    "args": {
      "External id": 2485504,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 2303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077684.034, "dur": 13.190,
    "args": {
      "External id": 2485505,"Record function id": 0, "Concrete Inputs": ["[12602]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077703.466, "dur": 34894.877,
    "args": {
      "External id": 2485506,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], [], []], "Ev Idx": 2305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077706.957, "dur": 34890.622,
    "args": {
      "External id": 2485507,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], []], "Ev Idx": 2306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077710.844, "dur": 4.729,
    "args": {
      "External id": 2485508,"Record function id": 0, "Concrete Inputs": ["[12602]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657077717.281, "dur": 34876.815,
    "args": {
      "External id": 2485509,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[12602], [12602], []], "Ev Idx": 2308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657112730.584, "dur": 8.312,
    "args": {
      "External id": 2485510,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[51384320], [], [], [], [], []], "Ev Idx": 2309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657112733.636, "dur": 4.960,
    "args": {
      "External id": 2485511,"Record function id": 0, "Concrete Inputs": ["[6423040]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657112766.500, "dur": 381.153,
    "args": {
      "External id": 2485512,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[6423040], [51384320], [], [], [], []], "Ev Idx": 2311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657112797.452, "dur": 344.755,
    "args": {
      "External id": 2485513,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 6423040, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[51384320], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 2312, "In msg nelems": 51384320
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657112809.684, "dur": 326.753,
    "args": {
      "External id": 2485514,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[51384320]], "Ev Idx": 2313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113173.647, "dur": 2.359,
    "args": {
      "External id": 2485515,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 2314, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113232.874, "dur": 7.849,
    "args": {
      "External id": 2485516,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113285.388, "dur": 1.618,
    "args": {
      "External id": 2485517,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "256"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113303.343, "dur": 1.232,
    "args": {
      "External id": 2485518,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113317.004, "dur": 0.972,
    "args": {
      "External id": 2485519,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1048832"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113330.278, "dur": 2.736,
    "args": {
      "External id": 2485520,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1573120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113344.229, "dur": 0.690,
    "args": {
      "External id": 2485521,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113355.863, "dur": 0.844,
    "args": {
      "External id": 2485522,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "2097664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113368.605, "dur": 1.671,
    "args": {
      "External id": 2485523,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "3539456"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113382.264, "dur": 2.701,
    "args": {
      "External id": 2485524,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "4981248"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113475.322, "dur": 2712.816,
    "args": {
      "External id": 2485525,"Record function id": 0, "Ev Idx": 2324
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.22)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113493.153, "dur": 1013.883,
    "args": {
      "External id": 2485526,"Record function id": 0, "Ev Idx": 2325
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.22)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113507.242, "dur": 333.487,
    "args": {
      "External id": 2485527,"Record function id": 0, "Ev Idx": 2326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113604.281, "dur": 4.766,
    "args": {
      "External id": 2485528,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 2327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113612.371, "dur": 0.729,
    "args": {
      "External id": 2485529,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 2328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113614.841, "dur": 0.655,
    "args": {
      "External id": 2485530,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 2329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113617.313, "dur": 0.767,
    "args": {
      "External id": 2485531,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 2330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113621.176, "dur": 0.895,
    "args": {
      "External id": 2485532,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 2331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113623.501, "dur": 0.676,
    "args": {
      "External id": 2485533,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 2332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113625.803, "dur": 3.470,
    "args": {
      "External id": 2485534,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 2333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113630.674, "dur": 0.640,
    "args": {
      "External id": 2485535,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 2334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113634.492, "dur": 0.916,
    "args": {
      "External id": 2485536,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 2335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113636.865, "dur": 0.834,
    "args": {
      "External id": 2485537,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 2336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113655.451, "dur": 152.033,
    "args": {
      "External id": 2485538,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 2337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113670.532, "dur": 132.543,
    "args": {
      "External id": 2485539,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 2338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113691.704, "dur": 12.574,
    "args": {
      "External id": 2485540,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113707.138, "dur": 69.716,
    "args": {
      "External id": 2485541,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 2340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113712.248, "dur": 64.251,
    "args": {
      "External id": 2485542,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 2341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113716.605, "dur": 6.191,
    "args": {
      "External id": 2485543,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113724.625, "dur": 50.993,
    "args": {
      "External id": 2485544,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 2343
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.21", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113918.356, "dur": 581.268,
    "args": {
      "External id": 2485545,"Record function id": 0, "Ev Idx": 2344
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.21)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657113933.661, "dur": 553.881,
    "args": {
      "External id": 2485546,"Record function id": 0, "Ev Idx": 2345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114030.932, "dur": 9.737,
    "args": {
      "External id": 2485547,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114056.537, "dur": 31.714,
    "args": {
      "External id": 2485548,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 2347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114062.147, "dur": 1.723,
    "args": {
      "External id": 2485549,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114066.272, "dur": 2.625,
    "args": {
      "External id": 2485550,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114070.967, "dur": 0.403,
    "args": {
      "External id": 2485551,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114073.296, "dur": 0.312,
    "args": {
      "External id": 2485552,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114075.318, "dur": 0.521,
    "args": {
      "External id": 2485553,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114077.378, "dur": 0.357,
    "args": {
      "External id": 2485554,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114079.491, "dur": 0.344,
    "args": {
      "External id": 2485555,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114081.545, "dur": 0.575,
    "args": {
      "External id": 2485556,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114083.868, "dur": 0.355,
    "args": {
      "External id": 2485557,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114102.525, "dur": 33.910,
    "args": {
      "External id": 2485558,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 2357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114169.799, "dur": 99.440,
    "args": {
      "External id": 2485559,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 2358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114180.153, "dur": 4.600,
    "args": {
      "External id": 2485560,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114189.776, "dur": 10.500,
    "args": {
      "External id": 2485561,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 2360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114193.771, "dur": 6.067,
    "args": {
      "External id": 2485562,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 2361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114197.817, "dur": 0.652,
    "args": {
      "External id": 2485563,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 2362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114206.836, "dur": 25.537,
    "args": {
      "External id": 2485564,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 2363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114209.004, "dur": 0.289,
    "args": {
      "External id": 2485565,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114211.590, "dur": 0.465,
    "args": {
      "External id": 2485566,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114213.607, "dur": 0.466,
    "args": {
      "External id": 2485567,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114216.241, "dur": 0.556,
    "args": {
      "External id": 2485568,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114218.370, "dur": 0.358,
    "args": {
      "External id": 2485569,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114220.476, "dur": 0.308,
    "args": {
      "External id": 2485570,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114222.410, "dur": 2.231,
    "args": {
      "External id": 2485571,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114226.284, "dur": 0.298,
    "args": {
      "External id": 2485572,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114228.149, "dur": 0.320,
    "args": {
      "External id": 2485573,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114243.569, "dur": 18.239,
    "args": {
      "External id": 2485574,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 2373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114314.044, "dur": 109.076,
    "args": {
      "External id": 2485575,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 2374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114337.389, "dur": 81.997,
    "args": {
      "External id": 2485576,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 2375, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114346.631, "dur": 68.636,
    "args": {
      "External id": 2485577,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 2376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114438.031, "dur": 1.986,
    "args": {
      "External id": 2485578,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 2377, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114513.482, "dur": 1655.032,
    "args": {
      "External id": 2485579,"Sequence number": 24740466, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 2378
    }
  },
  {
    "ph": "f", "id": 182, "pid": 1336759, "tid": 1381173, "ts": 1514657114513.482,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114641.882, "dur": 102.706,
    "args": {
      "External id": 2485580,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [5632, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 2379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114780.591, "dur": 37.436,
    "args": {
      "External id": 2485581,"kernel_hash": "cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/m5/cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [65536, 5632], [16, 4096, 5632], [16, 4096, 5632], [16, 4096, 5632], []], "Ev Idx": 2380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114835.223, "dur": 47.701,
    "args": {
      "External id": 2485582,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 2048], [0, 5632, 1], [11534336, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632], [1, 2048, 5632]], "Ev Idx": 2381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114891.975, "dur": 32.249,
    "args": {
      "External id": 2485583,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 2382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657114931.757, "dur": 45.762,
    "args": {
      "External id": 2485584,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 2383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657115032.623, "dur": 34.647,
    "args": {
      "External id": 2485585,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 2384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657115075.643, "dur": 45.989,
    "args": {
      "External id": 2485586,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 2385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657115145.318, "dur": 22.905,
    "args": {
      "External id": 2485587,"kernel_hash": "c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/47/c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 2386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657115195.169, "dur": 29.505,
    "args": {
      "External id": 2485588,"kernel_hash": "cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/xr/cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657115244.499, "dur": 18.585,
    "args": {
      "External id": 2485589,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 2388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657115276.864, "dur": 16.078,
    "args": {
      "External id": 2485590,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 2389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657115305.750, "dur": 31.375,
    "args": {
      "External id": 2485591,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 2390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657115340.517, "dur": 31.892,
    "args": {
      "External id": 2485592,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 2391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657115401.824, "dur": 183.001,
    "args": {
      "External id": 2485593,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 2392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657115476.308, "dur": 5.380,
    "args": {
      "External id": 2485594,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657115483.967, "dur": 3.855,
    "args": {
      "External id": 2485595,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657115616.878, "dur": 25.016,
    "args": {
      "External id": 2485596,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657115653.386, "dur": 14.700,
    "args": {
      "External id": 2485597,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657115677.073, "dur": 38.343,
    "args": {
      "External id": 2485598,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 2397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657115722.000, "dur": 34.890,
    "args": {
      "External id": 2485599,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 2398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657115764.268, "dur": 20.275,
    "args": {
      "External id": 2485600,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 2399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657115792.087, "dur": 30.027,
    "args": {
      "External id": 2485601,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 2400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657115829.488, "dur": 19.835,
    "args": {
      "External id": 2485602,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 2401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657115856.470, "dur": 29.644,
    "args": {
      "External id": 2485603,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 2402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 1336759, "tid": 1381173,
    "ts": 1514657115903.510, "dur": 20.178,
    "args": {
      "External id": 2485604,"kernel_hash": "cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/g4/cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [65536, 2048], []], "Ev Idx": 2403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657115939.875, "dur": 38.412,
    "args": {
      "External id": 2485605,"kernel_hash": "c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/4n/c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116037.960, "dur": 19.645,
    "args": {
      "External id": 2485606,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 2405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116075.069, "dur": 18.641,
    "args": {
      "External id": 2485607,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 2406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116118.681, "dur": 17.622,
    "args": {
      "External id": 2485608,"kernel_hash": "cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u7/cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [8388608, 2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [16, 4096, 2048], []], "Ev Idx": 2407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116213.599, "dur": 15.499,
    "args": {
      "External id": 2485609,"Record function id": 0, "Ev Idx": 2408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116217.232, "dur": 10.914,
    "args": {
      "External id": 2485610,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116221.144, "dur": 5.921,
    "args": {
      "External id": 2485611,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116222.834, "dur": 4.097,
    "args": {
      "External id": 2485612,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116232.995, "dur": 5.099,
    "args": {
      "External id": 2485613,"Record function id": 0, "Ev Idx": 2412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116234.558, "dur": 3.091,
    "args": {
      "External id": 2485614,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116235.125, "dur": 2.066,
    "args": {
      "External id": 2485615,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116235.970, "dur": 1.115,
    "args": {
      "External id": 2485616,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116241.204, "dur": 4.508,
    "args": {
      "External id": 2485617,"Record function id": 0, "Ev Idx": 2416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116242.686, "dur": 2.609,
    "args": {
      "External id": 2485618,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116243.283, "dur": 1.609,
    "args": {
      "External id": 2485619,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116243.810, "dur": 1.017,
    "args": {
      "External id": 2485620,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116248.884, "dur": 5.675,
    "args": {
      "External id": 2485621,"Record function id": 0, "Ev Idx": 2420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116250.104, "dur": 4.041,
    "args": {
      "External id": 2485622,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 2421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116250.803, "dur": 2.957,
    "args": {
      "External id": 2485623,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 2422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116251.193, "dur": 2.498,
    "args": {
      "External id": 2485624,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 2423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116257.690, "dur": 3.845,
    "args": {
      "External id": 2485625,"Record function id": 0, "Ev Idx": 2424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116258.820, "dur": 2.300,
    "args": {
      "External id": 2485626,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116259.307, "dur": 1.335,
    "args": {
      "External id": 2485627,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116259.697, "dur": 0.878,
    "args": {
      "External id": 2485628,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116267.784, "dur": 4.031,
    "args": {
      "External id": 2485629,"Record function id": 0, "Ev Idx": 2428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116269.171, "dur": 2.206,
    "args": {
      "External id": 2485630,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116269.717, "dur": 1.256,
    "args": {
      "External id": 2485631,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116270.160, "dur": 0.720,
    "args": {
      "External id": 2485632,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116275.011, "dur": 3.335,
    "args": {
      "External id": 2485633,"Record function id": 0, "Ev Idx": 2432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116276.079, "dur": 1.857,
    "args": {
      "External id": 2485634,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116276.538, "dur": 1.012,
    "args": {
      "External id": 2485635,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116276.814, "dur": 0.659,
    "args": {
      "External id": 2485636,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116281.393, "dur": 3.765,
    "args": {
      "External id": 2485637,"Record function id": 0, "Ev Idx": 2436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116282.749, "dur": 1.979,
    "args": {
      "External id": 2485638,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116283.301, "dur": 1.030,
    "args": {
      "External id": 2485639,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116283.650, "dur": 0.585,
    "args": {
      "External id": 2485640,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116288.175, "dur": 3.656,
    "args": {
      "External id": 2485641,"Record function id": 0, "Ev Idx": 2440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116289.257, "dur": 2.163,
    "args": {
      "External id": 2485642,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116289.822, "dur": 1.133,
    "args": {
      "External id": 2485643,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116290.334, "dur": 0.509,
    "args": {
      "External id": 2485644,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116295.255, "dur": 36027.423,
    "args": {
      "External id": 2485645,"Record function id": 0, "Sequence number": 24740465, "Fwd thread id": 1, "Ev Idx": 2444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116296.695, "dur": 36017.273,
    "args": {
      "External id": 2485646,"Sequence number": 24740465, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 2445
    }
  },
  {
    "ph": "f", "id": 183, "pid": 1336759, "tid": 1381173, "ts": 1514657116296.695,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.22)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116326.401, "dur": 38.157,
    "args": {
      "External id": 2485647,"Record function id": 0, "Ev Idx": 2446
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.22)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116372.404, "dur": 60.077,
    "args": {
      "External id": 2485648,"Record function id": 0, "Ev Idx": 2447
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.22)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116438.548, "dur": 35867.707,
    "args": {
      "External id": 2485649,"Record function id": 0, "Ev Idx": 2448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116524.803, "dur": 23.579,
    "args": {
      "External id": 2485650,"Record function id": 0, "Concrete Inputs": ["[51384320]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116559.928, "dur": 5.028,
    "args": {
      "External id": 2485651,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 2450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116581.167, "dur": 34892.716,
    "args": {
      "External id": 2485652,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 2451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116595.020, "dur": 34869.867,
    "args": {
      "External id": 2485653,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 2452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116651.181, "dur": 15.348,
    "args": {
      "External id": 2485654,"Record function id": 0, "Concrete Inputs": ["[12602]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116676.336, "dur": 34747.864,
    "args": {
      "External id": 2485655,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], [], []], "Ev Idx": 2454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116678.985, "dur": 34744.457,
    "args": {
      "External id": 2485656,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], []], "Ev Idx": 2455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116683.359, "dur": 4.878,
    "args": {
      "External id": 2485657,"Record function id": 0, "Concrete Inputs": ["[12602]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657116690.068, "dur": 34729.464,
    "args": {
      "External id": 2485658,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[12602], [12602], []], "Ev Idx": 2457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657151582.318, "dur": 9.418,
    "args": {
      "External id": 2485659,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[51384320], [], [], [], [], []], "Ev Idx": 2458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657151585.597, "dur": 5.609,
    "args": {
      "External id": 2485660,"Record function id": 0, "Concrete Inputs": ["[6423040]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657151625.175, "dur": 344.830,
    "args": {
      "External id": 2485661,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[6423040], [51384320], [], [], [], []], "Ev Idx": 2460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657151652.204, "dur": 312.805,
    "args": {
      "External id": 2485662,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 6423040, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[51384320], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 2461, "In msg nelems": 51384320
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657151663.479, "dur": 296.438,
    "args": {
      "External id": 2485663,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[51384320]], "Ev Idx": 2462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152022.893, "dur": 3.167,
    "args": {
      "External id": 2485664,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 2463, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152087.924, "dur": 6.387,
    "args": {
      "External id": 2485665,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152141.017, "dur": 1.451,
    "args": {
      "External id": 2485666,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "256"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152159.543, "dur": 1.089,
    "args": {
      "External id": 2485667,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152176.266, "dur": 1.057,
    "args": {
      "External id": 2485668,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1048832"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152189.521, "dur": 1.089,
    "args": {
      "External id": 2485669,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1573120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152203.341, "dur": 1.019,
    "args": {
      "External id": 2485670,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152216.556, "dur": 0.970,
    "args": {
      "External id": 2485671,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "2097664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152231.554, "dur": 1.894,
    "args": {
      "External id": 2485672,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "3539456"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152244.857, "dur": 0.994,
    "args": {
      "External id": 2485673,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "4981248"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152338.085, "dur": 2729.435,
    "args": {
      "External id": 2485674,"Record function id": 0, "Ev Idx": 2473
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.21)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152355.690, "dur": 1043.787,
    "args": {
      "External id": 2485675,"Record function id": 0, "Ev Idx": 2474
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.21)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152371.621, "dur": 338.262,
    "args": {
      "External id": 2485676,"Record function id": 0, "Ev Idx": 2475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152443.040, "dur": 4.086,
    "args": {
      "External id": 2485677,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 2476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152450.258, "dur": 0.937,
    "args": {
      "External id": 2485678,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 2477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152452.985, "dur": 0.919,
    "args": {
      "External id": 2485679,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 2478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152455.380, "dur": 2.878,
    "args": {
      "External id": 2485680,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 2479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152459.834, "dur": 0.786,
    "args": {
      "External id": 2485681,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 2480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152462.421, "dur": 0.747,
    "args": {
      "External id": 2485682,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 2481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152464.620, "dur": 1.777,
    "args": {
      "External id": 2485683,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 2482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152467.805, "dur": 1.164,
    "args": {
      "External id": 2485684,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 2483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152470.273, "dur": 0.814,
    "args": {
      "External id": 2485685,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 2484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152472.363, "dur": 0.914,
    "args": {
      "External id": 2485686,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 2485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152491.796, "dur": 180.519,
    "args": {
      "External id": 2485687,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 2486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152519.572, "dur": 147.205,
    "args": {
      "External id": 2485688,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 2487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152554.183, "dur": 12.384,
    "args": {
      "External id": 2485689,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152569.997, "dur": 68.526,
    "args": {
      "External id": 2485690,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 2489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152572.609, "dur": 65.572,
    "args": {
      "External id": 2485691,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 2490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152576.961, "dur": 7.995,
    "args": {
      "External id": 2485692,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152586.686, "dur": 50.921,
    "args": {
      "External id": 2485693,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 2492
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.20", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152790.072, "dur": 601.283,
    "args": {
      "External id": 2485694,"Record function id": 0, "Ev Idx": 2493
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.20)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152807.782, "dur": 570.267,
    "args": {
      "External id": 2485695,"Record function id": 0, "Ev Idx": 2494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152862.632, "dur": 4.463,
    "args": {
      "External id": 2485696,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152882.921, "dur": 34.980,
    "args": {
      "External id": 2485697,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 2496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152888.020, "dur": 1.431,
    "args": {
      "External id": 2485698,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152891.829, "dur": 0.619,
    "args": {
      "External id": 2485699,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152894.355, "dur": 0.482,
    "args": {
      "External id": 2485700,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152896.693, "dur": 0.508,
    "args": {
      "External id": 2485701,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152902.339, "dur": 0.534,
    "args": {
      "External id": 2485702,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152904.401, "dur": 0.359,
    "args": {
      "External id": 2485703,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152907.010, "dur": 2.557,
    "args": {
      "External id": 2485704,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152911.377, "dur": 0.367,
    "args": {
      "External id": 2485705,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152913.559, "dur": 0.372,
    "args": {
      "External id": 2485706,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657152927.708, "dur": 32.394,
    "args": {
      "External id": 2485707,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 2506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1381173,
    "ts": 1514657153030.091, "dur": 112.650,
    "args": {
      "External id": 2485708,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 2507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657153041.819, "dur": 4.698,
    "args": {
      "External id": 2485709,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1381173,
    "ts": 1514657153052.212, "dur": 11.781,
    "args": {
      "External id": 2485710,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 2509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514657153056.599, "dur": 6.953,
    "args": {
      "External id": 2485711,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 2510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657153060.907, "dur": 1.159,
    "args": {
      "External id": 2485712,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 2511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657153071.311, "dur": 31.118,
    "args": {
      "External id": 2485713,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 2512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657153074.195, "dur": 0.453,
    "args": {
      "External id": 2485714,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657153077.033, "dur": 0.464,
    "args": {
      "External id": 2485715,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657153079.327, "dur": 0.399,
    "args": {
      "External id": 2485716,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657153081.912, "dur": 2.485,
    "args": {
      "External id": 2485717,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657153089.509, "dur": 0.489,
    "args": {
      "External id": 2485718,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657153091.539, "dur": 0.325,
    "args": {
      "External id": 2485719,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657153093.499, "dur": 0.381,
    "args": {
      "External id": 2485720,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657153095.385, "dur": 0.377,
    "args": {
      "External id": 2485721,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657153097.482, "dur": 0.446,
    "args": {
      "External id": 2485722,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657153113.656, "dur": 21.344,
    "args": {
      "External id": 2485723,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 2522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657153193.757, "dur": 119.257,
    "args": {
      "External id": 2485724,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 2523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657153220.359, "dur": 89.334,
    "args": {
      "External id": 2485725,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 2524, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657153230.717, "dur": 74.718,
    "args": {
      "External id": 2485726,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 2525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657153326.651, "dur": 1.769,
    "args": {
      "External id": 2485727,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 2526, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657153406.620, "dur": 1642.998,
    "args": {
      "External id": 2485728,"Sequence number": 24740464, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 2527
    }
  },
  {
    "ph": "f", "id": 184, "pid": 1336759, "tid": 1381173, "ts": 1514657153406.620,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657153507.317, "dur": 121.055,
    "args": {
      "External id": 2485729,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [5632, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 2528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657153671.058, "dur": 36.734,
    "args": {
      "External id": 2485730,"kernel_hash": "cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/m5/cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [65536, 5632], [16, 4096, 5632], [16, 4096, 5632], [16, 4096, 5632], []], "Ev Idx": 2529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657153724.645, "dur": 50.072,
    "args": {
      "External id": 2485731,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 2048], [0, 5632, 1], [11534336, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632], [1, 2048, 5632]], "Ev Idx": 2530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657153783.792, "dur": 32.599,
    "args": {
      "External id": 2485732,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 2531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657153825.697, "dur": 45.467,
    "args": {
      "External id": 2485733,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 2532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657153879.278, "dur": 28.882,
    "args": {
      "External id": 2485734,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 2533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657153914.999, "dur": 43.155,
    "args": {
      "External id": 2485735,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 2534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657153979.734, "dur": 62.089,
    "args": {
      "External id": 2485736,"kernel_hash": "c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/47/c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 2535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657154064.520, "dur": 28.666,
    "args": {
      "External id": 2485737,"kernel_hash": "cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/xr/cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2536
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657154113.403, "dur": 18.398,
    "args": {
      "External id": 2485738,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 2537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657154143.546, "dur": 14.420,
    "args": {
      "External id": 2485739,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 2538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657154168.551, "dur": 32.570,
    "args": {
      "External id": 2485740,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 2539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657154204.439, "dur": 33.080,
    "args": {
      "External id": 2485741,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 2540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657154266.285, "dur": 173.734,
    "args": {
      "External id": 2485742,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 2541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657154343.753, "dur": 5.667,
    "args": {
      "External id": 2485743,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657154351.704, "dur": 2.122,
    "args": {
      "External id": 2485744,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657154491.843, "dur": 23.622,
    "args": {
      "External id": 2485745,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657154543.324, "dur": 17.828,
    "args": {
      "External id": 2485746,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657154573.169, "dur": 39.683,
    "args": {
      "External id": 2485747,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 2546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657154619.838, "dur": 35.608,
    "args": {
      "External id": 2485748,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 2547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657154666.046, "dur": 20.709,
    "args": {
      "External id": 2485749,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 2548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657154692.667, "dur": 31.047,
    "args": {
      "External id": 2485750,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 2549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657154730.283, "dur": 19.313,
    "args": {
      "External id": 2485751,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 2550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657154757.399, "dur": 29.262,
    "args": {
      "External id": 2485752,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 2551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 1336759, "tid": 1381173,
    "ts": 1514657154804.151, "dur": 20.929,
    "args": {
      "External id": 2485753,"kernel_hash": "cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/g4/cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [65536, 2048], []], "Ev Idx": 2552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657154857.279, "dur": 27.363,
    "args": {
      "External id": 2485754,"kernel_hash": "c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/4n/c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657154900.078, "dur": 17.030,
    "args": {
      "External id": 2485755,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 2554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657154934.201, "dur": 13.934,
    "args": {
      "External id": 2485756,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 2555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 1336759, "tid": 1381173,
    "ts": 1514657154963.992, "dur": 52.524,
    "args": {
      "External id": 2485757,"kernel_hash": "cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u7/cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [8388608, 2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [16, 4096, 2048], []], "Ev Idx": 2556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155090.157, "dur": 41.000,
    "args": {
      "External id": 2485758,"Record function id": 0, "Ev Idx": 2557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155093.821, "dur": 36.302,
    "args": {
      "External id": 2485759,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155097.944, "dur": 31.123,
    "args": {
      "External id": 2485760,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155124.723, "dur": 4.141,
    "args": {
      "External id": 2485761,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155135.264, "dur": 4.990,
    "args": {
      "External id": 2485762,"Record function id": 0, "Ev Idx": 2561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155136.620, "dur": 3.103,
    "args": {
      "External id": 2485763,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155137.383, "dur": 1.863,
    "args": {
      "External id": 2485764,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155138.014, "dur": 1.013,
    "args": {
      "External id": 2485765,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155143.489, "dur": 5.847,
    "args": {
      "External id": 2485766,"Record function id": 0, "Ev Idx": 2565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155144.745, "dur": 4.178,
    "args": {
      "External id": 2485767,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155145.279, "dur": 3.221,
    "args": {
      "External id": 2485768,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155145.577, "dur": 2.846,
    "args": {
      "External id": 2485769,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155152.384, "dur": 3.751,
    "args": {
      "External id": 2485770,"Record function id": 0, "Ev Idx": 2569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155153.643, "dur": 2.073,
    "args": {
      "External id": 2485771,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 2570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155154.350, "dur": 0.983,
    "args": {
      "External id": 2485772,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 2571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155154.627, "dur": 0.613,
    "args": {
      "External id": 2485773,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 2572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155159.188, "dur": 3.514,
    "args": {
      "External id": 2485774,"Record function id": 0, "Ev Idx": 2573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155160.336, "dur": 1.979,
    "args": {
      "External id": 2485775,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155160.837, "dur": 1.080,
    "args": {
      "External id": 2485776,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155161.170, "dur": 0.681,
    "args": {
      "External id": 2485777,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155165.712, "dur": 3.414,
    "args": {
      "External id": 2485778,"Record function id": 0, "Ev Idx": 2577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155166.673, "dur": 2.022,
    "args": {
      "External id": 2485779,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155167.224, "dur": 1.084,
    "args": {
      "External id": 2485780,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155167.695, "dur": 0.549,
    "args": {
      "External id": 2485781,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155172.715, "dur": 3.393,
    "args": {
      "External id": 2485782,"Record function id": 0, "Ev Idx": 2581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155173.866, "dur": 1.859,
    "args": {
      "External id": 2485783,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155174.362, "dur": 0.973,
    "args": {
      "External id": 2485784,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155174.629, "dur": 0.642,
    "args": {
      "External id": 2485785,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155179.140, "dur": 5.254,
    "args": {
      "External id": 2485786,"Record function id": 0, "Ev Idx": 2585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155180.255, "dur": 3.746,
    "args": {
      "External id": 2485787,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155180.795, "dur": 2.780,
    "args": {
      "External id": 2485788,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155182.848, "dur": 0.626,
    "args": {
      "External id": 2485789,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155187.353, "dur": 3.267,
    "args": {
      "External id": 2485790,"Record function id": 0, "Ev Idx": 2589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155188.378, "dur": 1.858,
    "args": {
      "External id": 2485791,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155188.867, "dur": 0.961,
    "args": {
      "External id": 2485792,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155189.177, "dur": 0.554,
    "args": {
      "External id": 2485793,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155194.080, "dur": 38654.375,
    "args": {
      "External id": 2485794,"Record function id": 0, "Sequence number": 24740463, "Fwd thread id": 1, "Ev Idx": 2593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155195.460, "dur": 38644.719,
    "args": {
      "External id": 2485795,"Sequence number": 24740463, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 2594
    }
  },
  {
    "ph": "f", "id": 185, "pid": 1336759, "tid": 1381173, "ts": 1514657155195.460,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.21)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155224.774, "dur": 36.431,
    "args": {
      "External id": 2485796,"Record function id": 0, "Ev Idx": 2595
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.21)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155268.724, "dur": 65.960,
    "args": {
      "External id": 2485797,"Record function id": 0, "Ev Idx": 2596
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.21)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155340.647, "dur": 38491.282,
    "args": {
      "External id": 2485798,"Record function id": 0, "Ev Idx": 2597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155429.465, "dur": 6.566,
    "args": {
      "External id": 2485799,"Record function id": 0, "Concrete Inputs": ["[51384320]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155445.453, "dur": 6.673,
    "args": {
      "External id": 2485800,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 2599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155471.170, "dur": 37475.719,
    "args": {
      "External id": 2485801,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 2600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155484.120, "dur": 37453.924,
    "args": {
      "External id": 2485802,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 2601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155553.321, "dur": 17.757,
    "args": {
      "External id": 2485803,"Record function id": 0, "Concrete Inputs": ["[12602]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155577.842, "dur": 37320.825,
    "args": {
      "External id": 2485804,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], [], []], "Ev Idx": 2603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155580.657, "dur": 37317.424,
    "args": {
      "External id": 2485805,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], []], "Ev Idx": 2604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155585.084, "dur": 5.910,
    "args": {
      "External id": 2485806,"Record function id": 0, "Concrete Inputs": ["[12602]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657155593.004, "dur": 37301.843,
    "args": {
      "External id": 2485807,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[12602], [12602], []], "Ev Idx": 2606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657193061.126, "dur": 9.064,
    "args": {
      "External id": 2485808,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[51384320], [], [], [], [], []], "Ev Idx": 2607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657193064.604, "dur": 5.106,
    "args": {
      "External id": 2485809,"Record function id": 0, "Concrete Inputs": ["[6423040]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657193099.215, "dur": 404.861,
    "args": {
      "External id": 2485810,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[6423040], [51384320], [], [], [], []], "Ev Idx": 2609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657193125.912, "dur": 373.107,
    "args": {
      "External id": 2485811,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 6423040, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[51384320], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 2610, "In msg nelems": 51384320
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657193137.534, "dur": 355.364,
    "args": {
      "External id": 2485812,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[51384320]], "Ev Idx": 2611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657193525.821, "dur": 16.648,
    "args": {
      "External id": 2485813,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 2612, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657193606.365, "dur": 6.803,
    "args": {
      "External id": 2485814,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657193657.203, "dur": 1.704,
    "args": {
      "External id": 2485815,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "256"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657193675.725, "dur": 3.395,
    "args": {
      "External id": 2485816,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657193695.173, "dur": 1.355,
    "args": {
      "External id": 2485817,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1048832"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657193708.949, "dur": 1.016,
    "args": {
      "External id": 2485818,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1573120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657193721.365, "dur": 1.295,
    "args": {
      "External id": 2485819,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657193733.992, "dur": 3.085,
    "args": {
      "External id": 2485820,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "2097664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657193752.785, "dur": 2.859,
    "args": {
      "External id": 2485821,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "3539456"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657193766.579, "dur": 1.042,
    "args": {
      "External id": 2485822,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "4981248"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657193867.111, "dur": 2690.203,
    "args": {
      "External id": 2485823,"Record function id": 0, "Ev Idx": 2622
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.20)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657193885.790, "dur": 1017.850,
    "args": {
      "External id": 2485824,"Record function id": 0, "Ev Idx": 2623
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.20)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657193900.557, "dur": 363.984,
    "args": {
      "External id": 2485825,"Record function id": 0, "Ev Idx": 2624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657193973.886, "dur": 4.220,
    "args": {
      "External id": 2485826,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 2625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194018.034, "dur": 1.815,
    "args": {
      "External id": 2485827,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 2626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194023.380, "dur": 2.981,
    "args": {
      "External id": 2485828,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 2627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194028.163, "dur": 0.754,
    "args": {
      "External id": 2485829,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 2628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194030.688, "dur": 0.732,
    "args": {
      "External id": 2485830,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 2629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194033.009, "dur": 0.999,
    "args": {
      "External id": 2485831,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 2630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194035.676, "dur": 1.822,
    "args": {
      "External id": 2485832,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 2631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194039.134, "dur": 1.028,
    "args": {
      "External id": 2485833,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 2632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194041.820, "dur": 0.810,
    "args": {
      "External id": 2485834,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 2633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194044.031, "dur": 0.740,
    "args": {
      "External id": 2485835,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 2634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194066.524, "dur": 160.389,
    "args": {
      "External id": 2485836,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 2635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194082.982, "dur": 139.267,
    "args": {
      "External id": 2485837,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 2636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194108.008, "dur": 14.850,
    "args": {
      "External id": 2485838,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194125.945, "dur": 69.588,
    "args": {
      "External id": 2485839,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 2638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194128.828, "dur": 66.368,
    "args": {
      "External id": 2485840,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 2639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194133.296, "dur": 6.482,
    "args": {
      "External id": 2485841,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194141.880, "dur": 52.845,
    "args": {
      "External id": 2485842,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 2641
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.19", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194351.399, "dur": 544.338,
    "args": {
      "External id": 2485843,"Record function id": 0, "Ev Idx": 2642
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.19)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194367.699, "dur": 515.922,
    "args": {
      "External id": 2485844,"Record function id": 0, "Ev Idx": 2643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194421.673, "dur": 5.149,
    "args": {
      "External id": 2485845,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194442.396, "dur": 31.145,
    "args": {
      "External id": 2485846,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 2645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194447.469, "dur": 1.314,
    "args": {
      "External id": 2485847,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194451.046, "dur": 0.664,
    "args": {
      "External id": 2485848,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194453.342, "dur": 0.417,
    "args": {
      "External id": 2485849,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194455.402, "dur": 0.755,
    "args": {
      "External id": 2485850,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194458.031, "dur": 0.491,
    "args": {
      "External id": 2485851,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194460.313, "dur": 2.502,
    "args": {
      "External id": 2485852,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194464.662, "dur": 0.341,
    "args": {
      "External id": 2485853,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194466.685, "dur": 0.323,
    "args": {
      "External id": 2485854,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194468.679, "dur": 0.365,
    "args": {
      "External id": 2485855,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194483.064, "dur": 27.727,
    "args": {
      "External id": 2485856,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 2655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194559.021, "dur": 105.827,
    "args": {
      "External id": 2485857,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 2656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194570.479, "dur": 4.140,
    "args": {
      "External id": 2485858,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194580.136, "dur": 11.258,
    "args": {
      "External id": 2485859,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 2658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194584.516, "dur": 6.480,
    "args": {
      "External id": 2485860,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 2659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194588.669, "dur": 0.759,
    "args": {
      "External id": 2485861,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 2660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194598.943, "dur": 26.554,
    "args": {
      "External id": 2485862,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 2661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194602.023, "dur": 0.585,
    "args": {
      "External id": 2485863,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194604.508, "dur": 0.420,
    "args": {
      "External id": 2485864,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194606.595, "dur": 2.215,
    "args": {
      "External id": 2485865,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194610.331, "dur": 0.533,
    "args": {
      "External id": 2485866,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194612.585, "dur": 0.330,
    "args": {
      "External id": 2485867,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194614.494, "dur": 0.418,
    "args": {
      "External id": 2485868,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194616.763, "dur": 0.732,
    "args": {
      "External id": 2485869,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194619.331, "dur": 0.430,
    "args": {
      "External id": 2485870,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194621.622, "dur": 0.361,
    "args": {
      "External id": 2485871,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194636.934, "dur": 20.436,
    "args": {
      "External id": 2485872,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 2671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194710.983, "dur": 109.869,
    "args": {
      "External id": 2485873,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 2672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194733.351, "dur": 84.282,
    "args": {
      "External id": 2485874,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 2673, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194743.208, "dur": 69.930,
    "args": {
      "External id": 2485875,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 2674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194833.855, "dur": 2.122,
    "args": {
      "External id": 2485876,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 2675, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657194909.341, "dur": 1612.907,
    "args": {
      "External id": 2485877,"Sequence number": 24740462, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 2676
    }
  },
  {
    "ph": "f", "id": 186, "pid": 1336759, "tid": 1381173, "ts": 1514657194909.341,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657195052.768, "dur": 104.092,
    "args": {
      "External id": 2485878,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [5632, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 2677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657195193.517, "dur": 37.794,
    "args": {
      "External id": 2485879,"kernel_hash": "cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/m5/cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [65536, 5632], [16, 4096, 5632], [16, 4096, 5632], [16, 4096, 5632], []], "Ev Idx": 2678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657195248.012, "dur": 48.968,
    "args": {
      "External id": 2485880,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 2048], [0, 5632, 1], [11534336, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632], [1, 2048, 5632]], "Ev Idx": 2679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657195307.966, "dur": 35.350,
    "args": {
      "External id": 2485881,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 2680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657195350.745, "dur": 47.781,
    "args": {
      "External id": 2485882,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 2681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657195406.813, "dur": 26.963,
    "args": {
      "External id": 2485883,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 2682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657195441.198, "dur": 44.793,
    "args": {
      "External id": 2485884,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 2683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657195512.075, "dur": 38.240,
    "args": {
      "External id": 2485885,"kernel_hash": "c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/47/c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 2684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657195572.899, "dur": 30.143,
    "args": {
      "External id": 2485886,"kernel_hash": "cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/xr/cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657195623.096, "dur": 17.958,
    "args": {
      "External id": 2485887,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 2686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657195653.910, "dur": 14.522,
    "args": {
      "External id": 2485888,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 2687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657195678.341, "dur": 31.038,
    "args": {
      "External id": 2485889,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 2688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657195712.771, "dur": 32.701,
    "args": {
      "External id": 2485890,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 2689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657195774.643, "dur": 174.032,
    "args": {
      "External id": 2485891,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 2690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657195854.662, "dur": 6.340,
    "args": {
      "External id": 2485892,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657195863.465, "dur": 2.517,
    "args": {
      "External id": 2485893,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196020.068, "dur": 27.270,
    "args": {
      "External id": 2485894,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196060.332, "dur": 14.811,
    "args": {
      "External id": 2485895,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196085.808, "dur": 39.071,
    "args": {
      "External id": 2485896,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 2695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196131.893, "dur": 35.368,
    "args": {
      "External id": 2485897,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 2696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196177.092, "dur": 20.288,
    "args": {
      "External id": 2485898,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 2697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196202.873, "dur": 29.556,
    "args": {
      "External id": 2485899,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 2698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196239.709, "dur": 19.996,
    "args": {
      "External id": 2485900,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 2699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196275.288, "dur": 41.246,
    "args": {
      "External id": 2485901,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 2700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196340.692, "dur": 22.007,
    "args": {
      "External id": 2485902,"kernel_hash": "cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/g4/cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [65536, 2048], []], "Ev Idx": 2701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196379.762, "dur": 24.071,
    "args": {
      "External id": 2485903,"kernel_hash": "c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/4n/c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196418.133, "dur": 16.479,
    "args": {
      "External id": 2485904,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 2703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196452.529, "dur": 13.842,
    "args": {
      "External id": 2485905,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 2704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196478.618, "dur": 15.791,
    "args": {
      "External id": 2485906,"kernel_hash": "cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u7/cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [8388608, 2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [16, 4096, 2048], []], "Ev Idx": 2705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196580.111, "dur": 15.253,
    "args": {
      "External id": 2485907,"Record function id": 0, "Ev Idx": 2706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196583.511, "dur": 10.819,
    "args": {
      "External id": 2485908,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196587.532, "dur": 5.805,
    "args": {
      "External id": 2485909,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196589.127, "dur": 4.050,
    "args": {
      "External id": 2485910,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196599.217, "dur": 7.255,
    "args": {
      "External id": 2485911,"Record function id": 0, "Ev Idx": 2710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196600.641, "dur": 5.353,
    "args": {
      "External id": 2485912,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196601.491, "dur": 4.025,
    "args": {
      "External id": 2485913,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196602.206, "dur": 3.188,
    "args": {
      "External id": 2485914,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196609.710, "dur": 4.785,
    "args": {
      "External id": 2485915,"Record function id": 0, "Ev Idx": 2714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196611.247, "dur": 2.824,
    "args": {
      "External id": 2485916,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196611.906, "dur": 1.751,
    "args": {
      "External id": 2485917,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196612.442, "dur": 1.094,
    "args": {
      "External id": 2485918,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196617.797, "dur": 3.938,
    "args": {
      "External id": 2485919,"Record function id": 0, "Ev Idx": 2718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196618.962, "dur": 2.360,
    "args": {
      "External id": 2485920,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 2719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196619.599, "dur": 1.341,
    "args": {
      "External id": 2485921,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 2720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196620.157, "dur": 0.672,
    "args": {
      "External id": 2485922,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 2721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196624.754, "dur": 3.758,
    "args": {
      "External id": 2485923,"Record function id": 0, "Ev Idx": 2722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196625.942, "dur": 2.171,
    "args": {
      "External id": 2485924,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196626.595, "dur": 1.125,
    "args": {
      "External id": 2485925,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196626.935, "dur": 0.720,
    "args": {
      "External id": 2485926,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196631.932, "dur": 4.123,
    "args": {
      "External id": 2485927,"Record function id": 0, "Ev Idx": 2726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196633.192, "dur": 2.435,
    "args": {
      "External id": 2485928,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196633.650, "dur": 1.501,
    "args": {
      "External id": 2485929,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196634.267, "dur": 0.787,
    "args": {
      "External id": 2485930,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196639.417, "dur": 3.949,
    "args": {
      "External id": 2485931,"Record function id": 0, "Ev Idx": 2730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196640.751, "dur": 2.215,
    "args": {
      "External id": 2485932,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196641.457, "dur": 1.052,
    "args": {
      "External id": 2485933,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196641.839, "dur": 0.575,
    "args": {
      "External id": 2485934,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196646.537, "dur": 3.414,
    "args": {
      "External id": 2485935,"Record function id": 0, "Ev Idx": 2734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196647.763, "dur": 1.794,
    "args": {
      "External id": 2485936,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196648.187, "dur": 0.942,
    "args": {
      "External id": 2485937,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196648.487, "dur": 0.574,
    "args": {
      "External id": 2485938,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196653.099, "dur": 5.161,
    "args": {
      "External id": 2485939,"Record function id": 0, "Ev Idx": 2738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196654.168, "dur": 3.675,
    "args": {
      "External id": 2485940,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196654.772, "dur": 2.675,
    "args": {
      "External id": 2485941,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196655.148, "dur": 2.213,
    "args": {
      "External id": 2485942,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196662.035, "dur": 36434.193,
    "args": {
      "External id": 2485943,"Record function id": 0, "Sequence number": 24740461, "Fwd thread id": 1, "Ev Idx": 2742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196663.520, "dur": 36424.113,
    "args": {
      "External id": 2485944,"Sequence number": 24740461, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 2743
    }
  },
  {
    "ph": "f", "id": 187, "pid": 1336759, "tid": 1381173, "ts": 1514657196663.520,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.20)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196692.584, "dur": 37.140,
    "args": {
      "External id": 2485945,"Record function id": 0, "Ev Idx": 2744
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.20)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196737.358, "dur": 59.638,
    "args": {
      "External id": 2485946,"Record function id": 0, "Ev Idx": 2745
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.20)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196802.596, "dur": 36276.715,
    "args": {
      "External id": 2485947,"Record function id": 0, "Ev Idx": 2746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196887.349, "dur": 6.348,
    "args": {
      "External id": 2485948,"Record function id": 0, "Concrete Inputs": ["[51384320]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196902.649, "dur": 4.467,
    "args": {
      "External id": 2485949,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 2748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196922.401, "dur": 35300.473,
    "args": {
      "External id": 2485950,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 2749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657196935.966, "dur": 35277.466,
    "args": {
      "External id": 2485951,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 2750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657197054.108, "dur": 14.591,
    "args": {
      "External id": 2485952,"Record function id": 0, "Concrete Inputs": ["[12602]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657197075.471, "dur": 35099.503,
    "args": {
      "External id": 2485953,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], [], []], "Ev Idx": 2752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657197078.088, "dur": 35096.352,
    "args": {
      "External id": 2485954,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], []], "Ev Idx": 2753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657197082.176, "dur": 6.456,
    "args": {
      "External id": 2485955,"Record function id": 0, "Concrete Inputs": ["[12602]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657197090.555, "dur": 35080.607,
    "args": {
      "External id": 2485956,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[12602], [12602], []], "Ev Idx": 2755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657232313.335, "dur": 9.011,
    "args": {
      "External id": 2485957,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[51384320], [], [], [], [], []], "Ev Idx": 2756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657232316.795, "dur": 5.214,
    "args": {
      "External id": 2485958,"Record function id": 0, "Concrete Inputs": ["[6423040]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657232352.486, "dur": 376.486,
    "args": {
      "External id": 2485959,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[6423040], [51384320], [], [], [], []], "Ev Idx": 2758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657232382.722, "dur": 340.668,
    "args": {
      "External id": 2485960,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 6423040, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[51384320], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 2759, "In msg nelems": 51384320
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657232393.826, "dur": 323.151,
    "args": {
      "External id": 2485961,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[51384320]], "Ev Idx": 2760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657232750.228, "dur": 2.549,
    "args": {
      "External id": 2485962,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 2761, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657232812.731, "dur": 6.654,
    "args": {
      "External id": 2485963,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657232864.123, "dur": 3.722,
    "args": {
      "External id": 2485964,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "256"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657232884.299, "dur": 1.137,
    "args": {
      "External id": 2485965,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657232897.967, "dur": 1.286,
    "args": {
      "External id": 2485966,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1048832"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657232912.012, "dur": 0.885,
    "args": {
      "External id": 2485967,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1573120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657232924.171, "dur": 2.858,
    "args": {
      "External id": 2485968,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657232939.607, "dur": 0.848,
    "args": {
      "External id": 2485969,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "2097664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657232952.797, "dur": 1.908,
    "args": {
      "External id": 2485970,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "3539456"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657232966.094, "dur": 0.984,
    "args": {
      "External id": 2485971,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "4981248"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233113.457, "dur": 2719.518,
    "args": {
      "External id": 2485972,"Record function id": 0, "Ev Idx": 2771
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.19)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233132.985, "dur": 1045.110,
    "args": {
      "External id": 2485973,"Record function id": 0, "Ev Idx": 2772
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.19)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233150.129, "dur": 324.082,
    "args": {
      "External id": 2485974,"Record function id": 0, "Ev Idx": 2773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233228.450, "dur": 6.439,
    "args": {
      "External id": 2485975,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 2774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233238.274, "dur": 1.066,
    "args": {
      "External id": 2485976,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 2775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233241.194, "dur": 1.194,
    "args": {
      "External id": 2485977,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 2776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233244.233, "dur": 1.250,
    "args": {
      "External id": 2485978,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 2777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233247.238, "dur": 0.766,
    "args": {
      "External id": 2485979,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 2778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233257.943, "dur": 0.995,
    "args": {
      "External id": 2485980,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 2779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233260.799, "dur": 1.981,
    "args": {
      "External id": 2485981,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 2780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233264.740, "dur": 0.720,
    "args": {
      "External id": 2485982,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 2781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233266.845, "dur": 2.511,
    "args": {
      "External id": 2485983,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 2782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233270.687, "dur": 0.750,
    "args": {
      "External id": 2485984,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 2783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233289.715, "dur": 152.510,
    "args": {
      "External id": 2485985,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 2784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233305.455, "dur": 132.231,
    "args": {
      "External id": 2485986,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 2785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233323.889, "dur": 11.571,
    "args": {
      "External id": 2485987,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233338.789, "dur": 68.303,
    "args": {
      "External id": 2485988,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 2787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233341.517, "dur": 65.249,
    "args": {
      "External id": 2485989,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 2788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233345.553, "dur": 8.594,
    "args": {
      "External id": 2485990,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233356.182, "dur": 49.816,
    "args": {
      "External id": 2485991,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 2790
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.18", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233580.510, "dur": 589.473,
    "args": {
      "External id": 2485992,"Record function id": 0, "Ev Idx": 2791
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.18)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233598.141, "dur": 554.772,
    "args": {
      "External id": 2485993,"Record function id": 0, "Ev Idx": 2792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233658.604, "dur": 5.594,
    "args": {
      "External id": 2485994,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233680.117, "dur": 34.946,
    "args": {
      "External id": 2485995,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 2794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233688.838, "dur": 1.667,
    "args": {
      "External id": 2485996,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233692.757, "dur": 0.688,
    "args": {
      "External id": 2485997,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233695.054, "dur": 0.677,
    "args": {
      "External id": 2485998,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233697.349, "dur": 2.363,
    "args": {
      "External id": 2485999,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233701.410, "dur": 0.454,
    "args": {
      "External id": 2486000,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233703.456, "dur": 0.441,
    "args": {
      "External id": 2486001,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233705.466, "dur": 0.341,
    "args": {
      "External id": 2486002,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233707.360, "dur": 0.417,
    "args": {
      "External id": 2486003,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233709.611, "dur": 0.373,
    "args": {
      "External id": 2486004,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233725.430, "dur": 30.420,
    "args": {
      "External id": 2486005,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 2804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233787.711, "dur": 100.152,
    "args": {
      "External id": 2486006,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 2805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233798.309, "dur": 3.371,
    "args": {
      "External id": 2486007,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233806.519, "dur": 9.619,
    "args": {
      "External id": 2486008,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 2807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233810.676, "dur": 5.029,
    "args": {
      "External id": 2486009,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 2808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233814.262, "dur": 0.379,
    "args": {
      "External id": 2486010,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 2809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233823.539, "dur": 28.627,
    "args": {
      "External id": 2486011,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 2810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233825.889, "dur": 2.339,
    "args": {
      "External id": 2486012,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233830.084, "dur": 0.509,
    "args": {
      "External id": 2486013,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233832.467, "dur": 0.570,
    "args": {
      "External id": 2486014,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233835.041, "dur": 0.396,
    "args": {
      "External id": 2486015,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233837.537, "dur": 0.411,
    "args": {
      "External id": 2486016,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233839.536, "dur": 0.388,
    "args": {
      "External id": 2486017,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233842.128, "dur": 0.406,
    "args": {
      "External id": 2486018,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233843.960, "dur": 0.562,
    "args": {
      "External id": 2486019,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233846.117, "dur": 2.168,
    "args": {
      "External id": 2486020,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233861.854, "dur": 18.654,
    "args": {
      "External id": 2486021,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 2820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233930.453, "dur": 147.569,
    "args": {
      "External id": 2486022,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 2821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233951.625, "dur": 122.163,
    "args": {
      "External id": 2486023,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 2822, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657233960.383, "dur": 108.983,
    "args": {
      "External id": 2486024,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 2823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657234093.716, "dur": 1.950,
    "args": {
      "External id": 2486025,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 2824, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657234185.534, "dur": 1627.567,
    "args": {
      "External id": 2486026,"Sequence number": 24740460, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 2825
    }
  },
  {
    "ph": "f", "id": 188, "pid": 1336759, "tid": 1381173, "ts": 1514657234185.534,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657234289.021, "dur": 100.383,
    "args": {
      "External id": 2486027,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [5632, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 2826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657234424.671, "dur": 38.896,
    "args": {
      "External id": 2486028,"kernel_hash": "cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/m5/cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [65536, 5632], [16, 4096, 5632], [16, 4096, 5632], [16, 4096, 5632], []], "Ev Idx": 2827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657234483.316, "dur": 68.349,
    "args": {
      "External id": 2486029,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 2048], [0, 5632, 1], [11534336, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632], [1, 2048, 5632]], "Ev Idx": 2828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657234565.269, "dur": 37.044,
    "args": {
      "External id": 2486030,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 2829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657234610.582, "dur": 46.462,
    "args": {
      "External id": 2486031,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 2830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657234665.864, "dur": 28.294,
    "args": {
      "External id": 2486032,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 2831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657234703.865, "dur": 42.174,
    "args": {
      "External id": 2486033,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 2832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657234771.567, "dur": 24.461,
    "args": {
      "External id": 2486034,"kernel_hash": "c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/47/c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 2833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657234814.662, "dur": 28.566,
    "args": {
      "External id": 2486035,"kernel_hash": "cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/xr/cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657234862.128, "dur": 18.456,
    "args": {
      "External id": 2486036,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 2835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657234894.239, "dur": 14.663,
    "args": {
      "External id": 2486037,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 2836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657234922.334, "dur": 27.889,
    "args": {
      "External id": 2486038,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 2837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657234953.922, "dur": 71.343,
    "args": {
      "External id": 2486039,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 2838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235059.809, "dur": 172.684,
    "args": {
      "External id": 2486040,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 2839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235134.862, "dur": 5.958,
    "args": {
      "External id": 2486041,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235143.060, "dur": 2.581,
    "args": {
      "External id": 2486042,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235263.454, "dur": 27.449,
    "args": {
      "External id": 2486043,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235303.467, "dur": 15.206,
    "args": {
      "External id": 2486044,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235327.431, "dur": 42.192,
    "args": {
      "External id": 2486045,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 2844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235378.348, "dur": 39.521,
    "args": {
      "External id": 2486046,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 2845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235425.189, "dur": 20.163,
    "args": {
      "External id": 2486047,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 2846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235450.830, "dur": 43.910,
    "args": {
      "External id": 2486048,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 2847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235506.885, "dur": 42.373,
    "args": {
      "External id": 2486049,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 2848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235561.673, "dur": 38.195,
    "args": {
      "External id": 2486050,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 2849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235620.450, "dur": 23.690,
    "args": {
      "External id": 2486051,"kernel_hash": "cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/g4/cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [65536, 2048], []], "Ev Idx": 2850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235665.505, "dur": 22.475,
    "args": {
      "External id": 2486052,"kernel_hash": "c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/4n/c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235702.231, "dur": 17.348,
    "args": {
      "External id": 2486053,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 2852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235734.406, "dur": 16.863,
    "args": {
      "External id": 2486054,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 2853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235764.228, "dur": 16.108,
    "args": {
      "External id": 2486055,"kernel_hash": "cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u7/cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [8388608, 2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [16, 4096, 2048], []], "Ev Idx": 2854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235855.815, "dur": 17.293,
    "args": {
      "External id": 2486056,"Record function id": 0, "Ev Idx": 2855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235859.145, "dur": 13.092,
    "args": {
      "External id": 2486057,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235863.459, "dur": 7.804,
    "args": {
      "External id": 2486058,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235865.285, "dur": 5.820,
    "args": {
      "External id": 2486059,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235877.619, "dur": 5.427,
    "args": {
      "External id": 2486060,"Record function id": 0, "Ev Idx": 2859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235879.191, "dur": 3.360,
    "args": {
      "External id": 2486061,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235880.129, "dur": 1.963,
    "args": {
      "External id": 2486062,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235880.966, "dur": 0.962,
    "args": {
      "External id": 2486063,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235886.408, "dur": 4.347,
    "args": {
      "External id": 2486064,"Record function id": 0, "Ev Idx": 2863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235887.844, "dur": 2.500,
    "args": {
      "External id": 2486065,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235888.540, "dur": 1.364,
    "args": {
      "External id": 2486066,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235888.877, "dur": 0.940,
    "args": {
      "External id": 2486067,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 2866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235894.235, "dur": 3.902,
    "args": {
      "External id": 2486068,"Record function id": 0, "Ev Idx": 2867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235895.554, "dur": 2.184,
    "args": {
      "External id": 2486069,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 2868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235896.311, "dur": 1.043,
    "args": {
      "External id": 2486070,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 2869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235896.615, "dur": 0.645,
    "args": {
      "External id": 2486071,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 2870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235901.766, "dur": 3.795,
    "args": {
      "External id": 2486072,"Record function id": 0, "Ev Idx": 2871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235902.930, "dur": 2.200,
    "args": {
      "External id": 2486073,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235903.763, "dur": 0.963,
    "args": {
      "External id": 2486074,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235904.049, "dur": 0.610,
    "args": {
      "External id": 2486075,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235908.921, "dur": 4.535,
    "args": {
      "External id": 2486076,"Record function id": 0, "Ev Idx": 2875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235910.400, "dur": 2.625,
    "args": {
      "External id": 2486077,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235911.126, "dur": 1.457,
    "args": {
      "External id": 2486078,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235911.781, "dur": 0.698,
    "args": {
      "External id": 2486079,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235916.871, "dur": 4.359,
    "args": {
      "External id": 2486080,"Record function id": 0, "Ev Idx": 2879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235918.103, "dur": 2.697,
    "args": {
      "External id": 2486081,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235918.928, "dur": 1.463,
    "args": {
      "External id": 2486082,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235919.577, "dur": 0.698,
    "args": {
      "External id": 2486083,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235924.471, "dur": 5.657,
    "args": {
      "External id": 2486084,"Record function id": 0, "Ev Idx": 2883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235925.797, "dur": 3.921,
    "args": {
      "External id": 2486085,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235926.295, "dur": 3.010,
    "args": {
      "External id": 2486086,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235926.636, "dur": 2.602,
    "args": {
      "External id": 2486087,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 2886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235933.383, "dur": 3.758,
    "args": {
      "External id": 2486088,"Record function id": 0, "Ev Idx": 2887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235934.421, "dur": 2.262,
    "args": {
      "External id": 2486089,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235934.946, "dur": 1.308,
    "args": {
      "External id": 2486090,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235935.391, "dur": 0.777,
    "args": {
      "External id": 2486091,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 2890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235940.787, "dur": 35903.427,
    "args": {
      "External id": 2486092,"Record function id": 0, "Sequence number": 24740459, "Fwd thread id": 1, "Ev Idx": 2891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235942.177, "dur": 35892.736,
    "args": {
      "External id": 2486093,"Sequence number": 24740459, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 2892
    }
  },
  {
    "ph": "f", "id": 189, "pid": 1336759, "tid": 1381173, "ts": 1514657235942.177,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.19)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657235970.097, "dur": 72.491,
    "args": {
      "External id": 2486094,"Record function id": 0, "Ev Idx": 2893
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.19)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657236051.824, "dur": 61.009,
    "args": {
      "External id": 2486095,"Record function id": 0, "Ev Idx": 2894
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.19)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657236119.028, "dur": 35707.670,
    "args": {
      "External id": 2486096,"Record function id": 0, "Ev Idx": 2895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657236205.759, "dur": 6.715,
    "args": {
      "External id": 2486097,"Record function id": 0, "Concrete Inputs": ["[51384320]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657236222.919, "dur": 5.198,
    "args": {
      "External id": 2486098,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 2897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657236243.952, "dur": 34837.729,
    "args": {
      "External id": 2486099,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 2898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657236257.035, "dur": 34814.186,
    "args": {
      "External id": 2486100,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 2899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657236349.001, "dur": 14.199,
    "args": {
      "External id": 2486101,"Record function id": 0, "Concrete Inputs": ["[12602]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657236369.682, "dur": 34657.688,
    "args": {
      "External id": 2486102,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], [], []], "Ev Idx": 2901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657236372.373, "dur": 34654.181,
    "args": {
      "External id": 2486103,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], []], "Ev Idx": 2902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657236376.515, "dur": 5.808,
    "args": {
      "External id": 2486104,"Record function id": 0, "Concrete Inputs": ["[12602]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657236384.310, "dur": 34637.874,
    "args": {
      "External id": 2486105,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[12602], [12602], []], "Ev Idx": 2904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657271179.165, "dur": 10.769,
    "args": {
      "External id": 2486106,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[51384320], [], [], [], [], []], "Ev Idx": 2905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657271182.772, "dur": 6.832,
    "args": {
      "External id": 2486107,"Record function id": 0, "Concrete Inputs": ["[6423040]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657271219.997, "dur": 289.483,
    "args": {
      "External id": 2486108,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[6423040], [51384320], [], [], [], []], "Ev Idx": 2907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657271246.463, "dur": 258.498,
    "args": {
      "External id": 2486109,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 6423040, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[51384320], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 2908, "In msg nelems": 51384320
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657271257.770, "dur": 241.865,
    "args": {
      "External id": 2486110,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[51384320]], "Ev Idx": 2909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657271542.255, "dur": 3.432,
    "args": {
      "External id": 2486111,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 2910, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657271608.573, "dur": 6.477,
    "args": {
      "External id": 2486112,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657271665.222, "dur": 1.288,
    "args": {
      "External id": 2486113,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "256"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657271683.288, "dur": 1.060,
    "args": {
      "External id": 2486114,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657271697.228, "dur": 1.053,
    "args": {
      "External id": 2486115,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1048832"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657271709.065, "dur": 1.016,
    "args": {
      "External id": 2486116,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1573120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657271723.767, "dur": 1.241,
    "args": {
      "External id": 2486117,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657271737.327, "dur": 0.996,
    "args": {
      "External id": 2486118,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "2097664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657271750.712, "dur": 1.834,
    "args": {
      "External id": 2486119,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "3539456"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657271764.271, "dur": 0.828,
    "args": {
      "External id": 2486120,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "4981248"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657271860.773, "dur": 2743.477,
    "args": {
      "External id": 2486121,"Record function id": 0, "Ev Idx": 2920
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.18)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657271879.394, "dur": 1046.863,
    "args": {
      "External id": 2486122,"Record function id": 0, "Ev Idx": 2921
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.18)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657271894.132, "dur": 359.105,
    "args": {
      "External id": 2486123,"Record function id": 0, "Ev Idx": 2922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657271968.808, "dur": 3.985,
    "args": {
      "External id": 2486124,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 2923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657271976.139, "dur": 1.038,
    "args": {
      "External id": 2486125,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 2924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657271979.082, "dur": 1.085,
    "args": {
      "External id": 2486126,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 2925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272019.503, "dur": 1.742,
    "args": {
      "External id": 2486127,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 2926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272024.451, "dur": 0.959,
    "args": {
      "External id": 2486128,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 2927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272027.309, "dur": 0.966,
    "args": {
      "External id": 2486129,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 2928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272030.426, "dur": 2.019,
    "args": {
      "External id": 2486130,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 2929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272034.077, "dur": 2.969,
    "args": {
      "External id": 2486131,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 2930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272038.360, "dur": 0.963,
    "args": {
      "External id": 2486132,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 2931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272040.922, "dur": 0.859,
    "args": {
      "External id": 2486133,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 2932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272061.959, "dur": 156.428,
    "args": {
      "External id": 2486134,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 2933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272079.243, "dur": 134.676,
    "args": {
      "External id": 2486135,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 2934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272105.880, "dur": 12.459,
    "args": {
      "External id": 2486136,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272121.368, "dur": 65.620,
    "args": {
      "External id": 2486137,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 2936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272124.140, "dur": 62.492,
    "args": {
      "External id": 2486138,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 2937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272128.207, "dur": 5.825,
    "args": {
      "External id": 2486139,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272136.127, "dur": 49.920,
    "args": {
      "External id": 2486140,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 2939
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.17", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272337.543, "dur": 580.951,
    "args": {
      "External id": 2486141,"Record function id": 0, "Ev Idx": 2940
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.17)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272355.169, "dur": 549.940,
    "args": {
      "External id": 2486142,"Record function id": 0, "Ev Idx": 2941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272413.032, "dur": 4.633,
    "args": {
      "External id": 2486143,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272432.473, "dur": 36.552,
    "args": {
      "External id": 2486144,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 2943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272437.916, "dur": 1.715,
    "args": {
      "External id": 2486145,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272441.770, "dur": 0.502,
    "args": {
      "External id": 2486146,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272444.242, "dur": 2.792,
    "args": {
      "External id": 2486147,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272448.633, "dur": 0.470,
    "args": {
      "External id": 2486148,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272450.947, "dur": 0.346,
    "args": {
      "External id": 2486149,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272453.458, "dur": 0.414,
    "args": {
      "External id": 2486150,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272459.706, "dur": 0.444,
    "args": {
      "External id": 2486151,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272462.260, "dur": 0.343,
    "args": {
      "External id": 2486152,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272464.308, "dur": 0.437,
    "args": {
      "External id": 2486153,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272478.827, "dur": 31.320,
    "args": {
      "External id": 2486154,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 2953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272566.755, "dur": 111.982,
    "args": {
      "External id": 2486155,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 2954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272578.244, "dur": 4.996,
    "args": {
      "External id": 2486156,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272588.784, "dur": 12.830,
    "args": {
      "External id": 2486157,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 2956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272593.364, "dur": 7.826,
    "args": {
      "External id": 2486158,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 2957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272597.324, "dur": 2.389,
    "args": {
      "External id": 2486159,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 2958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272609.308, "dur": 25.139,
    "args": {
      "External id": 2486160,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 2959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272612.020, "dur": 0.380,
    "args": {
      "External id": 2486161,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272614.136, "dur": 0.399,
    "args": {
      "External id": 2486162,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272616.259, "dur": 0.350,
    "args": {
      "External id": 2486163,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272618.412, "dur": 0.449,
    "args": {
      "External id": 2486164,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272620.506, "dur": 0.334,
    "args": {
      "External id": 2486165,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272622.601, "dur": 0.530,
    "args": {
      "External id": 2486166,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272624.986, "dur": 0.509,
    "args": {
      "External id": 2486167,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272626.912, "dur": 2.423,
    "args": {
      "External id": 2486168,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272630.890, "dur": 0.385,
    "args": {
      "External id": 2486169,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 2968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272649.695, "dur": 21.426,
    "args": {
      "External id": 2486170,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 2969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272726.125, "dur": 112.680,
    "args": {
      "External id": 2486171,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 2970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272752.529, "dur": 82.725,
    "args": {
      "External id": 2486172,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 2971, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272762.361, "dur": 68.878,
    "args": {
      "External id": 2486173,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 2972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272855.084, "dur": 1.690,
    "args": {
      "External id": 2486174,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 2973, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657272932.955, "dur": 1651.992,
    "args": {
      "External id": 2486175,"Sequence number": 24740458, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 2974
    }
  },
  {
    "ph": "f", "id": 190, "pid": 1336759, "tid": 1381173, "ts": 1514657272932.955,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657273088.118, "dur": 102.978,
    "args": {
      "External id": 2486176,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [5632, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 2975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657273230.233, "dur": 39.140,
    "args": {
      "External id": 2486177,"kernel_hash": "cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/m5/cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [65536, 5632], [16, 4096, 5632], [16, 4096, 5632], [16, 4096, 5632], []], "Ev Idx": 2976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657273289.393, "dur": 48.362,
    "args": {
      "External id": 2486178,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 2048], [0, 5632, 1], [11534336, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632], [1, 2048, 5632]], "Ev Idx": 2977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657273347.904, "dur": 31.953,
    "args": {
      "External id": 2486179,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 2978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657273387.894, "dur": 45.755,
    "args": {
      "External id": 2486180,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 2979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657273442.177, "dur": 28.337,
    "args": {
      "External id": 2486181,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 2980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657273481.414, "dur": 43.424,
    "args": {
      "External id": 2486182,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 2981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657273568.563, "dur": 24.427,
    "args": {
      "External id": 2486183,"kernel_hash": "c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/47/c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 2982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657273611.749, "dur": 26.791,
    "args": {
      "External id": 2486184,"kernel_hash": "cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/xr/cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657273657.902, "dur": 17.994,
    "args": {
      "External id": 2486185,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 2984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657273688.650, "dur": 14.809,
    "args": {
      "External id": 2486186,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 2985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657273715.177, "dur": 31.510,
    "args": {
      "External id": 2486187,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 2986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657273750.219, "dur": 37.093,
    "args": {
      "External id": 2486188,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 2987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657273818.674, "dur": 206.414,
    "args": {
      "External id": 2486189,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 2988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657273893.390, "dur": 6.163,
    "args": {
      "External id": 2486190,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657273901.597, "dur": 2.423,
    "args": {
      "External id": 2486191,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 2990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274058.931, "dur": 25.628,
    "args": {
      "External id": 2486192,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274095.832, "dur": 14.877,
    "args": {
      "External id": 2486193,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 2992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274121.756, "dur": 39.564,
    "args": {
      "External id": 2486194,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 2993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274168.927, "dur": 35.331,
    "args": {
      "External id": 2486195,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 2994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274212.185, "dur": 20.211,
    "args": {
      "External id": 2486196,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 2995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274238.097, "dur": 30.311,
    "args": {
      "External id": 2486197,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 2996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274277.087, "dur": 34.616,
    "args": {
      "External id": 2486198,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 2997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274325.368, "dur": 35.418,
    "args": {
      "External id": 2486199,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 2998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274378.856, "dur": 21.491,
    "args": {
      "External id": 2486200,"kernel_hash": "cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/g4/cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [65536, 2048], []], "Ev Idx": 2999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274416.764, "dur": 22.763,
    "args": {
      "External id": 2486201,"kernel_hash": "c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/4n/c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274456.845, "dur": 16.887,
    "args": {
      "External id": 2486202,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 3001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274487.914, "dur": 16.770,
    "args": {
      "External id": 2486203,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 3002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274517.296, "dur": 30.823,
    "args": {
      "External id": 2486204,"kernel_hash": "cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u7/cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [8388608, 2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [16, 4096, 2048], []], "Ev Idx": 3003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274630.538, "dur": 14.952,
    "args": {
      "External id": 2486205,"Record function id": 0, "Ev Idx": 3004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274633.934, "dur": 10.643,
    "args": {
      "External id": 2486206,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274638.143, "dur": 5.515,
    "args": {
      "External id": 2486207,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274639.669, "dur": 3.851,
    "args": {
      "External id": 2486208,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274649.367, "dur": 4.903,
    "args": {
      "External id": 2486209,"Record function id": 0, "Ev Idx": 3008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274650.963, "dur": 2.908,
    "args": {
      "External id": 2486210,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274651.556, "dur": 1.857,
    "args": {
      "External id": 2486211,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274652.193, "dur": 1.105,
    "args": {
      "External id": 2486212,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274657.718, "dur": 4.111,
    "args": {
      "External id": 2486213,"Record function id": 0, "Ev Idx": 3012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274658.990, "dur": 2.433,
    "args": {
      "External id": 2486214,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274659.533, "dur": 1.443,
    "args": {
      "External id": 2486215,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274659.870, "dur": 1.026,
    "args": {
      "External id": 2486216,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274665.062, "dur": 4.313,
    "args": {
      "External id": 2486217,"Record function id": 0, "Ev Idx": 3016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274666.217, "dur": 2.750,
    "args": {
      "External id": 2486218,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 3017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274666.796, "dur": 1.749,
    "args": {
      "External id": 2486219,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 3018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274667.280, "dur": 1.202,
    "args": {
      "External id": 2486220,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 3019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274672.649, "dur": 6.852,
    "args": {
      "External id": 2486221,"Record function id": 0, "Ev Idx": 3020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274673.762, "dur": 5.351,
    "args": {
      "External id": 2486222,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274674.328, "dur": 4.377,
    "args": {
      "External id": 2486223,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274674.759, "dur": 3.883,
    "args": {
      "External id": 2486224,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274682.778, "dur": 3.726,
    "args": {
      "External id": 2486225,"Record function id": 0, "Ev Idx": 3024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274683.927, "dur": 2.176,
    "args": {
      "External id": 2486226,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274684.500, "dur": 1.177,
    "args": {
      "External id": 2486227,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274684.918, "dur": 0.653,
    "args": {
      "External id": 2486228,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274689.476, "dur": 5.979,
    "args": {
      "External id": 2486229,"Record function id": 0, "Ev Idx": 3028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274690.760, "dur": 4.306,
    "args": {
      "External id": 2486230,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274691.208, "dur": 3.439,
    "args": {
      "External id": 2486231,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274691.478, "dur": 3.082,
    "args": {
      "External id": 2486232,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274698.624, "dur": 3.427,
    "args": {
      "External id": 2486233,"Record function id": 0, "Ev Idx": 3032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274699.660, "dur": 2.007,
    "args": {
      "External id": 2486234,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274700.201, "dur": 1.065,
    "args": {
      "External id": 2486235,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274700.451, "dur": 0.729,
    "args": {
      "External id": 2486236,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274705.245, "dur": 3.491,
    "args": {
      "External id": 2486237,"Record function id": 0, "Ev Idx": 3036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274706.408, "dur": 1.930,
    "args": {
      "External id": 2486238,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274706.865, "dur": 1.081,
    "args": {
      "External id": 2486239,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274707.304, "dur": 0.550,
    "args": {
      "External id": 2486240,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274712.560, "dur": 36221.944,
    "args": {
      "External id": 2486241,"Record function id": 0, "Sequence number": 24740457, "Fwd thread id": 1, "Ev Idx": 3040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274714.360, "dur": 36211.771,
    "args": {
      "External id": 2486242,"Sequence number": 24740457, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 3041
    }
  },
  {
    "ph": "f", "id": 191, "pid": 1336759, "tid": 1381173, "ts": 1514657274714.360,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.18)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274745.512, "dur": 36.117,
    "args": {
      "External id": 2486243,"Record function id": 0, "Ev Idx": 3042
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.18)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274789.306, "dur": 59.266,
    "args": {
      "External id": 2486244,"Record function id": 0, "Ev Idx": 3043
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.18)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274854.679, "dur": 36064.279,
    "args": {
      "External id": 2486245,"Record function id": 0, "Ev Idx": 3044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274940.182, "dur": 6.379,
    "args": {
      "External id": 2486246,"Record function id": 0, "Concrete Inputs": ["[51384320]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274955.833, "dur": 4.361,
    "args": {
      "External id": 2486247,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 3046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657274975.071, "dur": 35079.054,
    "args": {
      "External id": 2486248,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 3047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657275061.904, "dur": 34982.290,
    "args": {
      "External id": 2486249,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 3048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657275113.986, "dur": 14.479,
    "args": {
      "External id": 2486250,"Record function id": 0, "Concrete Inputs": ["[12602]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657275135.044, "dur": 34842.439,
    "args": {
      "External id": 2486251,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], [], []], "Ev Idx": 3050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657275137.973, "dur": 34838.517,
    "args": {
      "External id": 2486252,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], []], "Ev Idx": 3051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657275142.298, "dur": 6.199,
    "args": {
      "External id": 2486253,"Record function id": 0, "Concrete Inputs": ["[12602]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657275150.487, "dur": 34822.682,
    "args": {
      "External id": 2486254,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[12602], [12602], []], "Ev Idx": 3053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657310145.672, "dur": 8.886,
    "args": {
      "External id": 2486255,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[51384320], [], [], [], [], []], "Ev Idx": 3054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657310149.096, "dur": 5.094,
    "args": {
      "External id": 2486256,"Record function id": 0, "Concrete Inputs": ["[6423040]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657310184.441, "dur": 425.425,
    "args": {
      "External id": 2486257,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[6423040], [51384320], [], [], [], []], "Ev Idx": 3056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657310211.401, "dur": 392.471,
    "args": {
      "External id": 2486258,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 6423040, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[51384320], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 3057, "In msg nelems": 51384320
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657310223.159, "dur": 374.221,
    "args": {
      "External id": 2486259,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[51384320]], "Ev Idx": 3058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657310632.172, "dur": 2.813,
    "args": {
      "External id": 2486260,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 3059, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657310698.102, "dur": 8.913,
    "args": {
      "External id": 2486261,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657310753.050, "dur": 2.449,
    "args": {
      "External id": 2486262,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "256"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657310774.234, "dur": 0.818,
    "args": {
      "External id": 2486263,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657310787.923, "dur": 1.001,
    "args": {
      "External id": 2486264,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1048832"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657310800.417, "dur": 2.794,
    "args": {
      "External id": 2486265,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1573120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657310815.154, "dur": 0.978,
    "args": {
      "External id": 2486266,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657310828.294, "dur": 1.022,
    "args": {
      "External id": 2486267,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "2097664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657310841.598, "dur": 1.731,
    "args": {
      "External id": 2486268,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "3539456"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657310855.388, "dur": 2.676,
    "args": {
      "External id": 2486269,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "4981248"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657310948.899, "dur": 2708.171,
    "args": {
      "External id": 2486270,"Record function id": 0, "Ev Idx": 3069
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.17)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657310966.761, "dur": 1084.898,
    "args": {
      "External id": 2486271,"Record function id": 0, "Ev Idx": 3070
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.17)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311016.460, "dur": 342.748,
    "args": {
      "External id": 2486272,"Record function id": 0, "Ev Idx": 3071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311127.557, "dur": 4.852,
    "args": {
      "External id": 2486273,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 3072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311135.988, "dur": 0.933,
    "args": {
      "External id": 2486274,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 3073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311138.946, "dur": 0.999,
    "args": {
      "External id": 2486275,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 3074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311141.547, "dur": 1.044,
    "args": {
      "External id": 2486276,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 3075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311144.018, "dur": 1.056,
    "args": {
      "External id": 2486277,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 3076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311148.289, "dur": 0.854,
    "args": {
      "External id": 2486278,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 3077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311150.795, "dur": 3.087,
    "args": {
      "External id": 2486279,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 3078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311155.222, "dur": 0.915,
    "args": {
      "External id": 2486280,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 3079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311157.560, "dur": 0.780,
    "args": {
      "External id": 2486281,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 3080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311159.970, "dur": 0.821,
    "args": {
      "External id": 2486282,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 3081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311179.813, "dur": 145.999,
    "args": {
      "External id": 2486283,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 3082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311196.685, "dur": 124.556,
    "args": {
      "External id": 2486284,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 3083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311213.766, "dur": 12.422,
    "args": {
      "External id": 2486285,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311229.682, "dur": 64.913,
    "args": {
      "External id": 2486286,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 3085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311232.341, "dur": 61.889,
    "args": {
      "External id": 2486287,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 3086
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311236.511, "dur": 5.665,
    "args": {
      "External id": 2486288,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311244.169, "dur": 49.377,
    "args": {
      "External id": 2486289,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 3088
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.16", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311436.020, "dur": 607.529,
    "args": {
      "External id": 2486290,"Record function id": 0, "Ev Idx": 3089
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.16)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311451.693, "dur": 577.801,
    "args": {
      "External id": 2486291,"Record function id": 0, "Ev Idx": 3090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311506.474, "dur": 4.212,
    "args": {
      "External id": 2486292,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311525.243, "dur": 46.671,
    "args": {
      "External id": 2486293,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 3092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311546.476, "dur": 1.814,
    "args": {
      "External id": 2486294,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311550.987, "dur": 2.138,
    "args": {
      "External id": 2486295,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311554.789, "dur": 0.451,
    "args": {
      "External id": 2486296,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311556.775, "dur": 0.410,
    "args": {
      "External id": 2486297,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311558.834, "dur": 0.404,
    "args": {
      "External id": 2486298,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311560.732, "dur": 0.432,
    "args": {
      "External id": 2486299,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311562.999, "dur": 0.414,
    "args": {
      "External id": 2486300,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311565.088, "dur": 0.346,
    "args": {
      "External id": 2486301,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311566.939, "dur": 0.400,
    "args": {
      "External id": 2486302,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311582.805, "dur": 31.852,
    "args": {
      "External id": 2486303,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 3102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311649.989, "dur": 115.208,
    "args": {
      "External id": 2486304,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 3103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311660.416, "dur": 5.472,
    "args": {
      "External id": 2486305,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311670.820, "dur": 10.058,
    "args": {
      "External id": 2486306,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 3105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311675.141, "dur": 5.333,
    "args": {
      "External id": 2486307,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 3106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311678.800, "dur": 0.579,
    "args": {
      "External id": 2486308,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 3107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311688.341, "dur": 28.698,
    "args": {
      "External id": 2486309,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 3108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311694.385, "dur": 0.522,
    "args": {
      "External id": 2486310,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311696.503, "dur": 0.389,
    "args": {
      "External id": 2486311,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311698.670, "dur": 0.426,
    "args": {
      "External id": 2486312,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311700.776, "dur": 0.444,
    "args": {
      "External id": 2486313,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311702.974, "dur": 0.358,
    "args": {
      "External id": 2486314,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311705.001, "dur": 0.392,
    "args": {
      "External id": 2486315,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311706.896, "dur": 2.263,
    "args": {
      "External id": 2486316,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311710.815, "dur": 0.462,
    "args": {
      "External id": 2486317,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311712.760, "dur": 0.568,
    "args": {
      "External id": 2486318,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311734.369, "dur": 23.160,
    "args": {
      "External id": 2486319,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 3118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311811.927, "dur": 112.563,
    "args": {
      "External id": 2486320,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 3119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311837.161, "dur": 83.908,
    "args": {
      "External id": 2486321,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 3120, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311847.213, "dur": 69.387,
    "args": {
      "External id": 2486322,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 3121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657311940.812, "dur": 2.290,
    "args": {
      "External id": 2486323,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 3122, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657312060.109, "dur": 1577.744,
    "args": {
      "External id": 2486324,"Sequence number": 24740456, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 3123
    }
  },
  {
    "ph": "f", "id": 192, "pid": 1336759, "tid": 1381173, "ts": 1514657312060.109,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657312162.725, "dur": 103.421,
    "args": {
      "External id": 2486325,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [5632, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 3124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657312301.972, "dur": 36.523,
    "args": {
      "External id": 2486326,"kernel_hash": "cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/m5/cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [65536, 5632], [16, 4096, 5632], [16, 4096, 5632], [16, 4096, 5632], []], "Ev Idx": 3125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657312357.114, "dur": 47.115,
    "args": {
      "External id": 2486327,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 2048], [0, 5632, 1], [11534336, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632], [1, 2048, 5632]], "Ev Idx": 3126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657312414.191, "dur": 32.059,
    "args": {
      "External id": 2486328,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 3127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657312453.157, "dur": 45.024,
    "args": {
      "External id": 2486329,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 3128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657312508.118, "dur": 44.955,
    "args": {
      "External id": 2486330,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 3129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657312563.782, "dur": 45.973,
    "args": {
      "External id": 2486331,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 3130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657312634.086, "dur": 21.035,
    "args": {
      "External id": 2486332,"kernel_hash": "c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/47/c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 3131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657312672.103, "dur": 27.367,
    "args": {
      "External id": 2486333,"kernel_hash": "cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/xr/cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657312718.461, "dur": 17.735,
    "args": {
      "External id": 2486334,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 3133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657312748.282, "dur": 14.089,
    "args": {
      "External id": 2486335,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 3134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657312774.034, "dur": 26.895,
    "args": {
      "External id": 2486336,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 3135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657312804.220, "dur": 32.194,
    "args": {
      "External id": 2486337,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 3136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657312862.894, "dur": 206.412,
    "args": {
      "External id": 2486338,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 3137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657312937.855, "dur": 5.790,
    "args": {
      "External id": 2486339,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657312945.745, "dur": 4.287,
    "args": {
      "External id": 2486340,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313103.950, "dur": 23.773,
    "args": {
      "External id": 2486341,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313155.216, "dur": 15.480,
    "args": {
      "External id": 2486342,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313179.968, "dur": 43.670,
    "args": {
      "External id": 2486343,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 3142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313231.021, "dur": 38.121,
    "args": {
      "External id": 2486344,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 3143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313276.061, "dur": 20.657,
    "args": {
      "External id": 2486345,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 3144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313304.307, "dur": 29.866,
    "args": {
      "External id": 2486346,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 3145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313340.781, "dur": 19.611,
    "args": {
      "External id": 2486347,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 3146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313368.117, "dur": 29.946,
    "args": {
      "External id": 2486348,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 3147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313425.688, "dur": 27.068,
    "args": {
      "External id": 2486349,"kernel_hash": "cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/g4/cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [65536, 2048], []], "Ev Idx": 3148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313471.118, "dur": 21.496,
    "args": {
      "External id": 2486350,"kernel_hash": "c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/4n/c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313505.854, "dur": 15.904,
    "args": {
      "External id": 2486351,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 3150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313558.018, "dur": 16.385,
    "args": {
      "External id": 2486352,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 3151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313590.138, "dur": 19.087,
    "args": {
      "External id": 2486353,"kernel_hash": "cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u7/cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [8388608, 2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [16, 4096, 2048], []], "Ev Idx": 3152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313683.806, "dur": 15.229,
    "args": {
      "External id": 2486354,"Record function id": 0, "Ev Idx": 3153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313687.057, "dur": 10.973,
    "args": {
      "External id": 2486355,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313691.118, "dur": 5.915,
    "args": {
      "External id": 2486356,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313692.959, "dur": 3.921,
    "args": {
      "External id": 2486357,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313703.006, "dur": 4.718,
    "args": {
      "External id": 2486358,"Record function id": 0, "Ev Idx": 3157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313704.337, "dur": 2.966,
    "args": {
      "External id": 2486359,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313704.875, "dur": 2.013,
    "args": {
      "External id": 2486360,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313705.732, "dur": 0.998,
    "args": {
      "External id": 2486361,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313711.018, "dur": 4.233,
    "args": {
      "External id": 2486362,"Record function id": 0, "Ev Idx": 3161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313712.293, "dur": 2.578,
    "args": {
      "External id": 2486363,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313712.972, "dur": 1.467,
    "args": {
      "External id": 2486364,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313713.250, "dur": 1.109,
    "args": {
      "External id": 2486365,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313718.385, "dur": 5.785,
    "args": {
      "External id": 2486366,"Record function id": 0, "Ev Idx": 3165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313719.601, "dur": 4.161,
    "args": {
      "External id": 2486367,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 3166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313720.166, "dur": 3.204,
    "args": {
      "External id": 2486368,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 3167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313720.501, "dur": 2.812,
    "args": {
      "External id": 2486369,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 3168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313727.268, "dur": 3.601,
    "args": {
      "External id": 2486370,"Record function id": 0, "Ev Idx": 3169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313728.345, "dur": 2.136,
    "args": {
      "External id": 2486371,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313728.824, "dur": 1.255,
    "args": {
      "External id": 2486372,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313729.279, "dur": 0.706,
    "args": {
      "External id": 2486373,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313734.033, "dur": 6.820,
    "args": {
      "External id": 2486374,"Record function id": 0, "Ev Idx": 3173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313735.141, "dur": 5.328,
    "args": {
      "External id": 2486375,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313735.668, "dur": 1.262,
    "args": {
      "External id": 2486376,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313736.270, "dur": 0.561,
    "args": {
      "External id": 2486377,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313744.095, "dur": 3.791,
    "args": {
      "External id": 2486378,"Record function id": 0, "Ev Idx": 3177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313745.268, "dur": 2.191,
    "args": {
      "External id": 2486379,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313745.734, "dur": 1.310,
    "args": {
      "External id": 2486380,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313746.069, "dur": 0.896,
    "args": {
      "External id": 2486381,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313751.129, "dur": 3.337,
    "args": {
      "External id": 2486382,"Record function id": 0, "Ev Idx": 3181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313752.252, "dur": 1.801,
    "args": {
      "External id": 2486383,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313752.750, "dur": 0.936,
    "args": {
      "External id": 2486384,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313753.012, "dur": 0.578,
    "args": {
      "External id": 2486385,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313757.783, "dur": 3.604,
    "args": {
      "External id": 2486386,"Record function id": 0, "Ev Idx": 3185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313758.845, "dur": 2.140,
    "args": {
      "External id": 2486387,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313759.294, "dur": 1.276,
    "args": {
      "External id": 2486388,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313759.788, "dur": 0.671,
    "args": {
      "External id": 2486389,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313764.926, "dur": 36178.978,
    "args": {
      "External id": 2486390,"Record function id": 0, "Sequence number": 24740455, "Fwd thread id": 1, "Ev Idx": 3189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313766.222, "dur": 36169.896,
    "args": {
      "External id": 2486391,"Sequence number": 24740455, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 3190
    }
  },
  {
    "ph": "f", "id": 193, "pid": 1336759, "tid": 1381173, "ts": 1514657313766.222,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.17)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313794.797, "dur": 38.343,
    "args": {
      "External id": 2486392,"Record function id": 0, "Ev Idx": 3191
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.17)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313841.458, "dur": 63.521,
    "args": {
      "External id": 2486393,"Record function id": 0, "Ev Idx": 3192
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.17)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657313910.892, "dur": 36017.474,
    "args": {
      "External id": 2486394,"Record function id": 0, "Ev Idx": 3193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657314038.942, "dur": 7.541,
    "args": {
      "External id": 2486395,"Record function id": 0, "Concrete Inputs": ["[51384320]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657314056.828, "dur": 5.122,
    "args": {
      "External id": 2486396,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 3195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657314077.346, "dur": 35036.205,
    "args": {
      "External id": 2486397,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 3196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657314090.861, "dur": 35013.631,
    "args": {
      "External id": 2486398,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 3197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657314145.920, "dur": 15.919,
    "args": {
      "External id": 2486399,"Record function id": 0, "Concrete Inputs": ["[12602]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657314168.079, "dur": 34898.581,
    "args": {
      "External id": 2486400,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], [], []], "Ev Idx": 3199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657314170.577, "dur": 34895.355,
    "args": {
      "External id": 2486401,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], []], "Ev Idx": 3200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657314175.014, "dur": 4.886,
    "args": {
      "External id": 2486402,"Record function id": 0, "Concrete Inputs": ["[12602]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657314181.818, "dur": 34880.862,
    "args": {
      "External id": 2486403,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[12602], [12602], []], "Ev Idx": 3202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657349204.239, "dur": 8.715,
    "args": {
      "External id": 2486404,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[51384320], [], [], [], [], []], "Ev Idx": 3203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657349207.412, "dur": 5.204,
    "args": {
      "External id": 2486405,"Record function id": 0, "Concrete Inputs": ["[6423040]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657349242.938, "dur": 377.281,
    "args": {
      "External id": 2486406,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[6423040], [51384320], [], [], [], []], "Ev Idx": 3205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657349269.987, "dur": 344.746,
    "args": {
      "External id": 2486407,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 6423040, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[51384320], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 3206, "In msg nelems": 51384320
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657349281.043, "dur": 327.596,
    "args": {
      "External id": 2486408,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[51384320]], "Ev Idx": 3207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657349642.154, "dur": 2.178,
    "args": {
      "External id": 2486409,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 3208, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657349705.041, "dur": 6.448,
    "args": {
      "External id": 2486410,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657349755.655, "dur": 1.421,
    "args": {
      "External id": 2486411,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "256"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657349774.708, "dur": 1.397,
    "args": {
      "External id": 2486412,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657349791.698, "dur": 1.067,
    "args": {
      "External id": 2486413,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1048832"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657349807.341, "dur": 1.305,
    "args": {
      "External id": 2486414,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1573120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657349819.949, "dur": 0.881,
    "args": {
      "External id": 2486415,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657349832.877, "dur": 1.066,
    "args": {
      "External id": 2486416,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "2097664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657349847.823, "dur": 2.190,
    "args": {
      "External id": 2486417,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "3539456"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657349864.695, "dur": 0.797,
    "args": {
      "External id": 2486418,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "4981248"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657349959.043, "dur": 2726.346,
    "args": {
      "External id": 2486419,"Record function id": 0, "Ev Idx": 3218
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.16)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657349977.506, "dur": 1062.225,
    "args": {
      "External id": 2486420,"Record function id": 0, "Ev Idx": 3219
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.16)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350034.516, "dur": 321.761,
    "args": {
      "External id": 2486421,"Record function id": 0, "Ev Idx": 3220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350114.976, "dur": 4.557,
    "args": {
      "External id": 2486422,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 3221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350123.116, "dur": 0.930,
    "args": {
      "External id": 2486423,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 3222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350125.870, "dur": 0.960,
    "args": {
      "External id": 2486424,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 3223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350128.454, "dur": 2.878,
    "args": {
      "External id": 2486425,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 3224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350132.556, "dur": 0.882,
    "args": {
      "External id": 2486426,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 3225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350135.093, "dur": 0.932,
    "args": {
      "External id": 2486427,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 3226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350137.496, "dur": 1.753,
    "args": {
      "External id": 2486428,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 3227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350140.649, "dur": 0.937,
    "args": {
      "External id": 2486429,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 3228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350143.033, "dur": 0.891,
    "args": {
      "External id": 2486430,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 3229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350145.486, "dur": 0.646,
    "args": {
      "External id": 2486431,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 3230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350164.989, "dur": 157.292,
    "args": {
      "External id": 2486432,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 3231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350180.445, "dur": 137.131,
    "args": {
      "External id": 2486433,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 3232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350207.668, "dur": 11.909,
    "args": {
      "External id": 2486434,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350222.944, "dur": 67.988,
    "args": {
      "External id": 2486435,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 3234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350225.705, "dur": 64.842,
    "args": {
      "External id": 2486436,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 3235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350230.037, "dur": 7.140,
    "args": {
      "External id": 2486437,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350239.203, "dur": 50.744,
    "args": {
      "External id": 2486438,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 3237
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.15", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350435.888, "dur": 593.852,
    "args": {
      "External id": 2486439,"Record function id": 0, "Ev Idx": 3238
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.15)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350454.105, "dur": 518.334,
    "args": {
      "External id": 2486440,"Record function id": 0, "Ev Idx": 3239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350509.254, "dur": 4.909,
    "args": {
      "External id": 2486441,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350547.306, "dur": 32.487,
    "args": {
      "External id": 2486442,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 3241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350553.227, "dur": 1.541,
    "args": {
      "External id": 2486443,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350557.208, "dur": 0.513,
    "args": {
      "External id": 2486444,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350559.626, "dur": 0.687,
    "args": {
      "External id": 2486445,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350562.261, "dur": 0.715,
    "args": {
      "External id": 2486446,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350564.789, "dur": 0.441,
    "args": {
      "External id": 2486447,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350567.007, "dur": 0.447,
    "args": {
      "External id": 2486448,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350569.079, "dur": 2.188,
    "args": {
      "External id": 2486449,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350572.760, "dur": 0.316,
    "args": {
      "External id": 2486450,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350574.690, "dur": 0.375,
    "args": {
      "External id": 2486451,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350590.133, "dur": 32.548,
    "args": {
      "External id": 2486452,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 3251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350656.018, "dur": 101.912,
    "args": {
      "External id": 2486453,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 3252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350667.443, "dur": 4.170,
    "args": {
      "External id": 2486454,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350676.528, "dur": 9.866,
    "args": {
      "External id": 2486455,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 3254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350680.578, "dur": 5.364,
    "args": {
      "External id": 2486456,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 3255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350684.345, "dur": 0.456,
    "args": {
      "External id": 2486457,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 3256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350694.016, "dur": 25.342,
    "args": {
      "External id": 2486458,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 3257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350696.479, "dur": 0.613,
    "args": {
      "External id": 2486459,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350698.747, "dur": 0.410,
    "args": {
      "External id": 2486460,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350701.272, "dur": 0.370,
    "args": {
      "External id": 2486461,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350703.430, "dur": 2.244,
    "args": {
      "External id": 2486462,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350707.496, "dur": 0.350,
    "args": {
      "External id": 2486463,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350709.428, "dur": 0.305,
    "args": {
      "External id": 2486464,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350711.387, "dur": 0.363,
    "args": {
      "External id": 2486465,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350713.453, "dur": 0.350,
    "args": {
      "External id": 2486466,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350715.273, "dur": 0.525,
    "args": {
      "External id": 2486467,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350731.734, "dur": 18.741,
    "args": {
      "External id": 2486468,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 3267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350803.704, "dur": 106.363,
    "args": {
      "External id": 2486469,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 3268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350826.106, "dur": 80.389,
    "args": {
      "External id": 2486470,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 3269, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350836.341, "dur": 65.958,
    "args": {
      "External id": 2486471,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 3270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657350923.896, "dur": 1.751,
    "args": {
      "External id": 2486472,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 3271, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657351049.110, "dur": 1617.101,
    "args": {
      "External id": 2486473,"Sequence number": 24740454, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 3272
    }
  },
  {
    "ph": "f", "id": 194, "pid": 1336759, "tid": 1381173, "ts": 1514657351049.110,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657351155.253, "dur": 103.237,
    "args": {
      "External id": 2486474,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [5632, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 3273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657351297.431, "dur": 37.558,
    "args": {
      "External id": 2486475,"kernel_hash": "cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/m5/cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [65536, 5632], [16, 4096, 5632], [16, 4096, 5632], [16, 4096, 5632], []], "Ev Idx": 3274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657351352.402, "dur": 49.034,
    "args": {
      "External id": 2486476,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 2048], [0, 5632, 1], [11534336, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632], [1, 2048, 5632]], "Ev Idx": 3275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657351411.713, "dur": 32.412,
    "args": {
      "External id": 2486477,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 3276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657351456.400, "dur": 45.395,
    "args": {
      "External id": 2486478,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 3277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657351546.971, "dur": 35.193,
    "args": {
      "External id": 2486479,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 3278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657351593.333, "dur": 42.522,
    "args": {
      "External id": 2486480,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 3279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657351660.847, "dur": 22.812,
    "args": {
      "External id": 2486481,"kernel_hash": "c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/47/c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 3280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657351703.507, "dur": 28.796,
    "args": {
      "External id": 2486482,"kernel_hash": "cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/xr/cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657351753.500, "dur": 18.667,
    "args": {
      "External id": 2486483,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 3282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657351786.984, "dur": 14.744,
    "args": {
      "External id": 2486484,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 3283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657351811.810, "dur": 28.198,
    "args": {
      "External id": 2486485,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 3284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657351843.518, "dur": 33.372,
    "args": {
      "External id": 2486486,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 3285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657351906.604, "dur": 208.783,
    "args": {
      "External id": 2486487,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 3286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352017.619, "dur": 6.791,
    "args": {
      "External id": 2486488,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352027.464, "dur": 2.111,
    "args": {
      "External id": 2486489,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352147.647, "dur": 24.126,
    "args": {
      "External id": 2486490,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352183.362, "dur": 14.499,
    "args": {
      "External id": 2486491,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352206.844, "dur": 40.687,
    "args": {
      "External id": 2486492,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 3291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352254.839, "dur": 35.337,
    "args": {
      "External id": 2486493,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 3292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352299.342, "dur": 20.599,
    "args": {
      "External id": 2486494,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 3293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352326.165, "dur": 29.948,
    "args": {
      "External id": 2486495,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 3294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352362.388, "dur": 19.483,
    "args": {
      "External id": 2486496,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 3295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352389.364, "dur": 29.236,
    "args": {
      "External id": 2486497,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 3296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352435.111, "dur": 33.230,
    "args": {
      "External id": 2486498,"kernel_hash": "cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/g4/cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [65536, 2048], []], "Ev Idx": 3297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352490.665, "dur": 24.901,
    "args": {
      "External id": 2486499,"kernel_hash": "c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/4n/c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352547.582, "dur": 19.755,
    "args": {
      "External id": 2486500,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 3299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352587.146, "dur": 18.904,
    "args": {
      "External id": 2486501,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 3300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352619.269, "dur": 15.633,
    "args": {
      "External id": 2486502,"kernel_hash": "cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u7/cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [8388608, 2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [16, 4096, 2048], []], "Ev Idx": 3301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352707.314, "dur": 18.682,
    "args": {
      "External id": 2486503,"Record function id": 0, "Ev Idx": 3302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352710.590, "dur": 14.379,
    "args": {
      "External id": 2486504,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352714.596, "dur": 9.451,
    "args": {
      "External id": 2486505,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352716.512, "dur": 7.380,
    "args": {
      "External id": 2486506,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352730.354, "dur": 5.322,
    "args": {
      "External id": 2486507,"Record function id": 0, "Ev Idx": 3306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352731.914, "dur": 3.315,
    "args": {
      "External id": 2486508,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352732.712, "dur": 2.077,
    "args": {
      "External id": 2486509,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352733.692, "dur": 0.934,
    "args": {
      "External id": 2486510,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352738.963, "dur": 5.963,
    "args": {
      "External id": 2486511,"Record function id": 0, "Ev Idx": 3310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352740.107, "dur": 4.387,
    "args": {
      "External id": 2486512,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352740.761, "dur": 3.320,
    "args": {
      "External id": 2486513,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352741.253, "dur": 2.746,
    "args": {
      "External id": 2486514,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352748.254, "dur": 4.906,
    "args": {
      "External id": 2486515,"Record function id": 0, "Ev Idx": 3314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352749.699, "dur": 3.060,
    "args": {
      "External id": 2486516,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 3315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352750.689, "dur": 1.680,
    "args": {
      "External id": 2486517,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 3316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352751.403, "dur": 0.869,
    "args": {
      "External id": 2486518,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 3317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352756.585, "dur": 4.035,
    "args": {
      "External id": 2486519,"Record function id": 0, "Ev Idx": 3318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352757.858, "dur": 2.359,
    "args": {
      "External id": 2486520,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352758.505, "dur": 1.310,
    "args": {
      "External id": 2486521,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352759.021, "dur": 0.695,
    "args": {
      "External id": 2486522,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352763.794, "dur": 3.949,
    "args": {
      "External id": 2486523,"Record function id": 0, "Ev Idx": 3322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352764.853, "dur": 2.435,
    "args": {
      "External id": 2486524,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352765.304, "dur": 1.616,
    "args": {
      "External id": 2486525,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352766.009, "dur": 0.804,
    "args": {
      "External id": 2486526,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352771.035, "dur": 3.760,
    "args": {
      "External id": 2486527,"Record function id": 0, "Ev Idx": 3326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352772.163, "dur": 2.243,
    "args": {
      "External id": 2486528,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352772.690, "dur": 1.328,
    "args": {
      "External id": 2486529,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352773.288, "dur": 0.644,
    "args": {
      "External id": 2486530,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352777.930, "dur": 3.819,
    "args": {
      "External id": 2486531,"Record function id": 0, "Ev Idx": 3330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352779.159, "dur": 2.175,
    "args": {
      "External id": 2486532,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352779.832, "dur": 1.106,
    "args": {
      "External id": 2486533,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352780.245, "dur": 0.627,
    "args": {
      "External id": 2486534,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352784.728, "dur": 3.457,
    "args": {
      "External id": 2486535,"Record function id": 0, "Ev Idx": 3334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352785.922, "dur": 1.867,
    "args": {
      "External id": 2486536,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352786.364, "dur": 1.062,
    "args": {
      "External id": 2486537,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352786.762, "dur": 0.570,
    "args": {
      "External id": 2486538,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352791.717, "dur": 37483.558,
    "args": {
      "External id": 2486539,"Record function id": 0, "Sequence number": 24740453, "Fwd thread id": 1, "Ev Idx": 3338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352793.018, "dur": 37473.989,
    "args": {
      "External id": 2486540,"Sequence number": 24740453, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 3339
    }
  },
  {
    "ph": "f", "id": 195, "pid": 1336759, "tid": 1381173, "ts": 1514657352793.018,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.16)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352821.379, "dur": 37.387,
    "args": {
      "External id": 2486541,"Record function id": 0, "Ev Idx": 3340
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.16)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352866.762, "dur": 62.484,
    "args": {
      "External id": 2486542,"Record function id": 0, "Ev Idx": 3341
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.16)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657352938.206, "dur": 37320.783,
    "args": {
      "External id": 2486543,"Record function id": 0, "Ev Idx": 3342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657353063.108, "dur": 7.783,
    "args": {
      "External id": 2486544,"Record function id": 0, "Concrete Inputs": ["[51384320]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657353081.365, "dur": 7.109,
    "args": {
      "External id": 2486545,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 3344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657353104.234, "dur": 36229.928,
    "args": {
      "External id": 2486546,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 3345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657353121.652, "dur": 36202.190,
    "args": {
      "External id": 2486547,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 3346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657353189.176, "dur": 13.804,
    "args": {
      "External id": 2486548,"Record function id": 0, "Concrete Inputs": ["[12602]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657353209.280, "dur": 36074.603,
    "args": {
      "External id": 2486549,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], [], []], "Ev Idx": 3348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657353211.929, "dur": 36071.226,
    "args": {
      "External id": 2486550,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], []], "Ev Idx": 3349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657353216.216, "dur": 5.026,
    "args": {
      "External id": 2486551,"Record function id": 0, "Concrete Inputs": ["[12602]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657353223.361, "dur": 36055.410,
    "args": {
      "External id": 2486552,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[12602], [12602], []], "Ev Idx": 3351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657389429.372, "dur": 10.374,
    "args": {
      "External id": 2486553,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[51384320], [], [], [], [], []], "Ev Idx": 3352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657389433.148, "dur": 6.224,
    "args": {
      "External id": 2486554,"Record function id": 0, "Concrete Inputs": ["[6423040]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657389470.309, "dur": 438.465,
    "args": {
      "External id": 2486555,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[6423040], [51384320], [], [], [], []], "Ev Idx": 3354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657389497.647, "dur": 405.790,
    "args": {
      "External id": 2486556,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 6423040, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[51384320], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 3355, "In msg nelems": 51384320
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657389509.254, "dur": 388.017,
    "args": {
      "External id": 2486557,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[51384320]], "Ev Idx": 3356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657389940.992, "dur": 2.315,
    "args": {
      "External id": 2486558,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 3357, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390039.614, "dur": 7.328,
    "args": {
      "External id": 2486559,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390093.098, "dur": 1.603,
    "args": {
      "External id": 2486560,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "256"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390112.388, "dur": 3.276,
    "args": {
      "External id": 2486561,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390128.850, "dur": 0.913,
    "args": {
      "External id": 2486562,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1048832"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390141.637, "dur": 1.016,
    "args": {
      "External id": 2486563,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1573120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390153.845, "dur": 0.961,
    "args": {
      "External id": 2486564,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390166.917, "dur": 2.650,
    "args": {
      "External id": 2486565,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "2097664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390182.111, "dur": 1.894,
    "args": {
      "External id": 2486566,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "3539456"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390195.217, "dur": 0.787,
    "args": {
      "External id": 2486567,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "4981248"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390291.918, "dur": 2670.230,
    "args": {
      "External id": 2486568,"Record function id": 0, "Ev Idx": 3367
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.15)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390310.298, "dur": 1020.103,
    "args": {
      "External id": 2486569,"Record function id": 0, "Ev Idx": 3368
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.15)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390327.356, "dur": 341.715,
    "args": {
      "External id": 2486570,"Record function id": 0, "Ev Idx": 3369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390408.099, "dur": 4.210,
    "args": {
      "External id": 2486571,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 3370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390415.535, "dur": 1.085,
    "args": {
      "External id": 2486572,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 3371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390418.925, "dur": 2.827,
    "args": {
      "External id": 2486573,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 3372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390423.764, "dur": 0.684,
    "args": {
      "External id": 2486574,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 3373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390426.122, "dur": 0.739,
    "args": {
      "External id": 2486575,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 3374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390428.617, "dur": 1.009,
    "args": {
      "External id": 2486576,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 3375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390431.360, "dur": 1.860,
    "args": {
      "External id": 2486577,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 3376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390434.651, "dur": 1.001,
    "args": {
      "External id": 2486578,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 3377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390437.219, "dur": 1.056,
    "args": {
      "External id": 2486579,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 3378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390439.614, "dur": 0.978,
    "args": {
      "External id": 2486580,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 3379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390458.348, "dur": 173.808,
    "args": {
      "External id": 2486581,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 3380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390474.488, "dur": 152.462,
    "args": {
      "External id": 2486582,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 3381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390497.514, "dur": 14.248,
    "args": {
      "External id": 2486583,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390515.112, "dur": 84.290,
    "args": {
      "External id": 2486584,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 3383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390518.029, "dur": 81.002,
    "args": {
      "External id": 2486585,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 3384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390522.202, "dur": 20.425,
    "args": {
      "External id": 2486586,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390545.959, "dur": 52.197,
    "args": {
      "External id": 2486587,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 3386
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.14", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390750.296, "dur": 572.269,
    "args": {
      "External id": 2486588,"Record function id": 0, "Ev Idx": 3387
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.14)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390767.154, "dur": 542.822,
    "args": {
      "External id": 2486589,"Record function id": 0, "Ev Idx": 3388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390824.038, "dur": 5.534,
    "args": {
      "External id": 2486590,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390844.675, "dur": 30.898,
    "args": {
      "External id": 2486591,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 3390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390849.585, "dur": 1.567,
    "args": {
      "External id": 2486592,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390853.667, "dur": 0.572,
    "args": {
      "External id": 2486593,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390855.853, "dur": 0.485,
    "args": {
      "External id": 2486594,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390858.215, "dur": 0.456,
    "args": {
      "External id": 2486595,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390860.344, "dur": 0.440,
    "args": {
      "External id": 2486596,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390862.372, "dur": 2.479,
    "args": {
      "External id": 2486597,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390866.598, "dur": 0.367,
    "args": {
      "External id": 2486598,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390868.664, "dur": 0.411,
    "args": {
      "External id": 2486599,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390870.748, "dur": 0.410,
    "args": {
      "External id": 2486600,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390886.046, "dur": 28.343,
    "args": {
      "External id": 2486601,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 3400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390945.100, "dur": 143.753,
    "args": {
      "External id": 2486602,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 3401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390955.014, "dur": 3.084,
    "args": {
      "External id": 2486603,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390962.880, "dur": 10.548,
    "args": {
      "External id": 2486604,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 3403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390967.310, "dur": 5.703,
    "args": {
      "External id": 2486605,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 3404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390971.269, "dur": 0.536,
    "args": {
      "External id": 2486606,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 3405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657390979.968, "dur": 66.333,
    "args": {
      "External id": 2486607,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 3406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657391020.712, "dur": 0.714,
    "args": {
      "External id": 2486608,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657391023.923, "dur": 0.395,
    "args": {
      "External id": 2486609,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657391026.196, "dur": 2.444,
    "args": {
      "External id": 2486610,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657391030.854, "dur": 0.421,
    "args": {
      "External id": 2486611,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657391032.739, "dur": 0.590,
    "args": {
      "External id": 2486612,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657391035.353, "dur": 0.482,
    "args": {
      "External id": 2486613,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657391037.569, "dur": 0.396,
    "args": {
      "External id": 2486614,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657391039.794, "dur": 0.449,
    "args": {
      "External id": 2486615,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657391041.834, "dur": 0.396,
    "args": {
      "External id": 2486616,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657391058.857, "dur": 21.596,
    "args": {
      "External id": 2486617,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 3416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657391137.133, "dur": 109.034,
    "args": {
      "External id": 2486618,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 3417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657391160.136, "dur": 82.583,
    "args": {
      "External id": 2486619,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 3418, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657391170.142, "dur": 68.152,
    "args": {
      "External id": 2486620,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 3419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657391259.827, "dur": 1.748,
    "args": {
      "External id": 2486621,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 3420, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657391336.935, "dur": 1603.524,
    "args": {
      "External id": 2486622,"Sequence number": 24740452, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 3421
    }
  },
  {
    "ph": "f", "id": 196, "pid": 1336759, "tid": 1381173, "ts": 1514657391336.935,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657391440.802, "dur": 120.265,
    "args": {
      "External id": 2486623,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [5632, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 3422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657391599.809, "dur": 38.139,
    "args": {
      "External id": 2486624,"kernel_hash": "cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/m5/cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [65536, 5632], [16, 4096, 5632], [16, 4096, 5632], [16, 4096, 5632], []], "Ev Idx": 3423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657391656.016, "dur": 53.719,
    "args": {
      "External id": 2486625,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 2048], [0, 5632, 1], [11534336, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632], [1, 2048, 5632]], "Ev Idx": 3424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657391721.581, "dur": 33.262,
    "args": {
      "External id": 2486626,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 3425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657391762.713, "dur": 44.670,
    "args": {
      "External id": 2486627,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 3426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657391815.333, "dur": 27.346,
    "args": {
      "External id": 2486628,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 3427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657391850.888, "dur": 42.164,
    "args": {
      "External id": 2486629,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 3428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657391917.195, "dur": 21.890,
    "args": {
      "External id": 2486630,"kernel_hash": "c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/47/c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 3429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657391962.144, "dur": 71.483,
    "args": {
      "External id": 2486631,"kernel_hash": "cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/xr/cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657392058.091, "dur": 19.861,
    "args": {
      "External id": 2486632,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 3431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657392091.521, "dur": 13.452,
    "args": {
      "External id": 2486633,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 3432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657392112.933, "dur": 31.381,
    "args": {
      "External id": 2486634,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 3433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657392147.684, "dur": 32.467,
    "args": {
      "External id": 2486635,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 3434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657392225.093, "dur": 164.617,
    "args": {
      "External id": 2486636,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 3435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657392298.642, "dur": 5.637,
    "args": {
      "External id": 2486637,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657392306.748, "dur": 2.306,
    "args": {
      "External id": 2486638,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657392419.788, "dur": 21.518,
    "args": {
      "External id": 2486639,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657392452.367, "dur": 14.307,
    "args": {
      "External id": 2486640,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657392475.578, "dur": 32.360,
    "args": {
      "External id": 2486641,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 3440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657392515.110, "dur": 56.866,
    "args": {
      "External id": 2486642,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 3441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657392584.341, "dur": 22.627,
    "args": {
      "External id": 2486643,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 3442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657392613.048, "dur": 33.436,
    "args": {
      "External id": 2486644,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 3443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657392653.336, "dur": 19.729,
    "args": {
      "External id": 2486645,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 3444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657392688.912, "dur": 41.109,
    "args": {
      "External id": 2486646,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 3445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 1336759, "tid": 1381173,
    "ts": 1514657392754.800, "dur": 22.867,
    "args": {
      "External id": 2486647,"kernel_hash": "cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/g4/cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [65536, 2048], []], "Ev Idx": 3446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657392794.903, "dur": 22.181,
    "args": {
      "External id": 2486648,"kernel_hash": "c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/4n/c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657392831.233, "dur": 20.634,
    "args": {
      "External id": 2486649,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 3448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657392865.743, "dur": 14.419,
    "args": {
      "External id": 2486650,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 3449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 1336759, "tid": 1381173,
    "ts": 1514657392895.626, "dur": 15.638,
    "args": {
      "External id": 2486651,"kernel_hash": "cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u7/cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [8388608, 2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [16, 4096, 2048], []], "Ev Idx": 3450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393023.651, "dur": 18.510,
    "args": {
      "External id": 2486652,"Record function id": 0, "Ev Idx": 3451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393028.833, "dur": 12.151,
    "args": {
      "External id": 2486653,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393033.152, "dur": 6.331,
    "args": {
      "External id": 2486654,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393034.941, "dur": 4.264,
    "args": {
      "External id": 2486655,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393048.288, "dur": 8.014,
    "args": {
      "External id": 2486656,"Record function id": 0, "Ev Idx": 3455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393049.835, "dur": 5.994,
    "args": {
      "External id": 2486657,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393050.773, "dur": 4.451,
    "args": {
      "External id": 2486658,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393051.707, "dur": 3.417,
    "args": {
      "External id": 2486659,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393059.593, "dur": 5.447,
    "args": {
      "External id": 2486660,"Record function id": 0, "Ev Idx": 3459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393061.016, "dur": 3.615,
    "args": {
      "External id": 2486661,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393061.948, "dur": 2.278,
    "args": {
      "External id": 2486662,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393063.146, "dur": 0.968,
    "args": {
      "External id": 2486663,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393068.243, "dur": 3.920,
    "args": {
      "External id": 2486664,"Record function id": 0, "Ev Idx": 3463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393069.508, "dur": 2.261,
    "args": {
      "External id": 2486665,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 3464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393070.033, "dur": 1.334,
    "args": {
      "External id": 2486666,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 3465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393070.513, "dur": 0.757,
    "args": {
      "External id": 2486667,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 3466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393075.375, "dur": 3.851,
    "args": {
      "External id": 2486668,"Record function id": 0, "Ev Idx": 3467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393076.753, "dur": 2.078,
    "args": {
      "External id": 2486669,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393077.365, "dur": 1.046,
    "args": {
      "External id": 2486670,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393077.780, "dur": 0.566,
    "args": {
      "External id": 2486671,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393085.259, "dur": 4.224,
    "args": {
      "External id": 2486672,"Record function id": 0, "Ev Idx": 3471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393086.485, "dur": 2.609,
    "args": {
      "External id": 2486673,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393087.146, "dur": 1.541,
    "args": {
      "External id": 2486674,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393087.661, "dur": 0.888,
    "args": {
      "External id": 2486675,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393092.660, "dur": 3.750,
    "args": {
      "External id": 2486676,"Record function id": 0, "Ev Idx": 3475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393093.602, "dur": 2.401,
    "args": {
      "External id": 2486677,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393094.305, "dur": 1.237,
    "args": {
      "External id": 2486678,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393094.780, "dur": 0.698,
    "args": {
      "External id": 2486679,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393099.435, "dur": 3.432,
    "args": {
      "External id": 2486680,"Record function id": 0, "Ev Idx": 3479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393100.552, "dur": 1.929,
    "args": {
      "External id": 2486681,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393101.034, "dur": 1.037,
    "args": {
      "External id": 2486682,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393101.282, "dur": 0.684,
    "args": {
      "External id": 2486683,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393105.935, "dur": 6.046,
    "args": {
      "External id": 2486684,"Record function id": 0, "Ev Idx": 3483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393106.988, "dur": 4.603,
    "args": {
      "External id": 2486685,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393107.493, "dur": 3.702,
    "args": {
      "External id": 2486686,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393108.183, "dur": 2.909,
    "args": {
      "External id": 2486687,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393115.705, "dur": 36408.832,
    "args": {
      "External id": 2486688,"Record function id": 0, "Sequence number": 24740451, "Fwd thread id": 1, "Ev Idx": 3487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393117.386, "dur": 36399.126,
    "args": {
      "External id": 2486689,"Sequence number": 24740451, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 3488
    }
  },
  {
    "ph": "f", "id": 197, "pid": 1336759, "tid": 1381173, "ts": 1514657393117.386,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.15)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393150.948, "dur": 35.779,
    "args": {
      "External id": 2486690,"Record function id": 0, "Ev Idx": 3489
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.15)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393193.880, "dur": 61.026,
    "args": {
      "External id": 2486691,"Record function id": 0, "Ev Idx": 3490
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.15)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393263.893, "dur": 36244.640,
    "args": {
      "External id": 2486692,"Record function id": 0, "Ev Idx": 3491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393349.371, "dur": 6.435,
    "args": {
      "External id": 2486693,"Record function id": 0, "Concrete Inputs": ["[51384320]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393365.019, "dur": 4.691,
    "args": {
      "External id": 2486694,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 3493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393384.700, "dur": 35288.637,
    "args": {
      "External id": 2486695,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 3494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393397.945, "dur": 35266.503,
    "args": {
      "External id": 2486696,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 3495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393454.620, "dur": 13.053,
    "args": {
      "External id": 2486697,"Record function id": 0, "Concrete Inputs": ["[12602]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393473.998, "dur": 35152.102,
    "args": {
      "External id": 2486698,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], [], []], "Ev Idx": 3497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393476.802, "dur": 35148.637,
    "args": {
      "External id": 2486699,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], []], "Ev Idx": 3498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393481.176, "dur": 5.580,
    "args": {
      "External id": 2486700,"Record function id": 0, "Concrete Inputs": ["[12602]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657393488.982, "dur": 35132.904,
    "args": {
      "External id": 2486701,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[12602], [12602], []], "Ev Idx": 3500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657428762.939, "dur": 8.998,
    "args": {
      "External id": 2486702,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[51384320], [], [], [], [], []], "Ev Idx": 3501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657428766.595, "dur": 4.979,
    "args": {
      "External id": 2486703,"Record function id": 0, "Concrete Inputs": ["[6423040]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657428800.038, "dur": 407.013,
    "args": {
      "External id": 2486704,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[6423040], [51384320], [], [], [], []], "Ev Idx": 3503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657428826.182, "dur": 375.116,
    "args": {
      "External id": 2486705,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 6423040, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[51384320], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 3504, "In msg nelems": 51384320
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657428837.074, "dur": 358.005,
    "args": {
      "External id": 2486706,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[51384320]], "Ev Idx": 3505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429231.621, "dur": 2.435,
    "args": {
      "External id": 2486707,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 3506, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429297.626, "dur": 6.492,
    "args": {
      "External id": 2486708,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429348.282, "dur": 3.669,
    "args": {
      "External id": 2486709,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "256"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429368.143, "dur": 1.053,
    "args": {
      "External id": 2486710,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429381.273, "dur": 1.109,
    "args": {
      "External id": 2486711,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1048832"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429393.737, "dur": 0.981,
    "args": {
      "External id": 2486712,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1573120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429406.259, "dur": 2.924,
    "args": {
      "External id": 2486713,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429421.067, "dur": 0.833,
    "args": {
      "External id": 2486714,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "2097664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429433.216, "dur": 2.035,
    "args": {
      "External id": 2486715,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "3539456"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429446.310, "dur": 1.022,
    "args": {
      "External id": 2486716,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "4981248"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429556.704, "dur": 2699.504,
    "args": {
      "External id": 2486717,"Record function id": 0, "Ev Idx": 3516
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.14)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429577.156, "dur": 1008.689,
    "args": {
      "External id": 2486718,"Record function id": 0, "Ev Idx": 3517
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.14)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429593.789, "dur": 309.734,
    "args": {
      "External id": 2486719,"Record function id": 0, "Ev Idx": 3518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429668.577, "dur": 6.318,
    "args": {
      "External id": 2486720,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 3519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429678.359, "dur": 1.230,
    "args": {
      "External id": 2486721,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 3520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429681.266, "dur": 1.109,
    "args": {
      "External id": 2486722,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 3521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429684.229, "dur": 0.957,
    "args": {
      "External id": 2486723,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 3522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429686.771, "dur": 0.834,
    "args": {
      "External id": 2486724,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 3523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429689.886, "dur": 0.921,
    "args": {
      "External id": 2486725,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 3524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429692.533, "dur": 1.891,
    "args": {
      "External id": 2486726,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 3525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429695.801, "dur": 0.777,
    "args": {
      "External id": 2486727,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 3526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429697.967, "dur": 2.888,
    "args": {
      "External id": 2486728,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 3527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429702.470, "dur": 0.879,
    "args": {
      "External id": 2486729,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 3528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429721.934, "dur": 150.731,
    "args": {
      "External id": 2486730,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 3529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429737.259, "dur": 131.089,
    "args": {
      "External id": 2486731,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 3530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429760.906, "dur": 12.669,
    "args": {
      "External id": 2486732,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429776.790, "dur": 64.343,
    "args": {
      "External id": 2486733,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 3532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429779.532, "dur": 61.228,
    "args": {
      "External id": 2486734,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 3533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429783.635, "dur": 5.647,
    "args": {
      "External id": 2486735,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429791.248, "dur": 48.850,
    "args": {
      "External id": 2486736,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 3535
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.13", "pid": 1336759, "tid": 1381173,
    "ts": 1514657429978.937, "dur": 599.078,
    "args": {
      "External id": 2486737,"Record function id": 0, "Ev Idx": 3536
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.13)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430033.480, "dur": 530.466,
    "args": {
      "External id": 2486738,"Record function id": 0, "Ev Idx": 3537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430092.694, "dur": 6.476,
    "args": {
      "External id": 2486739,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430114.658, "dur": 31.933,
    "args": {
      "External id": 2486740,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 3539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430120.219, "dur": 1.605,
    "args": {
      "External id": 2486741,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430124.191, "dur": 0.488,
    "args": {
      "External id": 2486742,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430126.681, "dur": 0.469,
    "args": {
      "External id": 2486743,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430128.699, "dur": 2.468,
    "args": {
      "External id": 2486744,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430133.357, "dur": 0.357,
    "args": {
      "External id": 2486745,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430135.225, "dur": 0.468,
    "args": {
      "External id": 2486746,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430137.392, "dur": 0.529,
    "args": {
      "External id": 2486747,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430139.611, "dur": 0.341,
    "args": {
      "External id": 2486748,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430141.722, "dur": 0.559,
    "args": {
      "External id": 2486749,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430156.578, "dur": 31.875,
    "args": {
      "External id": 2486750,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 3549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430220.611, "dur": 110.970,
    "args": {
      "External id": 2486751,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 3550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430230.701, "dur": 3.040,
    "args": {
      "External id": 2486752,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430245.123, "dur": 10.246,
    "args": {
      "External id": 2486753,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 3552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430249.676, "dur": 5.224,
    "args": {
      "External id": 2486754,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 3553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430253.226, "dur": 0.462,
    "args": {
      "External id": 2486755,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 3554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430264.803, "dur": 28.247,
    "args": {
      "External id": 2486756,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 3555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430267.449, "dur": 2.564,
    "args": {
      "External id": 2486757,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430271.543, "dur": 0.529,
    "args": {
      "External id": 2486758,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430273.984, "dur": 0.417,
    "args": {
      "External id": 2486759,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430276.056, "dur": 0.442,
    "args": {
      "External id": 2486760,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430278.275, "dur": 0.404,
    "args": {
      "External id": 2486761,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430280.070, "dur": 0.397,
    "args": {
      "External id": 2486762,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430282.380, "dur": 0.612,
    "args": {
      "External id": 2486763,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430284.712, "dur": 0.526,
    "args": {
      "External id": 2486764,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430286.826, "dur": 2.755,
    "args": {
      "External id": 2486765,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430304.287, "dur": 19.356,
    "args": {
      "External id": 2486766,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 3565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430375.843, "dur": 105.045,
    "args": {
      "External id": 2486767,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 3566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430396.755, "dur": 80.666,
    "args": {
      "External id": 2486768,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 3567, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430406.899, "dur": 66.622,
    "args": {
      "External id": 2486769,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 3568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430496.159, "dur": 1.906,
    "args": {
      "External id": 2486770,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 3569, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430594.016, "dur": 1639.293,
    "args": {
      "External id": 2486771,"Sequence number": 24740450, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 3570
    }
  },
  {
    "ph": "f", "id": 198, "pid": 1336759, "tid": 1381173, "ts": 1514657430594.016,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430699.771, "dur": 102.517,
    "args": {
      "External id": 2486772,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [5632, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 3571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430836.942, "dur": 39.892,
    "args": {
      "External id": 2486773,"kernel_hash": "cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/m5/cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [65536, 5632], [16, 4096, 5632], [16, 4096, 5632], [16, 4096, 5632], []], "Ev Idx": 3572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430896.583, "dur": 51.273,
    "args": {
      "External id": 2486774,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 2048], [0, 5632, 1], [11534336, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632], [1, 2048, 5632]], "Ev Idx": 3573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657430957.472, "dur": 71.276,
    "args": {
      "External id": 2486775,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 3574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657431040.683, "dur": 49.768,
    "args": {
      "External id": 2486776,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 3575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657431098.527, "dur": 27.919,
    "args": {
      "External id": 2486777,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 3576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657431134.712, "dur": 40.763,
    "args": {
      "External id": 2486778,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 3577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657431198.063, "dur": 21.738,
    "args": {
      "External id": 2486779,"kernel_hash": "c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/47/c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 3578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657431236.444, "dur": 26.931,
    "args": {
      "External id": 2486780,"kernel_hash": "cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/xr/cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657431282.291, "dur": 17.569,
    "args": {
      "External id": 2486781,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 3580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657431314.531, "dur": 13.692,
    "args": {
      "External id": 2486782,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 3581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657431337.351, "dur": 32.222,
    "args": {
      "External id": 2486783,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 3582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657431372.702, "dur": 33.051,
    "args": {
      "External id": 2486784,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 3583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657431454.370, "dur": 187.517,
    "args": {
      "External id": 2486785,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 3584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657431547.016, "dur": 7.010,
    "args": {
      "External id": 2486786,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657431556.465, "dur": 2.057,
    "args": {
      "External id": 2486787,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657431674.661, "dur": 23.301,
    "args": {
      "External id": 2486788,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657431713.117, "dur": 14.777,
    "args": {
      "External id": 2486789,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657431740.275, "dur": 43.158,
    "args": {
      "External id": 2486790,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 3589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657431793.143, "dur": 34.869,
    "args": {
      "External id": 2486791,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 3590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657431835.152, "dur": 20.338,
    "args": {
      "External id": 2486792,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 3591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657431861.336, "dur": 29.857,
    "args": {
      "External id": 2486793,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 3592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657431897.252, "dur": 20.416,
    "args": {
      "External id": 2486794,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 3593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657431927.349, "dur": 29.833,
    "args": {
      "External id": 2486795,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 3594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 1336759, "tid": 1381173,
    "ts": 1514657431973.292, "dur": 74.893,
    "args": {
      "External id": 2486796,"kernel_hash": "cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/g4/cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [65536, 2048], []], "Ev Idx": 3595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432081.523, "dur": 22.179,
    "args": {
      "External id": 2486797,"kernel_hash": "c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/4n/c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432124.300, "dur": 16.818,
    "args": {
      "External id": 2486798,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 3597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432159.226, "dur": 13.638,
    "args": {
      "External id": 2486799,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 3598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432188.638, "dur": 15.446,
    "args": {
      "External id": 2486800,"kernel_hash": "cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u7/cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [8388608, 2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [16, 4096, 2048], []], "Ev Idx": 3599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432279.370, "dur": 20.690,
    "args": {
      "External id": 2486801,"Record function id": 0, "Ev Idx": 3600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432282.610, "dur": 16.306,
    "args": {
      "External id": 2486802,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432286.710, "dur": 11.228,
    "args": {
      "External id": 2486803,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432288.300, "dur": 9.473,
    "args": {
      "External id": 2486804,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432303.785, "dur": 4.448,
    "args": {
      "External id": 2486805,"Record function id": 0, "Ev Idx": 3604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432304.932, "dur": 2.841,
    "args": {
      "External id": 2486806,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432305.521, "dur": 1.783,
    "args": {
      "External id": 2486807,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432306.147, "dur": 1.031,
    "args": {
      "External id": 2486808,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432311.472, "dur": 4.354,
    "args": {
      "External id": 2486809,"Record function id": 0, "Ev Idx": 3608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432312.746, "dur": 2.662,
    "args": {
      "External id": 2486810,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432313.440, "dur": 1.525,
    "args": {
      "External id": 2486811,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432313.872, "dur": 1.028,
    "args": {
      "External id": 2486812,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432318.905, "dur": 3.644,
    "args": {
      "External id": 2486813,"Record function id": 0, "Ev Idx": 3612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432319.997, "dur": 2.121,
    "args": {
      "External id": 2486814,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 3613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432320.488, "dur": 1.215,
    "args": {
      "External id": 2486815,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 3614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432320.902, "dur": 0.693,
    "args": {
      "External id": 2486816,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 3615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432325.598, "dur": 3.449,
    "args": {
      "External id": 2486817,"Record function id": 0, "Ev Idx": 3616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432326.858, "dur": 1.794,
    "args": {
      "External id": 2486818,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432327.331, "dur": 0.887,
    "args": {
      "External id": 2486819,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432327.583, "dur": 0.571,
    "args": {
      "External id": 2486820,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432332.164, "dur": 3.964,
    "args": {
      "External id": 2486821,"Record function id": 0, "Ev Idx": 3620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432333.398, "dur": 2.291,
    "args": {
      "External id": 2486822,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432333.963, "dur": 1.261,
    "args": {
      "External id": 2486823,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432334.402, "dur": 0.729,
    "args": {
      "External id": 2486824,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432339.235, "dur": 3.702,
    "args": {
      "External id": 2486825,"Record function id": 0, "Ev Idx": 3624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432340.444, "dur": 2.073,
    "args": {
      "External id": 2486826,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432340.958, "dur": 1.083,
    "args": {
      "External id": 2486827,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432341.224, "dur": 0.717,
    "args": {
      "External id": 2486828,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432345.976, "dur": 5.275,
    "args": {
      "External id": 2486829,"Record function id": 0, "Ev Idx": 3628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432346.964, "dur": 3.867,
    "args": {
      "External id": 2486830,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432347.550, "dur": 2.859,
    "args": {
      "External id": 2486831,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432347.810, "dur": 2.523,
    "args": {
      "External id": 2486832,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432357.547, "dur": 3.800,
    "args": {
      "External id": 2486833,"Record function id": 0, "Ev Idx": 3632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432358.740, "dur": 2.195,
    "args": {
      "External id": 2486834,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432359.247, "dur": 1.269,
    "args": {
      "External id": 2486835,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432359.640, "dur": 0.787,
    "args": {
      "External id": 2486836,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432365.032, "dur": 35546.655,
    "args": {
      "External id": 2486837,"Record function id": 0, "Sequence number": 24740449, "Fwd thread id": 1, "Ev Idx": 3636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432366.243, "dur": 35537.010,
    "args": {
      "External id": 2486838,"Sequence number": 24740449, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 3637
    }
  },
  {
    "ph": "f", "id": 199, "pid": 1336759, "tid": 1381173, "ts": 1514657432366.243,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.14)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432395.651, "dur": 41.153,
    "args": {
      "External id": 2486839,"Record function id": 0, "Ev Idx": 3638
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.14)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432444.346, "dur": 65.135,
    "args": {
      "External id": 2486840,"Record function id": 0, "Ev Idx": 3639
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.14)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432515.597, "dur": 35379.771,
    "args": {
      "External id": 2486841,"Record function id": 0, "Ev Idx": 3640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432624.175, "dur": 6.701,
    "args": {
      "External id": 2486842,"Record function id": 0, "Concrete Inputs": ["[51384320]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432641.246, "dur": 5.210,
    "args": {
      "External id": 2486843,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 3642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432661.959, "dur": 34445.857,
    "args": {
      "External id": 2486844,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 3643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432675.309, "dur": 34422.610,
    "args": {
      "External id": 2486845,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 3644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432716.746, "dur": 13.281,
    "args": {
      "External id": 2486846,"Record function id": 0, "Concrete Inputs": ["[12602]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432736.189, "dur": 34321.400,
    "args": {
      "External id": 2486847,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], [], []], "Ev Idx": 3646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432738.534, "dur": 34318.308,
    "args": {
      "External id": 2486848,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], []], "Ev Idx": 3647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432742.836, "dur": 5.120,
    "args": {
      "External id": 2486849,"Record function id": 0, "Concrete Inputs": ["[12602]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657432749.723, "dur": 34303.899,
    "args": {
      "External id": 2486850,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[12602], [12602], []], "Ev Idx": 3649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657467197.697, "dur": 9.374,
    "args": {
      "External id": 2486851,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[51384320], [], [], [], [], []], "Ev Idx": 3650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657467201.007, "dur": 5.733,
    "args": {
      "External id": 2486852,"Record function id": 0, "Concrete Inputs": ["[6423040]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657467236.108, "dur": 367.529,
    "args": {
      "External id": 2486853,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[6423040], [51384320], [], [], [], []], "Ev Idx": 3652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657467263.340, "dur": 334.681,
    "args": {
      "External id": 2486854,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 6423040, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[51384320], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 3653, "In msg nelems": 51384320
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657467274.125, "dur": 317.881,
    "args": {
      "External id": 2486855,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[51384320]], "Ev Idx": 3654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657467625.311, "dur": 2.357,
    "args": {
      "External id": 2486856,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 3655, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657467687.431, "dur": 6.575,
    "args": {
      "External id": 2486857,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657467740.913, "dur": 1.260,
    "args": {
      "External id": 2486858,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "256"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657467756.673, "dur": 1.012,
    "args": {
      "External id": 2486859,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657467770.187, "dur": 1.106,
    "args": {
      "External id": 2486860,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1048832"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657467782.153, "dur": 1.017,
    "args": {
      "External id": 2486861,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1573120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657467795.759, "dur": 0.824,
    "args": {
      "External id": 2486862,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657467807.329, "dur": 0.859,
    "args": {
      "External id": 2486863,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "2097664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657467819.409, "dur": 1.804,
    "args": {
      "External id": 2486864,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "3539456"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657467831.364, "dur": 0.945,
    "args": {
      "External id": 2486865,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "4981248"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657467926.577, "dur": 2713.418,
    "args": {
      "External id": 2486866,"Record function id": 0, "Ev Idx": 3665
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.13)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657467944.311, "dur": 1000.375,
    "args": {
      "External id": 2486867,"Record function id": 0, "Ev Idx": 3666
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.13)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657467959.390, "dur": 361.651,
    "args": {
      "External id": 2486868,"Record function id": 0, "Ev Idx": 3667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468074.620, "dur": 4.671,
    "args": {
      "External id": 2486869,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 3668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468082.659, "dur": 0.752,
    "args": {
      "External id": 2486870,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 3669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468085.270, "dur": 0.812,
    "args": {
      "External id": 2486871,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 3670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468087.964, "dur": 0.850,
    "args": {
      "External id": 2486872,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 3671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468090.245, "dur": 1.031,
    "args": {
      "External id": 2486873,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 3672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468092.580, "dur": 0.766,
    "args": {
      "External id": 2486874,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 3673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468094.695, "dur": 1.887,
    "args": {
      "External id": 2486875,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 3674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468098.457, "dur": 2.470,
    "args": {
      "External id": 2486876,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 3675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468102.275, "dur": 0.693,
    "args": {
      "External id": 2486877,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 3676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468104.538, "dur": 0.619,
    "args": {
      "External id": 2486878,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 3677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468125.956, "dur": 161.717,
    "args": {
      "External id": 2486879,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 3678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468141.885, "dur": 141.227,
    "args": {
      "External id": 2486880,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 3679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468169.867, "dur": 12.493,
    "args": {
      "External id": 2486881,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468185.487, "dur": 70.223,
    "args": {
      "External id": 2486882,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 3681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468188.146, "dur": 67.198,
    "args": {
      "External id": 2486883,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 3682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468192.499, "dur": 5.297,
    "args": {
      "External id": 2486884,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468199.818, "dur": 54.759,
    "args": {
      "External id": 2486885,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 3684
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.12", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468401.316, "dur": 535.683,
    "args": {
      "External id": 2486886,"Record function id": 0, "Ev Idx": 3685
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.12)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468417.850, "dur": 506.246,
    "args": {
      "External id": 2486887,"Record function id": 0, "Ev Idx": 3686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468471.209, "dur": 4.557,
    "args": {
      "External id": 2486888,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468489.976, "dur": 27.967,
    "args": {
      "External id": 2486889,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 3688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468495.031, "dur": 1.734,
    "args": {
      "External id": 2486890,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468499.056, "dur": 0.415,
    "args": {
      "External id": 2486891,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468501.062, "dur": 2.139,
    "args": {
      "External id": 2486892,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468504.649, "dur": 0.359,
    "args": {
      "External id": 2486893,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468506.446, "dur": 0.374,
    "args": {
      "External id": 2486894,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468508.218, "dur": 0.414,
    "args": {
      "External id": 2486895,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468509.862, "dur": 0.362,
    "args": {
      "External id": 2486896,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468511.679, "dur": 0.447,
    "args": {
      "External id": 2486897,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468513.735, "dur": 0.333,
    "args": {
      "External id": 2486898,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468527.018, "dur": 48.933,
    "args": {
      "External id": 2486899,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 3698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468610.292, "dur": 98.391,
    "args": {
      "External id": 2486900,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 3699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468620.978, "dur": 4.340,
    "args": {
      "External id": 2486901,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468630.538, "dur": 12.384,
    "args": {
      "External id": 2486902,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 3701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468634.695, "dur": 7.788,
    "args": {
      "External id": 2486903,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 3702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468638.360, "dur": 2.676,
    "args": {
      "External id": 2486904,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 3703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468649.686, "dur": 22.132,
    "args": {
      "External id": 2486905,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 3704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468651.783, "dur": 0.657,
    "args": {
      "External id": 2486906,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468654.094, "dur": 0.416,
    "args": {
      "External id": 2486907,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468656.034, "dur": 0.325,
    "args": {
      "External id": 2486908,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468657.743, "dur": 0.364,
    "args": {
      "External id": 2486909,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468659.346, "dur": 0.369,
    "args": {
      "External id": 2486910,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468661.203, "dur": 0.353,
    "args": {
      "External id": 2486911,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468662.639, "dur": 0.371,
    "args": {
      "External id": 2486912,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468664.388, "dur": 2.028,
    "args": {
      "External id": 2486913,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468667.988, "dur": 0.381,
    "args": {
      "External id": 2486914,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468682.915, "dur": 18.243,
    "args": {
      "External id": 2486915,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 3714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468753.346, "dur": 107.601,
    "args": {
      "External id": 2486916,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 3715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468775.758, "dur": 81.538,
    "args": {
      "External id": 2486917,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 3716, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468785.537, "dur": 67.430,
    "args": {
      "External id": 2486918,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 3717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468874.660, "dur": 1.839,
    "args": {
      "External id": 2486919,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 3718, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657468951.601, "dur": 1668.417,
    "args": {
      "External id": 2486920,"Sequence number": 24740448, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 3719
    }
  },
  {
    "ph": "f", "id": 200, "pid": 1336759, "tid": 1381173, "ts": 1514657468951.601,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657469097.609, "dur": 105.265,
    "args": {
      "External id": 2486921,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [5632, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 3720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657469240.434, "dur": 37.450,
    "args": {
      "External id": 2486922,"kernel_hash": "cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/m5/cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [65536, 5632], [16, 4096, 5632], [16, 4096, 5632], [16, 4096, 5632], []], "Ev Idx": 3721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657469296.371, "dur": 48.364,
    "args": {
      "External id": 2486923,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 2048], [0, 5632, 1], [11534336, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632], [1, 2048, 5632]], "Ev Idx": 3722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657469354.697, "dur": 32.827,
    "args": {
      "External id": 2486924,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 3723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657469395.379, "dur": 45.192,
    "args": {
      "External id": 2486925,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 3724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657469449.392, "dur": 27.419,
    "args": {
      "External id": 2486926,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 3725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657469486.500, "dur": 59.053,
    "args": {
      "External id": 2486927,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 3726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657469572.103, "dur": 23.957,
    "args": {
      "External id": 2486928,"kernel_hash": "c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/47/c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 3727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657469614.106, "dur": 26.629,
    "args": {
      "External id": 2486929,"kernel_hash": "cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/xr/cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657469659.567, "dur": 18.486,
    "args": {
      "External id": 2486930,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 3729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657469691.489, "dur": 13.937,
    "args": {
      "External id": 2486931,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 3730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657469715.548, "dur": 30.130,
    "args": {
      "External id": 2486932,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 3731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657469749.477, "dur": 32.739,
    "args": {
      "External id": 2486933,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 3732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657469812.467, "dur": 166.592,
    "args": {
      "External id": 2486934,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 3733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657469889.965, "dur": 6.144,
    "args": {
      "External id": 2486935,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657469898.140, "dur": 1.969,
    "args": {
      "External id": 2486936,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470064.164, "dur": 28.319,
    "args": {
      "External id": 2486937,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470122.078, "dur": 15.627,
    "args": {
      "External id": 2486938,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470149.963, "dur": 43.501,
    "args": {
      "External id": 2486939,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 3738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470201.068, "dur": 38.224,
    "args": {
      "External id": 2486940,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 3739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470246.567, "dur": 20.464,
    "args": {
      "External id": 2486941,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 3740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470272.997, "dur": 30.380,
    "args": {
      "External id": 2486942,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 3741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470311.362, "dur": 20.413,
    "args": {
      "External id": 2486943,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 3742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470338.977, "dur": 44.435,
    "args": {
      "External id": 2486944,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 3743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470406.108, "dur": 23.026,
    "args": {
      "External id": 2486945,"kernel_hash": "cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/g4/cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [65536, 2048], []], "Ev Idx": 3744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470445.939, "dur": 25.009,
    "args": {
      "External id": 2486946,"kernel_hash": "c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/4n/c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470485.879, "dur": 16.593,
    "args": {
      "External id": 2486947,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 3746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470519.731, "dur": 33.358,
    "args": {
      "External id": 2486948,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 3747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470571.022, "dur": 17.382,
    "args": {
      "External id": 2486949,"kernel_hash": "cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u7/cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [8388608, 2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [16, 4096, 2048], []], "Ev Idx": 3748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470663.492, "dur": 15.472,
    "args": {
      "External id": 2486950,"Record function id": 0, "Ev Idx": 3749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470666.717, "dur": 11.287,
    "args": {
      "External id": 2486951,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470670.825, "dur": 6.164,
    "args": {
      "External id": 2486952,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470672.704, "dur": 4.158,
    "args": {
      "External id": 2486953,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470682.678, "dur": 8.566,
    "args": {
      "External id": 2486954,"Record function id": 0, "Ev Idx": 3753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470684.111, "dur": 6.701,
    "args": {
      "External id": 2486955,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470684.778, "dur": 5.457,
    "args": {
      "External id": 2486956,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470685.366, "dur": 4.737,
    "args": {
      "External id": 2486957,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470694.514, "dur": 4.146,
    "args": {
      "External id": 2486958,"Record function id": 0, "Ev Idx": 3757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470695.700, "dur": 2.566,
    "args": {
      "External id": 2486959,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470696.304, "dur": 1.473,
    "args": {
      "External id": 2486960,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470696.730, "dur": 0.969,
    "args": {
      "External id": 2486961,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470701.870, "dur": 3.893,
    "args": {
      "External id": 2486962,"Record function id": 0, "Ev Idx": 3761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470703.121, "dur": 2.239,
    "args": {
      "External id": 2486963,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 3762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470703.650, "dur": 1.307,
    "args": {
      "External id": 2486964,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 3763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470704.192, "dur": 0.648,
    "args": {
      "External id": 2486965,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 3764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470708.866, "dur": 3.770,
    "args": {
      "External id": 2486966,"Record function id": 0, "Ev Idx": 3765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470710.074, "dur": 2.155,
    "args": {
      "External id": 2486967,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470710.737, "dur": 1.098,
    "args": {
      "External id": 2486968,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470711.204, "dur": 0.564,
    "args": {
      "External id": 2486969,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470715.914, "dur": 3.588,
    "args": {
      "External id": 2486970,"Record function id": 0, "Ev Idx": 3769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470716.905, "dur": 2.201,
    "args": {
      "External id": 2486971,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470717.429, "dur": 1.168,
    "args": {
      "External id": 2486972,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470717.866, "dur": 0.627,
    "args": {
      "External id": 2486973,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470722.704, "dur": 5.391,
    "args": {
      "External id": 2486974,"Record function id": 0, "Ev Idx": 3773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470723.713, "dur": 3.971,
    "args": {
      "External id": 2486975,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470724.200, "dur": 3.046,
    "args": {
      "External id": 2486976,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470724.470, "dur": 2.703,
    "args": {
      "External id": 2486977,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470731.215, "dur": 3.665,
    "args": {
      "External id": 2486978,"Record function id": 0, "Ev Idx": 3777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470732.422, "dur": 2.025,
    "args": {
      "External id": 2486979,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470732.868, "dur": 1.144,
    "args": {
      "External id": 2486980,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470733.128, "dur": 0.789,
    "args": {
      "External id": 2486981,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470737.972, "dur": 3.358,
    "args": {
      "External id": 2486982,"Record function id": 0, "Ev Idx": 3781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470738.937, "dur": 1.985,
    "args": {
      "External id": 2486983,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470739.376, "dur": 1.096,
    "args": {
      "External id": 2486984,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470739.840, "dur": 0.545,
    "args": {
      "External id": 2486985,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470748.620, "dur": 38753.080,
    "args": {
      "External id": 2486986,"Record function id": 0, "Sequence number": 24740447, "Fwd thread id": 1, "Ev Idx": 3785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470749.984, "dur": 38742.389,
    "args": {
      "External id": 2486987,"Sequence number": 24740447, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 3786
    }
  },
  {
    "ph": "f", "id": 201, "pid": 1336759, "tid": 1381173, "ts": 1514657470749.984,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.13)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470780.449, "dur": 39.433,
    "args": {
      "External id": 2486988,"Record function id": 0, "Ev Idx": 3787
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.13)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470828.151, "dur": 62.973,
    "args": {
      "External id": 2486989,"Record function id": 0, "Ev Idx": 3788
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.13)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657470896.800, "dur": 38584.561,
    "args": {
      "External id": 2486990,"Record function id": 0, "Ev Idx": 3789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657471024.843, "dur": 8.072,
    "args": {
      "External id": 2486991,"Record function id": 0, "Concrete Inputs": ["[51384320]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657471047.406, "dur": 4.813,
    "args": {
      "External id": 2486992,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 3791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657471068.032, "dur": 37521.382,
    "args": {
      "External id": 2486993,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 3792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657471081.331, "dur": 37499.271,
    "args": {
      "External id": 2486994,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 3793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657471126.406, "dur": 13.628,
    "args": {
      "External id": 2486995,"Record function id": 0, "Concrete Inputs": ["[12602]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657471146.232, "dur": 37379.775,
    "args": {
      "External id": 2486996,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], [], []], "Ev Idx": 3795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657471148.813, "dur": 37376.438,
    "args": {
      "External id": 2486997,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], []], "Ev Idx": 3796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657471153.054, "dur": 5.976,
    "args": {
      "External id": 2486998,"Record function id": 0, "Concrete Inputs": ["[12602]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657471160.932, "dur": 37360.759,
    "args": {
      "External id": 2486999,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[12602], [12602], []], "Ev Idx": 3798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657508677.627, "dur": 8.545,
    "args": {
      "External id": 2487000,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[51384320], [], [], [], [], []], "Ev Idx": 3799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657508680.543, "dur": 5.277,
    "args": {
      "External id": 2487001,"Record function id": 0, "Concrete Inputs": ["[6423040]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657508713.083, "dur": 443.979,
    "args": {
      "External id": 2487002,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[6423040], [51384320], [], [], [], []], "Ev Idx": 3801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657508738.511, "dur": 413.020,
    "args": {
      "External id": 2487003,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 6423040, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[51384320], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 3802, "In msg nelems": 51384320
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657508749.510, "dur": 395.808,
    "args": {
      "External id": 2487004,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[51384320]], "Ev Idx": 3803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657509179.470, "dur": 2.268,
    "args": {
      "External id": 2487005,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 3804, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657509244.500, "dur": 8.466,
    "args": {
      "External id": 2487006,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657509296.709, "dur": 1.424,
    "args": {
      "External id": 2487007,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "256"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657509313.815, "dur": 1.063,
    "args": {
      "External id": 2487008,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657509327.897, "dur": 1.003,
    "args": {
      "External id": 2487009,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1048832"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657509339.937, "dur": 2.791,
    "args": {
      "External id": 2487010,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1573120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657509356.050, "dur": 1.057,
    "args": {
      "External id": 2487011,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657509368.826, "dur": 1.194,
    "args": {
      "External id": 2487012,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "2097664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657509382.532, "dur": 1.912,
    "args": {
      "External id": 2487013,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "3539456"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657509398.172, "dur": 3.217,
    "args": {
      "External id": 2487014,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "4981248"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657509523.944, "dur": 2733.552,
    "args": {
      "External id": 2487015,"Record function id": 0, "Ev Idx": 3814
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.12)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657509561.242, "dur": 1057.209,
    "args": {
      "External id": 2487016,"Record function id": 0, "Ev Idx": 3815
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.12)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657509578.637, "dur": 333.347,
    "args": {
      "External id": 2487017,"Record function id": 0, "Ev Idx": 3816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657509664.654, "dur": 5.212,
    "args": {
      "External id": 2487018,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 3817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657509673.470, "dur": 1.138,
    "args": {
      "External id": 2487019,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 3818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657509676.642, "dur": 0.945,
    "args": {
      "External id": 2487020,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 3819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657509679.037, "dur": 0.853,
    "args": {
      "External id": 2487021,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 3820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657509681.450, "dur": 1.009,
    "args": {
      "External id": 2487022,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 3821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657509684.152, "dur": 0.676,
    "args": {
      "External id": 2487023,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 3822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657509686.324, "dur": 4.028,
    "args": {
      "External id": 2487024,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 3823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657509692.086, "dur": 0.852,
    "args": {
      "External id": 2487025,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 3824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657509694.686, "dur": 0.642,
    "args": {
      "External id": 2487026,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 3825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657509696.724, "dur": 0.654,
    "args": {
      "External id": 2487027,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 3826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657509715.177, "dur": 160.445,
    "args": {
      "External id": 2487028,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 3827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657509730.985, "dur": 140.384,
    "args": {
      "External id": 2487029,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 3828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657509756.588, "dur": 13.795,
    "args": {
      "External id": 2487030,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657509773.599, "dur": 68.876,
    "args": {
      "External id": 2487031,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 3830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657509776.244, "dur": 65.869,
    "args": {
      "External id": 2487032,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 3831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657509780.223, "dur": 5.114,
    "args": {
      "External id": 2487033,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657509787.303, "dur": 54.019,
    "args": {
      "External id": 2487034,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 3833
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.11", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510038.574, "dur": 571.897,
    "args": {
      "External id": 2487035,"Record function id": 0, "Ev Idx": 3834
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.11)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510057.287, "dur": 539.665,
    "args": {
      "External id": 2487036,"Record function id": 0, "Ev Idx": 3835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510121.840, "dur": 6.534,
    "args": {
      "External id": 2487037,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510143.985, "dur": 30.062,
    "args": {
      "External id": 2487038,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 3837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510149.371, "dur": 1.947,
    "args": {
      "External id": 2487039,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510153.569, "dur": 2.078,
    "args": {
      "External id": 2487040,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510157.632, "dur": 0.370,
    "args": {
      "External id": 2487041,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510159.746, "dur": 0.352,
    "args": {
      "External id": 2487042,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510161.591, "dur": 0.396,
    "args": {
      "External id": 2487043,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510163.374, "dur": 0.452,
    "args": {
      "External id": 2487044,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510165.419, "dur": 0.372,
    "args": {
      "External id": 2487045,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510167.257, "dur": 0.427,
    "args": {
      "External id": 2487046,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510169.299, "dur": 0.388,
    "args": {
      "External id": 2487047,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510184.233, "dur": 34.945,
    "args": {
      "External id": 2487048,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 3847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510254.056, "dur": 101.863,
    "args": {
      "External id": 2487049,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 3848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510264.110, "dur": 5.722,
    "args": {
      "External id": 2487050,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510274.985, "dur": 10.861,
    "args": {
      "External id": 2487051,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 3850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510279.447, "dur": 5.989,
    "args": {
      "External id": 2487052,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 3851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510283.683, "dur": 0.615,
    "args": {
      "External id": 2487053,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 3852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510292.891, "dur": 23.427,
    "args": {
      "External id": 2487054,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 3853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510295.230, "dur": 0.521,
    "args": {
      "External id": 2487055,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510297.524, "dur": 0.445,
    "args": {
      "External id": 2487056,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510299.372, "dur": 0.490,
    "args": {
      "External id": 2487057,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510301.415, "dur": 0.322,
    "args": {
      "External id": 2487058,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510303.208, "dur": 0.375,
    "args": {
      "External id": 2487059,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510305.268, "dur": 0.389,
    "args": {
      "External id": 2487060,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510307.200, "dur": 1.981,
    "args": {
      "External id": 2487061,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510310.789, "dur": 0.398,
    "args": {
      "External id": 2487062,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510312.726, "dur": 0.391,
    "args": {
      "External id": 2487063,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510328.227, "dur": 20.314,
    "args": {
      "External id": 2487064,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 3863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510401.856, "dur": 107.014,
    "args": {
      "External id": 2487065,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 3864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510423.502, "dur": 81.638,
    "args": {
      "External id": 2487066,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 3865, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510432.941, "dur": 67.352,
    "args": {
      "External id": 2487067,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 3866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510522.085, "dur": 1.885,
    "args": {
      "External id": 2487068,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 3867, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510626.220, "dur": 1607.809,
    "args": {
      "External id": 2487069,"Sequence number": 24740446, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 3868
    }
  },
  {
    "ph": "f", "id": 202, "pid": 1336759, "tid": 1381173, "ts": 1514657510626.220,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510732.361, "dur": 102.877,
    "args": {
      "External id": 2487070,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [5632, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 3869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510871.200, "dur": 38.609,
    "args": {
      "External id": 2487071,"kernel_hash": "cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/m5/cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [65536, 5632], [16, 4096, 5632], [16, 4096, 5632], [16, 4096, 5632], []], "Ev Idx": 3870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657510928.697, "dur": 48.635,
    "args": {
      "External id": 2487072,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 2048], [0, 5632, 1], [11534336, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632], [1, 2048, 5632]], "Ev Idx": 3871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657511029.794, "dur": 38.684,
    "args": {
      "External id": 2487073,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 3872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657511077.152, "dur": 46.287,
    "args": {
      "External id": 2487074,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 3873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657511133.077, "dur": 27.717,
    "args": {
      "External id": 2487075,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 3874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657511172.082, "dur": 43.539,
    "args": {
      "External id": 2487076,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 3875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657511239.544, "dur": 22.931,
    "args": {
      "External id": 2487077,"kernel_hash": "c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/47/c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 3876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657511279.660, "dur": 27.418,
    "args": {
      "External id": 2487078,"kernel_hash": "cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/xr/cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657511326.291, "dur": 19.016,
    "args": {
      "External id": 2487079,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 3878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657511360.250, "dur": 14.589,
    "args": {
      "External id": 2487080,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 3879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657511385.475, "dur": 27.457,
    "args": {
      "External id": 2487081,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 3880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657511416.175, "dur": 32.470,
    "args": {
      "External id": 2487082,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 3881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657511475.356, "dur": 184.886,
    "args": {
      "External id": 2487083,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 3882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657511565.686, "dur": 6.597,
    "args": {
      "External id": 2487084,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657511574.672, "dur": 3.636,
    "args": {
      "External id": 2487085,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657511694.911, "dur": 22.311,
    "args": {
      "External id": 2487086,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657511728.238, "dur": 16.972,
    "args": {
      "External id": 2487087,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657511753.737, "dur": 37.566,
    "args": {
      "External id": 2487088,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 3887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657511798.568, "dur": 34.878,
    "args": {
      "External id": 2487089,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 3888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657511840.661, "dur": 20.526,
    "args": {
      "External id": 2487090,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 3889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657511869.426, "dur": 30.299,
    "args": {
      "External id": 2487091,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 3890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657511906.463, "dur": 20.462,
    "args": {
      "External id": 2487092,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 3891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657511933.906, "dur": 42.743,
    "args": {
      "External id": 2487093,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 3892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512035.922, "dur": 26.063,
    "args": {
      "External id": 2487094,"kernel_hash": "cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/g4/cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [65536, 2048], []], "Ev Idx": 3893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512082.946, "dur": 22.430,
    "args": {
      "External id": 2487095,"kernel_hash": "c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/4n/c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 3894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512120.468, "dur": 20.518,
    "args": {
      "External id": 2487096,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 3895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512156.619, "dur": 14.915,
    "args": {
      "External id": 2487097,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 3896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512188.906, "dur": 16.412,
    "args": {
      "External id": 2487098,"kernel_hash": "cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u7/cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [8388608, 2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [16, 4096, 2048], []], "Ev Idx": 3897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512279.553, "dur": 14.898,
    "args": {
      "External id": 2487099,"Record function id": 0, "Ev Idx": 3898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512282.674, "dur": 10.832,
    "args": {
      "External id": 2487100,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512286.752, "dur": 5.848,
    "args": {
      "External id": 2487101,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512288.399, "dur": 4.025,
    "args": {
      "External id": 2487102,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512301.272, "dur": 5.242,
    "args": {
      "External id": 2487103,"Record function id": 0, "Ev Idx": 3902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512302.520, "dur": 3.536,
    "args": {
      "External id": 2487104,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512303.363, "dur": 2.171,
    "args": {
      "External id": 2487105,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512304.210, "dur": 1.203,
    "args": {
      "External id": 2487106,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512309.689, "dur": 4.086,
    "args": {
      "External id": 2487107,"Record function id": 0, "Ev Idx": 3906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512310.789, "dur": 2.568,
    "args": {
      "External id": 2487108,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512311.356, "dur": 1.537,
    "args": {
      "External id": 2487109,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512311.749, "dur": 1.077,
    "args": {
      "External id": 2487110,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 3909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512316.983, "dur": 5.954,
    "args": {
      "External id": 2487111,"Record function id": 0, "Ev Idx": 3910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512318.335, "dur": 4.179,
    "args": {
      "External id": 2487112,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 3911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512318.928, "dur": 3.147,
    "args": {
      "External id": 2487113,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 3912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512319.563, "dur": 2.444,
    "args": {
      "External id": 2487114,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 3913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512325.979, "dur": 3.427,
    "args": {
      "External id": 2487115,"Record function id": 0, "Ev Idx": 3914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512326.962, "dur": 2.039,
    "args": {
      "External id": 2487116,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512327.420, "dur": 1.155,
    "args": {
      "External id": 2487117,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512327.795, "dur": 0.685,
    "args": {
      "External id": 2487118,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512332.518, "dur": 4.060,
    "args": {
      "External id": 2487119,"Record function id": 0, "Ev Idx": 3918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512333.601, "dur": 2.537,
    "args": {
      "External id": 2487120,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512334.064, "dur": 1.637,
    "args": {
      "External id": 2487121,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512334.552, "dur": 1.050,
    "args": {
      "External id": 2487122,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512339.773, "dur": 3.542,
    "args": {
      "External id": 2487123,"Record function id": 0, "Ev Idx": 3922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512340.987, "dur": 1.931,
    "args": {
      "External id": 2487124,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512341.462, "dur": 1.025,
    "args": {
      "External id": 2487125,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512341.729, "dur": 0.633,
    "args": {
      "External id": 2487126,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512346.312, "dur": 3.663,
    "args": {
      "External id": 2487127,"Record function id": 0, "Ev Idx": 3926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512347.433, "dur": 2.133,
    "args": {
      "External id": 2487128,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512347.979, "dur": 1.184,
    "args": {
      "External id": 2487129,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512348.373, "dur": 0.725,
    "args": {
      "External id": 2487130,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 3929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512352.900, "dur": 32.332,
    "args": {
      "External id": 2487131,"Record function id": 0, "Ev Idx": 3930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512353.954, "dur": 30.786,
    "args": {
      "External id": 2487132,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512354.603, "dur": 29.672,
    "args": {
      "External id": 2487133,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512383.459, "dur": 0.681,
    "args": {
      "External id": 2487134,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 3933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512389.031, "dur": 35465.059,
    "args": {
      "External id": 2487135,"Record function id": 0, "Sequence number": 24740445, "Fwd thread id": 1, "Ev Idx": 3934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512390.358, "dur": 35455.642,
    "args": {
      "External id": 2487136,"Sequence number": 24740445, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 3935
    }
  },
  {
    "ph": "f", "id": 203, "pid": 1336759, "tid": 1381173, "ts": 1514657512390.358,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.12)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512422.799, "dur": 36.773,
    "args": {
      "External id": 2487137,"Record function id": 0, "Ev Idx": 3936
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.12)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512466.633, "dur": 79.252,
    "args": {
      "External id": 2487138,"Record function id": 0, "Ev Idx": 3937
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.12)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512553.753, "dur": 35284.715,
    "args": {
      "External id": 2487139,"Record function id": 0, "Ev Idx": 3938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512641.825, "dur": 7.404,
    "args": {
      "External id": 2487140,"Record function id": 0, "Concrete Inputs": ["[51384320]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512659.362, "dur": 4.818,
    "args": {
      "External id": 2487141,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 3940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512679.383, "dur": 34359.653,
    "args": {
      "External id": 2487142,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 3941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512692.308, "dur": 34337.057,
    "args": {
      "External id": 2487143,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 3942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512731.569, "dur": 15.911,
    "args": {
      "External id": 2487144,"Record function id": 0, "Concrete Inputs": ["[12602]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512753.737, "dur": 34211.741,
    "args": {
      "External id": 2487145,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], [], []], "Ev Idx": 3944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512756.634, "dur": 34208.083,
    "args": {
      "External id": 2487146,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], []], "Ev Idx": 3945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512760.634, "dur": 4.907,
    "args": {
      "External id": 2487147,"Record function id": 0, "Concrete Inputs": ["[12602]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657512767.446, "dur": 34193.716,
    "args": {
      "External id": 2487148,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[12602], [12602], []], "Ev Idx": 3947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657547127.435, "dur": 10.000,
    "args": {
      "External id": 2487149,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[51384320], [], [], [], [], []], "Ev Idx": 3948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657547130.697, "dur": 6.289,
    "args": {
      "External id": 2487150,"Record function id": 0, "Concrete Inputs": ["[6423040]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657547166.979, "dur": 346.440,
    "args": {
      "External id": 2487151,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[6423040], [51384320], [], [], [], []], "Ev Idx": 3950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657547192.730, "dur": 316.211,
    "args": {
      "External id": 2487152,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 6423040, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[51384320], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 3951, "In msg nelems": 51384320
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657547203.785, "dur": 300.287,
    "args": {
      "External id": 2487153,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[51384320]], "Ev Idx": 3952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657547549.496, "dur": 3.179,
    "args": {
      "External id": 2487154,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 3953, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657547615.778, "dur": 6.713,
    "args": {
      "External id": 2487155,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657547674.342, "dur": 1.725,
    "args": {
      "External id": 2487156,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "256"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657547692.611, "dur": 1.170,
    "args": {
      "External id": 2487157,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657547708.463, "dur": 0.971,
    "args": {
      "External id": 2487158,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1048832"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657547721.612, "dur": 0.903,
    "args": {
      "External id": 2487159,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1573120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657547734.129, "dur": 0.943,
    "args": {
      "External id": 2487160,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657547747.507, "dur": 1.245,
    "args": {
      "External id": 2487161,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "2097664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657547761.948, "dur": 2.008,
    "args": {
      "External id": 2487162,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "3539456"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657547775.551, "dur": 0.904,
    "args": {
      "External id": 2487163,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "4981248"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657547869.703, "dur": 2615.698,
    "args": {
      "External id": 2487164,"Record function id": 0, "Ev Idx": 3963
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.11)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657547887.333, "dur": 978.071,
    "args": {
      "External id": 2487165,"Record function id": 0, "Ev Idx": 3964
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.11)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657547902.536, "dur": 345.398,
    "args": {
      "External id": 2487166,"Record function id": 0, "Ev Idx": 3965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657547977.507, "dur": 38.645,
    "args": {
      "External id": 2487167,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 3966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548021.833, "dur": 1.361,
    "args": {
      "External id": 2487168,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 3967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548024.992, "dur": 0.665,
    "args": {
      "External id": 2487169,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 3968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548027.338, "dur": 2.668,
    "args": {
      "External id": 2487170,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 3969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548031.340, "dur": 1.004,
    "args": {
      "External id": 2487171,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 3970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548033.864, "dur": 0.803,
    "args": {
      "External id": 2487172,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 3971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548036.240, "dur": 1.634,
    "args": {
      "External id": 2487173,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 3972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548039.222, "dur": 0.658,
    "args": {
      "External id": 2487174,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 3973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548041.380, "dur": 0.863,
    "args": {
      "External id": 2487175,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 3974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548043.652, "dur": 0.679,
    "args": {
      "External id": 2487176,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 3975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548063.163, "dur": 151.463,
    "args": {
      "External id": 2487177,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 3976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548078.926, "dur": 131.074,
    "args": {
      "External id": 2487178,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 3977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548100.227, "dur": 12.064,
    "args": {
      "External id": 2487179,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548115.829, "dur": 67.200,
    "args": {
      "External id": 2487180,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 3979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548118.488, "dur": 64.176,
    "args": {
      "External id": 2487181,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 3980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548122.571, "dur": 7.698,
    "args": {
      "External id": 2487182,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548132.200, "dur": 49.850,
    "args": {
      "External id": 2487183,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 3982
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.10", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548326.097, "dur": 531.340,
    "args": {
      "External id": 2487184,"Record function id": 0, "Ev Idx": 3983
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.10)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548341.943, "dur": 503.193,
    "args": {
      "External id": 2487185,"Record function id": 0, "Ev Idx": 3984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548399.430, "dur": 4.459,
    "args": {
      "External id": 2487186,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548418.556, "dur": 27.882,
    "args": {
      "External id": 2487187,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 3986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548423.542, "dur": 1.629,
    "args": {
      "External id": 2487188,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548427.002, "dur": 0.663,
    "args": {
      "External id": 2487189,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548429.313, "dur": 0.474,
    "args": {
      "External id": 2487190,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548431.134, "dur": 0.390,
    "args": {
      "External id": 2487191,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548433.048, "dur": 0.463,
    "args": {
      "External id": 2487192,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548434.934, "dur": 0.428,
    "args": {
      "External id": 2487193,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548436.914, "dur": 2.150,
    "args": {
      "External id": 2487194,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548440.572, "dur": 0.439,
    "args": {
      "External id": 2487195,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548442.287, "dur": 0.363,
    "args": {
      "External id": 2487196,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 3995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548455.270, "dur": 30.761,
    "args": {
      "External id": 2487197,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 3996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548516.248, "dur": 118.920,
    "args": {
      "External id": 2487198,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 3997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548525.408, "dur": 19.596,
    "args": {
      "External id": 2487199,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 3998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548551.597, "dur": 10.928,
    "args": {
      "External id": 2487200,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 3999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548555.744, "dur": 6.348,
    "args": {
      "External id": 2487201,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 4000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548559.735, "dur": 0.916,
    "args": {
      "External id": 2487202,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 4001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548570.582, "dur": 23.155,
    "args": {
      "External id": 2487203,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 4002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548572.805, "dur": 0.584,
    "args": {
      "External id": 2487204,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548575.328, "dur": 0.510,
    "args": {
      "External id": 2487205,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548577.328, "dur": 0.432,
    "args": {
      "External id": 2487206,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548579.257, "dur": 2.039,
    "args": {
      "External id": 2487207,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548582.685, "dur": 0.438,
    "args": {
      "External id": 2487208,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548584.528, "dur": 0.426,
    "args": {
      "External id": 2487209,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548586.347, "dur": 0.364,
    "args": {
      "External id": 2487210,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548588.175, "dur": 0.396,
    "args": {
      "External id": 2487211,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548590.134, "dur": 0.375,
    "args": {
      "External id": 2487212,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548604.170, "dur": 23.521,
    "args": {
      "External id": 2487213,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 4012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548680.439, "dur": 102.510,
    "args": {
      "External id": 2487214,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 4013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548701.228, "dur": 78.360,
    "args": {
      "External id": 2487215,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 4014, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548710.880, "dur": 64.983,
    "args": {
      "External id": 2487216,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 4015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548795.459, "dur": 1.959,
    "args": {
      "External id": 2487217,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 4016, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548872.476, "dur": 1594.549,
    "args": {
      "External id": 2487218,"Sequence number": 24740444, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 4017
    }
  },
  {
    "ph": "f", "id": 204, "pid": 1336759, "tid": 1381173, "ts": 1514657548872.476,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657548974.606, "dur": 146.647,
    "args": {
      "External id": 2487219,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [5632, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 4018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657549165.042, "dur": 39.151,
    "args": {
      "External id": 2487220,"kernel_hash": "cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/m5/cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [65536, 5632], [16, 4096, 5632], [16, 4096, 5632], [16, 4096, 5632], []], "Ev Idx": 4019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657549221.353, "dur": 52.846,
    "args": {
      "External id": 2487221,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 2048], [0, 5632, 1], [11534336, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632], [1, 2048, 5632]], "Ev Idx": 4020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657549283.943, "dur": 33.580,
    "args": {
      "External id": 2487222,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 4021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657549326.783, "dur": 46.549,
    "args": {
      "External id": 2487223,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 4022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657549381.067, "dur": 27.916,
    "args": {
      "External id": 2487224,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 4023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657549416.315, "dur": 42.674,
    "args": {
      "External id": 2487225,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 4024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657549480.211, "dur": 22.697,
    "args": {
      "External id": 2487226,"kernel_hash": "c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/47/c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 4025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657549523.823, "dur": 45.956,
    "args": {
      "External id": 2487227,"kernel_hash": "cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/xr/cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657549591.698, "dur": 18.627,
    "args": {
      "External id": 2487228,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 4027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657549623.762, "dur": 14.472,
    "args": {
      "External id": 2487229,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 4028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657549647.649, "dur": 30.121,
    "args": {
      "External id": 2487230,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 4029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657549681.290, "dur": 32.335,
    "args": {
      "External id": 2487231,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 4030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657549741.854, "dur": 165.805,
    "args": {
      "External id": 2487232,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 4031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657549818.281, "dur": 6.068,
    "args": {
      "External id": 2487233,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657549826.367, "dur": 2.132,
    "args": {
      "External id": 2487234,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657549937.491, "dur": 22.712,
    "args": {
      "External id": 2487235,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657549971.096, "dur": 53.235,
    "args": {
      "External id": 2487236,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550035.991, "dur": 39.980,
    "args": {
      "External id": 2487237,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 4036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550082.637, "dur": 35.332,
    "args": {
      "External id": 2487238,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 4037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550126.663, "dur": 20.350,
    "args": {
      "External id": 2487239,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 4038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550152.335, "dur": 29.215,
    "args": {
      "External id": 2487240,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 4039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550188.361, "dur": 19.928,
    "args": {
      "External id": 2487241,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 4040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550215.131, "dur": 29.755,
    "args": {
      "External id": 2487242,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 4041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550261.744, "dur": 22.021,
    "args": {
      "External id": 2487243,"kernel_hash": "cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/g4/cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [65536, 2048], []], "Ev Idx": 4042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550299.595, "dur": 24.395,
    "args": {
      "External id": 2487244,"kernel_hash": "c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/4n/c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550349.289, "dur": 25.267,
    "args": {
      "External id": 2487245,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 4044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550392.954, "dur": 14.432,
    "args": {
      "External id": 2487246,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 4045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550419.815, "dur": 15.772,
    "args": {
      "External id": 2487247,"kernel_hash": "cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u7/cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [8388608, 2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [16, 4096, 2048], []], "Ev Idx": 4046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550507.096, "dur": 18.331,
    "args": {
      "External id": 2487248,"Record function id": 0, "Ev Idx": 4047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550510.076, "dur": 14.549,
    "args": {
      "External id": 2487249,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550517.716, "dur": 5.855,
    "args": {
      "External id": 2487250,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550519.492, "dur": 3.944,
    "args": {
      "External id": 2487251,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550547.080, "dur": 7.961,
    "args": {
      "External id": 2487252,"Record function id": 0, "Ev Idx": 4051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550549.155, "dur": 5.181,
    "args": {
      "External id": 2487253,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550550.504, "dur": 2.943,
    "args": {
      "External id": 2487254,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550551.393, "dur": 1.815,
    "args": {
      "External id": 2487255,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550558.696, "dur": 6.526,
    "args": {
      "External id": 2487256,"Record function id": 0, "Ev Idx": 4055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550559.828, "dur": 4.987,
    "args": {
      "External id": 2487257,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550560.602, "dur": 3.751,
    "args": {
      "External id": 2487258,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550561.140, "dur": 3.148,
    "args": {
      "External id": 2487259,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550568.286, "dur": 3.938,
    "args": {
      "External id": 2487260,"Record function id": 0, "Ev Idx": 4059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550569.396, "dur": 2.408,
    "args": {
      "External id": 2487261,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 4060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550569.993, "dur": 1.395,
    "args": {
      "External id": 2487262,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 4061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550570.398, "dur": 0.929,
    "args": {
      "External id": 2487263,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 4062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550575.262, "dur": 4.385,
    "args": {
      "External id": 2487264,"Record function id": 0, "Ev Idx": 4063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550576.875, "dur": 2.367,
    "args": {
      "External id": 2487265,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550577.546, "dur": 1.283,
    "args": {
      "External id": 2487266,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550578.047, "dur": 0.715,
    "args": {
      "External id": 2487267,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550585.859, "dur": 4.160,
    "args": {
      "External id": 2487268,"Record function id": 0, "Ev Idx": 4067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550587.116, "dur": 2.470,
    "args": {
      "External id": 2487269,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550587.801, "dur": 1.369,
    "args": {
      "External id": 2487270,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550588.326, "dur": 0.730,
    "args": {
      "External id": 2487271,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550593.179, "dur": 3.489,
    "args": {
      "External id": 2487272,"Record function id": 0, "Ev Idx": 4071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550594.184, "dur": 2.089,
    "args": {
      "External id": 2487273,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550594.739, "dur": 1.119,
    "args": {
      "External id": 2487274,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550595.103, "dur": 0.672,
    "args": {
      "External id": 2487275,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550599.615, "dur": 3.444,
    "args": {
      "External id": 2487276,"Record function id": 0, "Ev Idx": 4075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550600.558, "dur": 2.107,
    "args": {
      "External id": 2487277,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550601.181, "dur": 1.063,
    "args": {
      "External id": 2487278,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550601.505, "dur": 0.673,
    "args": {
      "External id": 2487279,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550606.092, "dur": 3.611,
    "args": {
      "External id": 2487280,"Record function id": 0, "Ev Idx": 4079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550607.028, "dur": 2.283,
    "args": {
      "External id": 2487281,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550607.542, "dur": 1.348,
    "args": {
      "External id": 2487282,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550608.153, "dur": 0.650,
    "args": {
      "External id": 2487283,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550613.358, "dur": 35605.869,
    "args": {
      "External id": 2487284,"Record function id": 0, "Sequence number": 24740443, "Fwd thread id": 1, "Ev Idx": 4083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550614.515, "dur": 35596.765,
    "args": {
      "External id": 2487285,"Sequence number": 24740443, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 4084
    }
  },
  {
    "ph": "f", "id": 205, "pid": 1336759, "tid": 1381173, "ts": 1514657550614.515,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.11)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550647.246, "dur": 37.495,
    "args": {
      "External id": 2487286,"Record function id": 0, "Ev Idx": 4085
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.11)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550692.362, "dur": 63.795,
    "args": {
      "External id": 2487287,"Record function id": 0, "Ev Idx": 4086
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.11)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550765.163, "dur": 35437.691,
    "args": {
      "External id": 2487288,"Record function id": 0, "Ev Idx": 4087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550850.764, "dur": 6.605,
    "args": {
      "External id": 2487289,"Record function id": 0, "Concrete Inputs": ["[51384320]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550866.385, "dur": 6.937,
    "args": {
      "External id": 2487290,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 4089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550888.351, "dur": 34488.758,
    "args": {
      "External id": 2487291,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 4090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550901.072, "dur": 34466.674,
    "args": {
      "External id": 2487292,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 4091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657550967.468, "dur": 13.376,
    "args": {
      "External id": 2487293,"Record function id": 0, "Concrete Inputs": ["[12602]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657551023.387, "dur": 34305.053,
    "args": {
      "External id": 2487294,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], [], []], "Ev Idx": 4093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657551026.152, "dur": 34301.570,
    "args": {
      "External id": 2487295,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], []], "Ev Idx": 4094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657551030.308, "dur": 9.671,
    "args": {
      "External id": 2487296,"Record function id": 0, "Concrete Inputs": ["[12602]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657551060.664, "dur": 34263.722,
    "args": {
      "External id": 2487297,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[12602], [12602], []], "Ev Idx": 4096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657585468.279, "dur": 8.856,
    "args": {
      "External id": 2487298,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[51384320], [], [], [], [], []], "Ev Idx": 4097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657585471.372, "dur": 5.364,
    "args": {
      "External id": 2487299,"Record function id": 0, "Concrete Inputs": ["[6423040]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657585506.767, "dur": 364.293,
    "args": {
      "External id": 2487300,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[6423040], [51384320], [], [], [], []], "Ev Idx": 4099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657585543.326, "dur": 322.290,
    "args": {
      "External id": 2487301,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 6423040, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[51384320], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 4100, "In msg nelems": 51384320
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657585555.925, "dur": 304.545,
    "args": {
      "External id": 2487302,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[51384320]], "Ev Idx": 4101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657585890.489, "dur": 2.219,
    "args": {
      "External id": 2487303,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 4102, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657585950.580, "dur": 6.307,
    "args": {
      "External id": 2487304,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586034.011, "dur": 2.249,
    "args": {
      "External id": 2487305,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "256"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586053.965, "dur": 3.327,
    "args": {
      "External id": 2487306,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586070.032, "dur": 1.134,
    "args": {
      "External id": 2487307,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1048832"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586082.201, "dur": 0.985,
    "args": {
      "External id": 2487308,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1573120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586092.863, "dur": 0.823,
    "args": {
      "External id": 2487309,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586109.933, "dur": 2.610,
    "args": {
      "External id": 2487310,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "2097664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586124.853, "dur": 2.114,
    "args": {
      "External id": 2487311,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "3539456"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586137.219, "dur": 0.817,
    "args": {
      "External id": 2487312,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "4981248"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586235.155, "dur": 2647.225,
    "args": {
      "External id": 2487313,"Record function id": 0, "Ev Idx": 4112
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.10)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586254.457, "dur": 1013.716,
    "args": {
      "External id": 2487314,"Record function id": 0, "Ev Idx": 4113
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.10)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586270.788, "dur": 324.053,
    "args": {
      "External id": 2487315,"Record function id": 0, "Ev Idx": 4114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586346.172, "dur": 4.167,
    "args": {
      "External id": 2487316,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 4115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586353.388, "dur": 1.270,
    "args": {
      "External id": 2487317,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 4116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586356.151, "dur": 2.968,
    "args": {
      "External id": 2487318,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 4117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586360.389, "dur": 0.893,
    "args": {
      "External id": 2487319,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 4118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586362.527, "dur": 1.112,
    "args": {
      "External id": 2487320,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 4119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586365.217, "dur": 0.978,
    "args": {
      "External id": 2487321,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 4120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586367.744, "dur": 1.956,
    "args": {
      "External id": 2487322,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 4121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586370.969, "dur": 0.742,
    "args": {
      "External id": 2487323,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 4122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586372.990, "dur": 0.842,
    "args": {
      "External id": 2487324,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 4123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586375.318, "dur": 0.731,
    "args": {
      "External id": 2487325,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 4124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586393.475, "dur": 166.483,
    "args": {
      "External id": 2487326,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 4125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586408.477, "dur": 146.362,
    "args": {
      "External id": 2487327,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 4126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586426.722, "dur": 13.557,
    "args": {
      "External id": 2487328,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586443.393, "dur": 66.376,
    "args": {
      "External id": 2487329,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 4128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586445.973, "dur": 63.372,
    "args": {
      "External id": 2487330,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 4129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586449.863, "dur": 5.376,
    "args": {
      "External id": 2487331,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586456.988, "dur": 51.907,
    "args": {
      "External id": 2487332,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 4131
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.9", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586677.886, "dur": 582.510,
    "args": {
      "External id": 2487333,"Record function id": 0, "Ev Idx": 4132
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.9)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586698.858, "dur": 548.835,
    "args": {
      "External id": 2487334,"Record function id": 0, "Ev Idx": 4133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586758.332, "dur": 5.679,
    "args": {
      "External id": 2487335,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586779.918, "dur": 34.038,
    "args": {
      "External id": 2487336,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 4135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586784.950, "dur": 1.621,
    "args": {
      "External id": 2487337,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586788.724, "dur": 0.459,
    "args": {
      "External id": 2487338,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586790.957, "dur": 0.390,
    "args": {
      "External id": 2487339,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586792.819, "dur": 0.409,
    "args": {
      "External id": 2487340,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586794.992, "dur": 0.410,
    "args": {
      "External id": 2487341,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586796.972, "dur": 2.303,
    "args": {
      "External id": 2487342,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586805.306, "dur": 0.424,
    "args": {
      "External id": 2487343,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586807.333, "dur": 0.459,
    "args": {
      "External id": 2487344,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586809.075, "dur": 0.484,
    "args": {
      "External id": 2487345,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586823.622, "dur": 34.042,
    "args": {
      "External id": 2487346,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 4145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586888.166, "dur": 133.867,
    "args": {
      "External id": 2487347,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 4146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586898.171, "dur": 3.161,
    "args": {
      "External id": 2487348,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586906.098, "dur": 9.539,
    "args": {
      "External id": 2487349,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 4148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586910.290, "dur": 4.935,
    "args": {
      "External id": 2487350,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 4149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586913.563, "dur": 0.508,
    "args": {
      "External id": 2487351,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 4150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586925.862, "dur": 23.429,
    "args": {
      "External id": 2487352,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 4151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586928.144, "dur": 0.676,
    "args": {
      "External id": 2487353,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586930.407, "dur": 0.626,
    "args": {
      "External id": 2487354,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586932.551, "dur": 2.236,
    "args": {
      "External id": 2487355,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586936.374, "dur": 0.396,
    "args": {
      "External id": 2487356,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586938.282, "dur": 0.424,
    "args": {
      "External id": 2487357,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586940.094, "dur": 0.349,
    "args": {
      "External id": 2487358,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586941.973, "dur": 0.315,
    "args": {
      "External id": 2487359,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586943.776, "dur": 0.380,
    "args": {
      "External id": 2487360,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586945.603, "dur": 0.413,
    "args": {
      "External id": 2487361,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657586957.736, "dur": 20.299,
    "args": {
      "External id": 2487362,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 4161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657587071.333, "dur": 111.845,
    "args": {
      "External id": 2487363,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 4162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657587093.606, "dur": 86.070,
    "args": {
      "External id": 2487364,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 4163, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657587103.437, "dur": 72.298,
    "args": {
      "External id": 2487365,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 4164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657587197.193, "dur": 2.117,
    "args": {
      "External id": 2487366,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 4165, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657587275.555, "dur": 1588.980,
    "args": {
      "External id": 2487367,"Sequence number": 24740442, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 4166
    }
  },
  {
    "ph": "f", "id": 206, "pid": 1336759, "tid": 1381173, "ts": 1514657587275.555,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657587380.492, "dur": 100.269,
    "args": {
      "External id": 2487368,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [5632, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 4167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657587517.587, "dur": 59.331,
    "args": {
      "External id": 2487369,"kernel_hash": "cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/m5/cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [65536, 5632], [16, 4096, 5632], [16, 4096, 5632], [16, 4096, 5632], []], "Ev Idx": 4168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657587596.551, "dur": 54.894,
    "args": {
      "External id": 2487370,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 2048], [0, 5632, 1], [11534336, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632], [1, 2048, 5632]], "Ev Idx": 4169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657587663.221, "dur": 33.253,
    "args": {
      "External id": 2487371,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 4170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657587703.956, "dur": 45.834,
    "args": {
      "External id": 2487372,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 4171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657587757.381, "dur": 27.201,
    "args": {
      "External id": 2487373,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 4172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657587791.953, "dur": 41.751,
    "args": {
      "External id": 2487374,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 4173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657587858.177, "dur": 22.732,
    "args": {
      "External id": 2487375,"kernel_hash": "c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/47/c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 4174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657587909.924, "dur": 27.812,
    "args": {
      "External id": 2487376,"kernel_hash": "cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/xr/cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657587956.979, "dur": 18.715,
    "args": {
      "External id": 2487377,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 4176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588026.283, "dur": 17.693,
    "args": {
      "External id": 2487378,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 4177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588055.888, "dur": 33.414,
    "args": {
      "External id": 2487379,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 4178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588092.868, "dur": 34.894,
    "args": {
      "External id": 2487380,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 4179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588156.550, "dur": 171.448,
    "args": {
      "External id": 2487381,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 4180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588231.313, "dur": 5.569,
    "args": {
      "External id": 2487382,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588238.962, "dur": 2.417,
    "args": {
      "External id": 2487383,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588359.699, "dur": 22.607,
    "args": {
      "External id": 2487384,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588392.825, "dur": 14.841,
    "args": {
      "External id": 2487385,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588415.946, "dur": 32.349,
    "args": {
      "External id": 2487386,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 4185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588455.150, "dur": 33.978,
    "args": {
      "External id": 2487387,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 4186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588498.115, "dur": 20.304,
    "args": {
      "External id": 2487388,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 4187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588524.208, "dur": 48.799,
    "args": {
      "External id": 2487389,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 4188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588582.538, "dur": 21.111,
    "args": {
      "External id": 2487390,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 4189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588611.068, "dur": 30.142,
    "args": {
      "External id": 2487391,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 4190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588660.787, "dur": 36.783,
    "args": {
      "External id": 2487392,"kernel_hash": "cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/g4/cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [65536, 2048], []], "Ev Idx": 4191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588721.607, "dur": 24.842,
    "args": {
      "External id": 2487393,"kernel_hash": "c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/4n/c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588760.848, "dur": 16.634,
    "args": {
      "External id": 2487394,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 4193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588790.883, "dur": 14.542,
    "args": {
      "External id": 2487395,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 4194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588817.362, "dur": 16.251,
    "args": {
      "External id": 2487396,"kernel_hash": "cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u7/cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [8388608, 2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [16, 4096, 2048], []], "Ev Idx": 4195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588907.056, "dur": 14.923,
    "args": {
      "External id": 2487397,"Record function id": 0, "Ev Idx": 4196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588910.138, "dur": 10.796,
    "args": {
      "External id": 2487398,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588913.919, "dur": 6.033,
    "args": {
      "External id": 2487399,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588915.950, "dur": 3.853,
    "args": {
      "External id": 2487400,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588925.725, "dur": 7.077,
    "args": {
      "External id": 2487401,"Record function id": 0, "Ev Idx": 4200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588926.866, "dur": 5.506,
    "args": {
      "External id": 2487402,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588927.795, "dur": 4.072,
    "args": {
      "External id": 2487403,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588928.561, "dur": 3.245,
    "args": {
      "External id": 2487404,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588935.981, "dur": 4.252,
    "args": {
      "External id": 2487405,"Record function id": 0, "Ev Idx": 4204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588937.137, "dur": 2.691,
    "args": {
      "External id": 2487406,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588937.739, "dur": 1.658,
    "args": {
      "External id": 2487407,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588938.140, "dur": 1.192,
    "args": {
      "External id": 2487408,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588943.382, "dur": 4.068,
    "args": {
      "External id": 2487409,"Record function id": 0, "Ev Idx": 4208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588944.911, "dur": 2.122,
    "args": {
      "External id": 2487410,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 4209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588945.668, "dur": 0.977,
    "args": {
      "External id": 2487411,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 4210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588945.957, "dur": 0.620,
    "args": {
      "External id": 2487412,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 4211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588950.498, "dur": 3.287,
    "args": {
      "External id": 2487413,"Record function id": 0, "Ev Idx": 4212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588951.502, "dur": 1.865,
    "args": {
      "External id": 2487414,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588951.982, "dur": 0.983,
    "args": {
      "External id": 2487415,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588952.254, "dur": 0.644,
    "args": {
      "External id": 2487416,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588957.072, "dur": 7.296,
    "args": {
      "External id": 2487417,"Record function id": 0, "Ev Idx": 4216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588958.089, "dur": 5.831,
    "args": {
      "External id": 2487418,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588958.671, "dur": 4.813,
    "args": {
      "External id": 2487419,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588959.270, "dur": 4.131,
    "args": {
      "External id": 2487420,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588967.772, "dur": 3.877,
    "args": {
      "External id": 2487421,"Record function id": 0, "Ev Idx": 4220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588968.984, "dur": 2.247,
    "args": {
      "External id": 2487422,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588969.467, "dur": 1.360,
    "args": {
      "External id": 2487423,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588969.957, "dur": 0.769,
    "args": {
      "External id": 2487424,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588974.748, "dur": 3.706,
    "args": {
      "External id": 2487425,"Record function id": 0, "Ev Idx": 4224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588975.766, "dur": 2.266,
    "args": {
      "External id": 2487426,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588976.422, "dur": 1.156,
    "args": {
      "External id": 2487427,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657588976.843, "dur": 0.635,
    "args": {
      "External id": 2487428,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657589024.529, "dur": 9.284,
    "args": {
      "External id": 2487429,"Record function id": 0, "Ev Idx": 4228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657589026.668, "dur": 6.411,
    "args": {
      "External id": 2487430,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657589027.868, "dur": 4.307,
    "args": {
      "External id": 2487431,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657589028.881, "dur": 3.035,
    "args": {
      "External id": 2487432,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657589037.878, "dur": 36037.589,
    "args": {
      "External id": 2487433,"Record function id": 0, "Sequence number": 24740441, "Fwd thread id": 1, "Ev Idx": 4232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657589039.218, "dur": 36026.765,
    "args": {
      "External id": 2487434,"Sequence number": 24740441, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 4233
    }
  },
  {
    "ph": "f", "id": 207, "pid": 1336759, "tid": 1381173, "ts": 1514657589039.218,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.10)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657589069.841, "dur": 36.292,
    "args": {
      "External id": 2487435,"Record function id": 0, "Ev Idx": 4234
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.10)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657589113.690, "dur": 61.053,
    "args": {
      "External id": 2487436,"Record function id": 0, "Ev Idx": 4235
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.10)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657589184.339, "dur": 35873.075,
    "args": {
      "External id": 2487437,"Record function id": 0, "Ev Idx": 4236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657589274.722, "dur": 6.800,
    "args": {
      "External id": 2487438,"Record function id": 0, "Concrete Inputs": ["[51384320]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657589290.466, "dur": 4.870,
    "args": {
      "External id": 2487439,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 4238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657589309.439, "dur": 34869.168,
    "args": {
      "External id": 2487440,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 4239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657589321.856, "dur": 34847.617,
    "args": {
      "External id": 2487441,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 4240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657589382.103, "dur": 13.064,
    "args": {
      "External id": 2487442,"Record function id": 0, "Concrete Inputs": ["[12602]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657589401.557, "dur": 34728.599,
    "args": {
      "External id": 2487443,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], [], []], "Ev Idx": 4242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657589404.369, "dur": 34725.137,
    "args": {
      "External id": 2487444,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], []], "Ev Idx": 4243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657589408.608, "dur": 5.029,
    "args": {
      "External id": 2487445,"Record function id": 0, "Concrete Inputs": ["[12602]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657589415.469, "dur": 34711.022,
    "args": {
      "External id": 2487446,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[12602], [12602], []], "Ev Idx": 4245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657624265.895, "dur": 8.682,
    "args": {
      "External id": 2487447,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[51384320], [], [], [], [], []], "Ev Idx": 4246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657624269.080, "dur": 5.147,
    "args": {
      "External id": 2487448,"Record function id": 0, "Concrete Inputs": ["[6423040]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657624304.066, "dur": 423.243,
    "args": {
      "External id": 2487449,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[6423040], [51384320], [], [], [], []], "Ev Idx": 4248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657624330.638, "dur": 390.766,
    "args": {
      "External id": 2487450,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 6423040, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[51384320], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 4249, "In msg nelems": 51384320
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657624341.830, "dur": 373.038,
    "args": {
      "External id": 2487451,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[51384320]], "Ev Idx": 4250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657624749.607, "dur": 2.745,
    "args": {
      "External id": 2487452,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 4251, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657624813.566, "dur": 6.210,
    "args": {
      "External id": 2487453,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657624864.078, "dur": 3.959,
    "args": {
      "External id": 2487454,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "256"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657624883.984, "dur": 1.246,
    "args": {
      "External id": 2487455,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657624897.104, "dur": 1.039,
    "args": {
      "External id": 2487456,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1048832"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657624909.028, "dur": 1.157,
    "args": {
      "External id": 2487457,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1573120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657624921.279, "dur": 3.047,
    "args": {
      "External id": 2487458,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657624935.589, "dur": 0.826,
    "args": {
      "External id": 2487459,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "2097664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657624948.186, "dur": 1.892,
    "args": {
      "External id": 2487460,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "3539456"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657624959.703, "dur": 0.892,
    "args": {
      "External id": 2487461,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "4981248"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625091.999, "dur": 2632.070,
    "args": {
      "External id": 2487462,"Record function id": 0, "Ev Idx": 4261
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.9)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625110.397, "dur": 996.310,
    "args": {
      "External id": 2487463,"Record function id": 0, "Ev Idx": 4262
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.9)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625126.198, "dur": 306.458,
    "args": {
      "External id": 2487464,"Record function id": 0, "Ev Idx": 4263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625202.846, "dur": 6.369,
    "args": {
      "External id": 2487465,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 4264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625212.405, "dur": 1.077,
    "args": {
      "External id": 2487466,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 4265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625215.160, "dur": 0.940,
    "args": {
      "External id": 2487467,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 4266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625217.614, "dur": 0.866,
    "args": {
      "External id": 2487468,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 4267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625219.770, "dur": 0.903,
    "args": {
      "External id": 2487469,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 4268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625222.290, "dur": 0.699,
    "args": {
      "External id": 2487470,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 4269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625224.393, "dur": 2.096,
    "args": {
      "External id": 2487471,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 4270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625227.638, "dur": 1.040,
    "args": {
      "External id": 2487472,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 4271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625230.686, "dur": 2.728,
    "args": {
      "External id": 2487473,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 4272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625234.976, "dur": 0.709,
    "args": {
      "External id": 2487474,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 4273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625253.591, "dur": 152.010,
    "args": {
      "External id": 2487475,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 4274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625269.350, "dur": 131.721,
    "args": {
      "External id": 2487476,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 4275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625291.430, "dur": 12.145,
    "args": {
      "External id": 2487477,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625306.814, "dur": 68.613,
    "args": {
      "External id": 2487478,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 4277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625309.749, "dur": 65.303,
    "args": {
      "External id": 2487479,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 4278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625313.968, "dur": 5.510,
    "args": {
      "External id": 2487480,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625323.519, "dur": 50.916,
    "args": {
      "External id": 2487481,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 4280
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.8", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625507.335, "dur": 591.235,
    "args": {
      "External id": 2487482,"Record function id": 0, "Ev Idx": 4281
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.8)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625546.473, "dur": 539.588,
    "args": {
      "External id": 2487483,"Record function id": 0, "Ev Idx": 4282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625609.470, "dur": 5.948,
    "args": {
      "External id": 2487484,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625630.746, "dur": 29.096,
    "args": {
      "External id": 2487485,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 4284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625635.689, "dur": 1.779,
    "args": {
      "External id": 2487486,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625639.515, "dur": 0.514,
    "args": {
      "External id": 2487487,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625641.497, "dur": 0.540,
    "args": {
      "External id": 2487488,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625643.801, "dur": 2.077,
    "args": {
      "External id": 2487489,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625647.539, "dur": 0.695,
    "args": {
      "External id": 2487490,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625649.793, "dur": 0.415,
    "args": {
      "External id": 2487491,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625651.783, "dur": 0.368,
    "args": {
      "External id": 2487492,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625653.614, "dur": 0.411,
    "args": {
      "External id": 2487493,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625655.434, "dur": 0.380,
    "args": {
      "External id": 2487494,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625669.224, "dur": 33.434,
    "args": {
      "External id": 2487495,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 4294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625733.050, "dur": 97.301,
    "args": {
      "External id": 2487496,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 4295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625743.153, "dur": 2.860,
    "args": {
      "External id": 2487497,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625750.937, "dur": 9.976,
    "args": {
      "External id": 2487498,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 4297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625755.132, "dur": 5.362,
    "args": {
      "External id": 2487499,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 4298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625758.950, "dur": 0.411,
    "args": {
      "External id": 2487500,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 4299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625767.558, "dur": 24.412,
    "args": {
      "External id": 2487501,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 4300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625769.847, "dur": 1.919,
    "args": {
      "External id": 2487502,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625773.291, "dur": 0.688,
    "args": {
      "External id": 2487503,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625775.452, "dur": 0.425,
    "args": {
      "External id": 2487504,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625777.490, "dur": 0.578,
    "args": {
      "External id": 2487505,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625779.454, "dur": 0.502,
    "args": {
      "External id": 2487506,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625781.573, "dur": 0.653,
    "args": {
      "External id": 2487507,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625783.971, "dur": 0.440,
    "args": {
      "External id": 2487508,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625785.732, "dur": 0.428,
    "args": {
      "External id": 2487509,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625787.358, "dur": 1.989,
    "args": {
      "External id": 2487510,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625801.651, "dur": 21.259,
    "args": {
      "External id": 2487511,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 4310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625872.754, "dur": 101.522,
    "args": {
      "External id": 2487512,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 4311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625894.145, "dur": 76.782,
    "args": {
      "External id": 2487513,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 4312, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657625902.945, "dur": 64.374,
    "args": {
      "External id": 2487514,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 4313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657626029.873, "dur": 2.656,
    "args": {
      "External id": 2487515,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 4314, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657626113.670, "dur": 1589.843,
    "args": {
      "External id": 2487516,"Sequence number": 24740440, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 4315
    }
  },
  {
    "ph": "f", "id": 208, "pid": 1336759, "tid": 1381173, "ts": 1514657626113.670,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657626217.476, "dur": 105.236,
    "args": {
      "External id": 2487517,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [5632, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 4316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657626361.175, "dur": 37.451,
    "args": {
      "External id": 2487518,"kernel_hash": "cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/m5/cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [65536, 5632], [16, 4096, 5632], [16, 4096, 5632], [16, 4096, 5632], []], "Ev Idx": 4317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657626415.812, "dur": 46.049,
    "args": {
      "External id": 2487519,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 2048], [0, 5632, 1], [11534336, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632], [1, 2048, 5632]], "Ev Idx": 4318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657626471.292, "dur": 32.274,
    "args": {
      "External id": 2487520,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 4319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657626509.790, "dur": 64.402,
    "args": {
      "External id": 2487521,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 4320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657626584.607, "dur": 29.744,
    "args": {
      "External id": 2487522,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 4321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657626624.297, "dur": 40.548,
    "args": {
      "External id": 2487523,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 4322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657626689.707, "dur": 22.831,
    "args": {
      "External id": 2487524,"kernel_hash": "c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/47/c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 4323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657626730.231, "dur": 25.406,
    "args": {
      "External id": 2487525,"kernel_hash": "cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/xr/cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657626775.757, "dur": 18.286,
    "args": {
      "External id": 2487526,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 4325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657626808.048, "dur": 15.493,
    "args": {
      "External id": 2487527,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 4326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657626834.586, "dur": 27.853,
    "args": {
      "External id": 2487528,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 4327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657626866.225, "dur": 32.848,
    "args": {
      "External id": 2487529,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 4328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657626925.536, "dur": 208.667,
    "args": {
      "External id": 2487530,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 4329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627037.973, "dur": 6.293,
    "args": {
      "External id": 2487531,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627046.678, "dur": 2.069,
    "args": {
      "External id": 2487532,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627165.628, "dur": 26.979,
    "args": {
      "External id": 2487533,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627203.052, "dur": 13.702,
    "args": {
      "External id": 2487534,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627225.488, "dur": 41.537,
    "args": {
      "External id": 2487535,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 4334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627276.082, "dur": 37.301,
    "args": {
      "External id": 2487536,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 4335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627320.513, "dur": 21.147,
    "args": {
      "External id": 2487537,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 4336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627355.175, "dur": 31.249,
    "args": {
      "External id": 2487538,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 4337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627393.373, "dur": 20.184,
    "args": {
      "External id": 2487539,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 4338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627422.122, "dur": 29.960,
    "args": {
      "External id": 2487540,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 4339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627468.220, "dur": 20.256,
    "args": {
      "External id": 2487541,"kernel_hash": "cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/g4/cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [65536, 2048], []], "Ev Idx": 4340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627503.445, "dur": 21.839,
    "args": {
      "External id": 2487542,"kernel_hash": "c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/4n/c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627572.876, "dur": 29.134,
    "args": {
      "External id": 2487543,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 4342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627622.454, "dur": 18.631,
    "args": {
      "External id": 2487544,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 4343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627654.520, "dur": 15.929,
    "args": {
      "External id": 2487545,"kernel_hash": "cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u7/cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [8388608, 2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [16, 4096, 2048], []], "Ev Idx": 4344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627746.363, "dur": 20.483,
    "args": {
      "External id": 2487546,"Record function id": 0, "Ev Idx": 4345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627749.528, "dur": 16.232,
    "args": {
      "External id": 2487547,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627753.713, "dur": 11.172,
    "args": {
      "External id": 2487548,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627755.795, "dur": 8.946,
    "args": {
      "External id": 2487549,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627770.719, "dur": 5.475,
    "args": {
      "External id": 2487550,"Record function id": 0, "Ev Idx": 4349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627771.993, "dur": 3.783,
    "args": {
      "External id": 2487551,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627772.973, "dur": 2.331,
    "args": {
      "External id": 2487552,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627774.001, "dur": 1.187,
    "args": {
      "External id": 2487553,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627779.494, "dur": 5.056,
    "args": {
      "External id": 2487554,"Record function id": 0, "Ev Idx": 4353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627781.368, "dur": 2.792,
    "args": {
      "External id": 2487555,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627782.090, "dur": 1.670,
    "args": {
      "External id": 2487556,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627782.548, "dur": 1.108,
    "args": {
      "External id": 2487557,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627787.794, "dur": 3.820,
    "args": {
      "External id": 2487558,"Record function id": 0, "Ev Idx": 4357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627788.858, "dur": 2.364,
    "args": {
      "External id": 2487559,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 4358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627789.541, "dur": 1.326,
    "args": {
      "External id": 2487560,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 4359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627790.074, "dur": 0.693,
    "args": {
      "External id": 2487561,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 4360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627794.747, "dur": 3.663,
    "args": {
      "External id": 2487562,"Record function id": 0, "Ev Idx": 4361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627795.750, "dur": 2.272,
    "args": {
      "External id": 2487563,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627796.392, "dur": 1.218,
    "args": {
      "External id": 2487564,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627796.817, "dur": 0.731,
    "args": {
      "External id": 2487565,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627801.458, "dur": 3.947,
    "args": {
      "External id": 2487566,"Record function id": 0, "Ev Idx": 4365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627802.454, "dur": 2.493,
    "args": {
      "External id": 2487567,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627802.951, "dur": 1.595,
    "args": {
      "External id": 2487568,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627803.680, "dur": 0.748,
    "args": {
      "External id": 2487569,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627808.549, "dur": 3.215,
    "args": {
      "External id": 2487570,"Record function id": 0, "Ev Idx": 4369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627809.483, "dur": 1.878,
    "args": {
      "External id": 2487571,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627809.998, "dur": 0.952,
    "args": {
      "External id": 2487572,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627810.268, "dur": 0.620,
    "args": {
      "External id": 2487573,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627814.854, "dur": 9.040,
    "args": {
      "External id": 2487574,"Record function id": 0, "Ev Idx": 4373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627815.886, "dur": 4.329,
    "args": {
      "External id": 2487575,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627816.486, "dur": 3.337,
    "args": {
      "External id": 2487576,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627816.757, "dur": 2.980,
    "args": {
      "External id": 2487577,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627826.873, "dur": 3.659,
    "args": {
      "External id": 2487578,"Record function id": 0, "Ev Idx": 4377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627827.892, "dur": 2.235,
    "args": {
      "External id": 2487579,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627828.369, "dur": 1.330,
    "args": {
      "External id": 2487580,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627828.829, "dur": 0.752,
    "args": {
      "External id": 2487581,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627833.769, "dur": 35121.887,
    "args": {
      "External id": 2487582,"Record function id": 0, "Sequence number": 24740439, "Fwd thread id": 1, "Ev Idx": 4381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627834.905, "dur": 35112.883,
    "args": {
      "External id": 2487583,"Sequence number": 24740439, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 4382
    }
  },
  {
    "ph": "f", "id": 209, "pid": 1336759, "tid": 1381173, "ts": 1514657627834.905,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.9)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627861.814, "dur": 37.722,
    "args": {
      "External id": 2487584,"Record function id": 0, "Ev Idx": 4383
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.9)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627907.093, "dur": 59.731,
    "args": {
      "External id": 2487585,"Record function id": 0, "Ev Idx": 4384
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.9)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657627972.596, "dur": 34967.432,
    "args": {
      "External id": 2487586,"Record function id": 0, "Ev Idx": 4385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657628104.405, "dur": 6.790,
    "args": {
      "External id": 2487587,"Record function id": 0, "Concrete Inputs": ["[51384320]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657628121.752, "dur": 8.246,
    "args": {
      "External id": 2487588,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 4387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657628145.130, "dur": 34022.275,
    "args": {
      "External id": 2487589,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 4388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657628158.861, "dur": 33999.705,
    "args": {
      "External id": 2487590,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 4389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657628211.015, "dur": 13.579,
    "args": {
      "External id": 2487591,"Record function id": 0, "Concrete Inputs": ["[12602]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657628230.564, "dur": 33891.307,
    "args": {
      "External id": 2487592,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], [], []], "Ev Idx": 4391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657628233.439, "dur": 33887.731,
    "args": {
      "External id": 2487593,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], []], "Ev Idx": 4392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657628237.381, "dur": 5.222,
    "args": {
      "External id": 2487594,"Record function id": 0, "Concrete Inputs": ["[12602]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657628244.319, "dur": 33873.182,
    "args": {
      "External id": 2487595,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[12602], [12602], []], "Ev Idx": 4394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657662254.657, "dur": 9.888,
    "args": {
      "External id": 2487596,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[51384320], [], [], [], [], []], "Ev Idx": 4395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657662258.881, "dur": 5.297,
    "args": {
      "External id": 2487597,"Record function id": 0, "Concrete Inputs": ["[6423040]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657662291.837, "dur": 357.210,
    "args": {
      "External id": 2487598,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[6423040], [51384320], [], [], [], []], "Ev Idx": 4397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657662318.877, "dur": 324.835,
    "args": {
      "External id": 2487599,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 6423040, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[51384320], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 4398, "In msg nelems": 51384320
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657662329.855, "dur": 308.504,
    "args": {
      "External id": 2487600,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[51384320]], "Ev Idx": 4399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657662673.037, "dur": 2.479,
    "args": {
      "External id": 2487601,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 4400, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657662733.514, "dur": 6.456,
    "args": {
      "External id": 2487602,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657662783.796, "dur": 1.296,
    "args": {
      "External id": 2487603,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "256"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657662800.356, "dur": 1.166,
    "args": {
      "External id": 2487604,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657662812.919, "dur": 0.823,
    "args": {
      "External id": 2487605,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1048832"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657662826.876, "dur": 0.757,
    "args": {
      "External id": 2487606,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1573120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657662839.416, "dur": 0.897,
    "args": {
      "External id": 2487607,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657662851.040, "dur": 0.933,
    "args": {
      "External id": 2487608,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "2097664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657662862.836, "dur": 1.941,
    "args": {
      "External id": 2487609,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "3539456"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657662874.936, "dur": 0.687,
    "args": {
      "External id": 2487610,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "4981248"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657662969.907, "dur": 2669.162,
    "args": {
      "External id": 2487611,"Record function id": 0, "Ev Idx": 4410
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.8)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663024.556, "dur": 995.805,
    "args": {
      "External id": 2487612,"Record function id": 0, "Ev Idx": 4411
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.8)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663040.405, "dur": 308.305,
    "args": {
      "External id": 2487613,"Record function id": 0, "Ev Idx": 4412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663120.017, "dur": 4.610,
    "args": {
      "External id": 2487614,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 4413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663128.069, "dur": 0.965,
    "args": {
      "External id": 2487615,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 4414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663133.280, "dur": 1.131,
    "args": {
      "External id": 2487616,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 4415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663136.278, "dur": 0.947,
    "args": {
      "External id": 2487617,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 4416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663138.857, "dur": 0.870,
    "args": {
      "External id": 2487618,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 4417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663141.388, "dur": 0.490,
    "args": {
      "External id": 2487619,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 4418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663145.698, "dur": 1.573,
    "args": {
      "External id": 2487620,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 4419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663149.278, "dur": 2.824,
    "args": {
      "External id": 2487621,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 4420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663153.656, "dur": 1.083,
    "args": {
      "External id": 2487622,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 4421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663156.226, "dur": 0.938,
    "args": {
      "External id": 2487623,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 4422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663177.059, "dur": 142.811,
    "args": {
      "External id": 2487624,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 4423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663193.155, "dur": 122.468,
    "args": {
      "External id": 2487625,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 4424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663211.930, "dur": 11.821,
    "args": {
      "External id": 2487626,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663226.722, "dur": 62.769,
    "args": {
      "External id": 2487627,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 4426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663229.078, "dur": 60.072,
    "args": {
      "External id": 2487628,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 4427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663233.234, "dur": 5.720,
    "args": {
      "External id": 2487629,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663240.576, "dur": 47.945,
    "args": {
      "External id": 2487630,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 4429
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.7", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663424.491, "dur": 550.432,
    "args": {
      "External id": 2487631,"Record function id": 0, "Ev Idx": 4430
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.7)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663439.606, "dur": 522.947,
    "args": {
      "External id": 2487632,"Record function id": 0, "Ev Idx": 4431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663491.035, "dur": 4.757,
    "args": {
      "External id": 2487633,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663510.480, "dur": 48.760,
    "args": {
      "External id": 2487634,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 4433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663518.667, "dur": 2.350,
    "args": {
      "External id": 2487635,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663522.635, "dur": 0.731,
    "args": {
      "External id": 2487636,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663524.647, "dur": 2.275,
    "args": {
      "External id": 2487637,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663545.056, "dur": 0.637,
    "args": {
      "External id": 2487638,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663547.625, "dur": 0.577,
    "args": {
      "External id": 2487639,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663549.365, "dur": 0.634,
    "args": {
      "External id": 2487640,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663551.560, "dur": 0.471,
    "args": {
      "External id": 2487641,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663553.465, "dur": 0.612,
    "args": {
      "External id": 2487642,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663555.382, "dur": 0.318,
    "args": {
      "External id": 2487643,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663569.546, "dur": 34.773,
    "args": {
      "External id": 2487644,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 4443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663637.409, "dur": 106.166,
    "args": {
      "External id": 2487645,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 4444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663647.782, "dur": 3.830,
    "args": {
      "External id": 2487646,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663656.625, "dur": 11.812,
    "args": {
      "External id": 2487647,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 4446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663660.802, "dur": 7.226,
    "args": {
      "External id": 2487648,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 4447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663664.292, "dur": 2.576,
    "args": {
      "External id": 2487649,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 4448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663676.056, "dur": 24.077,
    "args": {
      "External id": 2487650,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 4449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663678.172, "dur": 0.375,
    "args": {
      "External id": 2487651,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663680.367, "dur": 0.536,
    "args": {
      "External id": 2487652,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663682.427, "dur": 0.520,
    "args": {
      "External id": 2487653,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663684.509, "dur": 0.489,
    "args": {
      "External id": 2487654,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663686.386, "dur": 0.395,
    "args": {
      "External id": 2487655,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663688.396, "dur": 0.560,
    "args": {
      "External id": 2487656,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663690.156, "dur": 0.520,
    "args": {
      "External id": 2487657,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663692.006, "dur": 2.510,
    "args": {
      "External id": 2487658,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663695.808, "dur": 0.411,
    "args": {
      "External id": 2487659,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663715.576, "dur": 20.479,
    "args": {
      "External id": 2487660,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 4459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663787.289, "dur": 110.064,
    "args": {
      "External id": 2487661,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 4460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663815.506, "dur": 78.442,
    "args": {
      "External id": 2487662,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 4461, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663825.047, "dur": 65.156,
    "args": {
      "External id": 2487663,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 4462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657663913.799, "dur": 1.980,
    "args": {
      "External id": 2487664,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 4463, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657664029.443, "dur": 1590.509,
    "args": {
      "External id": 2487665,"Sequence number": 24740438, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 4464
    }
  },
  {
    "ph": "f", "id": 210, "pid": 1336759, "tid": 1381173, "ts": 1514657664029.443,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657664135.147, "dur": 104.230,
    "args": {
      "External id": 2487666,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [5632, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 4465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657664281.008, "dur": 36.591,
    "args": {
      "External id": 2487667,"kernel_hash": "cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/m5/cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [65536, 5632], [16, 4096, 5632], [16, 4096, 5632], [16, 4096, 5632], []], "Ev Idx": 4466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657664334.831, "dur": 49.195,
    "args": {
      "External id": 2487668,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 2048], [0, 5632, 1], [11534336, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632], [1, 2048, 5632]], "Ev Idx": 4467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657664393.746, "dur": 32.070,
    "args": {
      "External id": 2487669,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 4468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657664432.491, "dur": 44.654,
    "args": {
      "External id": 2487670,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 4469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657664484.063, "dur": 26.787,
    "args": {
      "External id": 2487671,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 4470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657664520.084, "dur": 61.071,
    "args": {
      "External id": 2487672,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 4471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657664609.266, "dur": 24.348,
    "args": {
      "External id": 2487673,"kernel_hash": "c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/47/c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 4472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657664654.493, "dur": 27.262,
    "args": {
      "External id": 2487674,"kernel_hash": "cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/xr/cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657664701.766, "dur": 17.588,
    "args": {
      "External id": 2487675,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 4474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657664735.207, "dur": 14.274,
    "args": {
      "External id": 2487676,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 4475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657664759.237, "dur": 30.477,
    "args": {
      "External id": 2487677,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 4476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657664792.843, "dur": 32.854,
    "args": {
      "External id": 2487678,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 4477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657664852.374, "dur": 193.158,
    "args": {
      "External id": 2487679,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 4478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657664923.493, "dur": 5.918,
    "args": {
      "External id": 2487680,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657664931.099, "dur": 2.423,
    "args": {
      "External id": 2487681,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665081.872, "dur": 26.152,
    "args": {
      "External id": 2487682,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665120.148, "dur": 14.269,
    "args": {
      "External id": 2487683,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665145.028, "dur": 43.774,
    "args": {
      "External id": 2487684,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 4483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665194.258, "dur": 35.246,
    "args": {
      "External id": 2487685,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 4484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665239.433, "dur": 19.837,
    "args": {
      "External id": 2487686,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 4485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665263.765, "dur": 29.804,
    "args": {
      "External id": 2487687,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 4486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665304.520, "dur": 19.523,
    "args": {
      "External id": 2487688,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 4487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665331.087, "dur": 28.938,
    "args": {
      "External id": 2487689,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 4488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665379.539, "dur": 20.448,
    "args": {
      "External id": 2487690,"kernel_hash": "cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/g4/cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [65536, 2048], []], "Ev Idx": 4489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665418.657, "dur": 24.642,
    "args": {
      "External id": 2487691,"kernel_hash": "c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/4n/c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665466.105, "dur": 23.184,
    "args": {
      "External id": 2487692,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 4491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665515.259, "dur": 31.465,
    "args": {
      "External id": 2487693,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 4492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665569.765, "dur": 17.763,
    "args": {
      "External id": 2487694,"kernel_hash": "cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u7/cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [8388608, 2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [16, 4096, 2048], []], "Ev Idx": 4493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665661.339, "dur": 20.712,
    "args": {
      "External id": 2487695,"Record function id": 0, "Ev Idx": 4494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665664.493, "dur": 16.596,
    "args": {
      "External id": 2487696,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665668.755, "dur": 11.280,
    "args": {
      "External id": 2487697,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665670.258, "dur": 9.669,
    "args": {
      "External id": 2487698,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665685.821, "dur": 4.462,
    "args": {
      "External id": 2487699,"Record function id": 0, "Ev Idx": 4498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665686.982, "dur": 2.861,
    "args": {
      "External id": 2487700,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665687.775, "dur": 1.591,
    "args": {
      "External id": 2487701,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665688.430, "dur": 0.869,
    "args": {
      "External id": 2487702,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665693.510, "dur": 4.118,
    "args": {
      "External id": 2487703,"Record function id": 0, "Ev Idx": 4502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665694.772, "dur": 2.469,
    "args": {
      "External id": 2487704,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665695.463, "dur": 1.317,
    "args": {
      "External id": 2487705,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665695.932, "dur": 0.785,
    "args": {
      "External id": 2487706,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665700.747, "dur": 4.281,
    "args": {
      "External id": 2487707,"Record function id": 0, "Ev Idx": 4506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665701.913, "dur": 2.693,
    "args": {
      "External id": 2487708,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 4507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665702.798, "dur": 1.364,
    "args": {
      "External id": 2487709,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 4508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665703.329, "dur": 0.759,
    "args": {
      "External id": 2487710,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 4509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665708.074, "dur": 3.557,
    "args": {
      "External id": 2487711,"Record function id": 0, "Ev Idx": 4510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665709.121, "dur": 2.136,
    "args": {
      "External id": 2487712,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665709.829, "dur": 1.012,
    "args": {
      "External id": 2487713,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665710.184, "dur": 0.577,
    "args": {
      "External id": 2487714,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665714.633, "dur": 3.404,
    "args": {
      "External id": 2487715,"Record function id": 0, "Ev Idx": 4514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665715.809, "dur": 1.826,
    "args": {
      "External id": 2487716,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665716.300, "dur": 0.909,
    "args": {
      "External id": 2487717,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665716.577, "dur": 0.555,
    "args": {
      "External id": 2487718,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665721.179, "dur": 4.914,
    "args": {
      "External id": 2487719,"Record function id": 0, "Ev Idx": 4518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665722.071, "dur": 3.594,
    "args": {
      "External id": 2487720,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665722.551, "dur": 2.676,
    "args": {
      "External id": 2487721,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665724.585, "dur": 0.541,
    "args": {
      "External id": 2487722,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665729.186, "dur": 5.044,
    "args": {
      "External id": 2487723,"Record function id": 0, "Ev Idx": 4522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665730.267, "dur": 3.570,
    "args": {
      "External id": 2487724,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665730.750, "dur": 2.670,
    "args": {
      "External id": 2487725,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665731.006, "dur": 2.344,
    "args": {
      "External id": 2487726,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665740.497, "dur": 3.514,
    "args": {
      "External id": 2487727,"Record function id": 0, "Ev Idx": 4526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665741.438, "dur": 2.186,
    "args": {
      "External id": 2487728,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665741.903, "dur": 1.326,
    "args": {
      "External id": 2487729,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665742.168, "dur": 0.980,
    "args": {
      "External id": 2487730,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665747.361, "dur": 36528.935,
    "args": {
      "External id": 2487731,"Record function id": 0, "Sequence number": 24740437, "Fwd thread id": 1, "Ev Idx": 4530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665748.681, "dur": 36519.375,
    "args": {
      "External id": 2487732,"Sequence number": 24740437, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 4531
    }
  },
  {
    "ph": "f", "id": 211, "pid": 1336759, "tid": 1381173, "ts": 1514657665748.681,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.8)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665776.734, "dur": 40.242,
    "args": {
      "External id": 2487733,"Record function id": 0, "Ev Idx": 4532
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.8)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665824.795, "dur": 64.574,
    "args": {
      "External id": 2487734,"Record function id": 0, "Ev Idx": 4533
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.8)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657665895.430, "dur": 36364.465,
    "args": {
      "External id": 2487735,"Record function id": 0, "Ev Idx": 4534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657666020.232, "dur": 8.584,
    "args": {
      "External id": 2487736,"Record function id": 0, "Concrete Inputs": ["[51384320]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657666044.407, "dur": 5.221,
    "args": {
      "External id": 2487737,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 4536
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657666064.797, "dur": 35340.530,
    "args": {
      "External id": 2487738,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 4537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657666078.240, "dur": 35318.111,
    "args": {
      "External id": 2487739,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 4538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657666130.757, "dur": 13.507,
    "args": {
      "External id": 2487740,"Record function id": 0, "Concrete Inputs": ["[12602]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657666150.636, "dur": 35209.013,
    "args": {
      "External id": 2487741,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], [], []], "Ev Idx": 4540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657666153.288, "dur": 35205.599,
    "args": {
      "External id": 2487742,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], []], "Ev Idx": 4541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657666157.514, "dur": 5.338,
    "args": {
      "External id": 2487743,"Record function id": 0, "Concrete Inputs": ["[12602]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657666164.458, "dur": 35190.860,
    "args": {
      "External id": 2487744,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[12602], [12602], []], "Ev Idx": 4543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657701495.408, "dur": 8.858,
    "args": {
      "External id": 2487745,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[51384320], [], [], [], [], []], "Ev Idx": 4544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657701498.425, "dur": 5.416,
    "args": {
      "External id": 2487746,"Record function id": 0, "Concrete Inputs": ["[6423040]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657701545.577, "dur": 392.637,
    "args": {
      "External id": 2487747,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[6423040], [51384320], [], [], [], []], "Ev Idx": 4546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657701574.185, "dur": 358.753,
    "args": {
      "External id": 2487748,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 6423040, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[51384320], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 4547, "In msg nelems": 51384320
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657701585.696, "dur": 342.107,
    "args": {
      "External id": 2487749,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[51384320]], "Ev Idx": 4548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657701960.743, "dur": 2.191,
    "args": {
      "External id": 2487750,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 4549, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702055.373, "dur": 7.007,
    "args": {
      "External id": 2487751,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702106.375, "dur": 1.437,
    "args": {
      "External id": 2487752,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "256"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702122.080, "dur": 1.376,
    "args": {
      "External id": 2487753,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702134.695, "dur": 1.080,
    "args": {
      "External id": 2487754,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1048832"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702149.571, "dur": 0.723,
    "args": {
      "External id": 2487755,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1573120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702161.975, "dur": 0.692,
    "args": {
      "External id": 2487756,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702173.217, "dur": 0.796,
    "args": {
      "External id": 2487757,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "2097664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702185.128, "dur": 1.739,
    "args": {
      "External id": 2487758,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "3539456"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702197.118, "dur": 0.692,
    "args": {
      "External id": 2487759,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "4981248"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702291.697, "dur": 2678.304,
    "args": {
      "External id": 2487760,"Record function id": 0, "Ev Idx": 4559
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.7)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702310.512, "dur": 1037.468,
    "args": {
      "External id": 2487761,"Record function id": 0, "Ev Idx": 4560
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.7)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702326.169, "dur": 326.533,
    "args": {
      "External id": 2487762,"Record function id": 0, "Ev Idx": 4561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702403.668, "dur": 3.773,
    "args": {
      "External id": 2487763,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 4562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702410.402, "dur": 0.941,
    "args": {
      "External id": 2487764,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 4563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702413.471, "dur": 0.866,
    "args": {
      "External id": 2487765,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 4564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702416.320, "dur": 0.887,
    "args": {
      "External id": 2487766,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 4565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702419.986, "dur": 0.808,
    "args": {
      "External id": 2487767,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 4566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702422.085, "dur": 1.147,
    "args": {
      "External id": 2487768,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 4567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702424.863, "dur": 1.379,
    "args": {
      "External id": 2487769,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 4568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702427.628, "dur": 3.075,
    "args": {
      "External id": 2487770,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 4569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702433.302, "dur": 0.824,
    "args": {
      "External id": 2487771,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 4570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702435.542, "dur": 1.131,
    "args": {
      "External id": 2487772,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 4571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702455.528, "dur": 165.113,
    "args": {
      "External id": 2487773,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 4572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702470.997, "dur": 144.469,
    "args": {
      "External id": 2487774,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 4573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702488.705, "dur": 12.558,
    "args": {
      "External id": 2487775,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702504.208, "dur": 83.410,
    "args": {
      "External id": 2487776,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 4575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702506.882, "dur": 80.336,
    "args": {
      "External id": 2487777,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 4576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702511.104, "dur": 6.015,
    "args": {
      "External id": 2487778,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702519.934, "dur": 66.439,
    "args": {
      "External id": 2487779,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 4578
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.6", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702730.711, "dur": 609.936,
    "args": {
      "External id": 2487780,"Record function id": 0, "Ev Idx": 4579
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.6)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702747.304, "dur": 580.802,
    "args": {
      "External id": 2487781,"Record function id": 0, "Ev Idx": 4580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702801.992, "dur": 5.412,
    "args": {
      "External id": 2487782,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702822.034, "dur": 33.445,
    "args": {
      "External id": 2487783,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 4582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702829.991, "dur": 2.650,
    "args": {
      "External id": 2487784,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702834.861, "dur": 0.649,
    "args": {
      "External id": 2487785,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702836.934, "dur": 2.110,
    "args": {
      "External id": 2487786,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702841.191, "dur": 0.287,
    "args": {
      "External id": 2487787,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702842.793, "dur": 0.300,
    "args": {
      "External id": 2487788,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702844.410, "dur": 0.254,
    "args": {
      "External id": 2487789,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702846.345, "dur": 0.288,
    "args": {
      "External id": 2487790,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702847.866, "dur": 0.415,
    "args": {
      "External id": 2487791,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702850.154, "dur": 0.734,
    "args": {
      "External id": 2487792,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702865.248, "dur": 30.536,
    "args": {
      "External id": 2487793,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 4592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702925.908, "dur": 170.968,
    "args": {
      "External id": 2487794,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 4593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702935.924, "dur": 2.840,
    "args": {
      "External id": 2487795,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702943.525, "dur": 11.870,
    "args": {
      "External id": 2487796,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 4595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702947.239, "dur": 7.752,
    "args": {
      "External id": 2487797,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 4596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702950.874, "dur": 2.515,
    "args": {
      "External id": 2487798,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 4597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702962.171, "dur": 65.527,
    "args": {
      "External id": 2487799,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 4598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702964.314, "dur": 0.484,
    "args": {
      "External id": 2487800,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702966.687, "dur": 0.514,
    "args": {
      "External id": 2487801,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702968.415, "dur": 0.916,
    "args": {
      "External id": 2487802,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702970.637, "dur": 0.243,
    "args": {
      "External id": 2487803,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702972.325, "dur": 0.244,
    "args": {
      "External id": 2487804,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702974.464, "dur": 0.293,
    "args": {
      "External id": 2487805,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702976.507, "dur": 0.380,
    "args": {
      "External id": 2487806,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657702978.056, "dur": 2.331,
    "args": {
      "External id": 2487807,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657703021.607, "dur": 0.531,
    "args": {
      "External id": 2487808,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657703062.216, "dur": 26.291,
    "args": {
      "External id": 2487809,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 4608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657703145.144, "dur": 113.899,
    "args": {
      "External id": 2487810,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 4609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657703174.344, "dur": 81.227,
    "args": {
      "External id": 2487811,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 4610, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657703184.057, "dur": 66.652,
    "args": {
      "External id": 2487812,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 4611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657703275.803, "dur": 1.889,
    "args": {
      "External id": 2487813,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 4612, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657703354.535, "dur": 1592.935,
    "args": {
      "External id": 2487814,"Sequence number": 24740436, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 4613
    }
  },
  {
    "ph": "f", "id": 212, "pid": 1336759, "tid": 1381173, "ts": 1514657703354.535,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657703457.285, "dur": 121.559,
    "args": {
      "External id": 2487815,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [5632, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 4614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657703623.258, "dur": 37.386,
    "args": {
      "External id": 2487816,"kernel_hash": "cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/m5/cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [65536, 5632], [16, 4096, 5632], [16, 4096, 5632], [16, 4096, 5632], []], "Ev Idx": 4615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657703678.705, "dur": 51.307,
    "args": {
      "External id": 2487817,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 2048], [0, 5632, 1], [11534336, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632], [1, 2048, 5632]], "Ev Idx": 4616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657703740.080, "dur": 31.323,
    "args": {
      "External id": 2487818,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 4617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657703784.071, "dur": 43.695,
    "args": {
      "External id": 2487819,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 4618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657703834.615, "dur": 26.915,
    "args": {
      "External id": 2487820,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 4619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657703870.475, "dur": 41.102,
    "args": {
      "External id": 2487821,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 4620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657703934.877, "dur": 20.984,
    "args": {
      "External id": 2487822,"kernel_hash": "c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/47/c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 4621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657703973.617, "dur": 70.386,
    "args": {
      "External id": 2487823,"kernel_hash": "cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/xr/cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657704072.662, "dur": 19.428,
    "args": {
      "External id": 2487824,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 4623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657704107.719, "dur": 14.576,
    "args": {
      "External id": 2487825,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 4624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657704137.471, "dur": 32.442,
    "args": {
      "External id": 2487826,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 4625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657704173.428, "dur": 32.693,
    "args": {
      "External id": 2487827,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 4626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657704233.402, "dur": 164.314,
    "args": {
      "External id": 2487828,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 4627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657704307.815, "dur": 6.300,
    "args": {
      "External id": 2487829,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657704315.974, "dur": 2.576,
    "args": {
      "External id": 2487830,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657704429.934, "dur": 23.256,
    "args": {
      "External id": 2487831,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657704464.703, "dur": 13.968,
    "args": {
      "External id": 2487832,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657704486.400, "dur": 31.250,
    "args": {
      "External id": 2487833,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 4632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657704525.098, "dur": 58.728,
    "args": {
      "External id": 2487834,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 4633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657704593.286, "dur": 25.277,
    "args": {
      "External id": 2487835,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 4634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657704623.210, "dur": 28.786,
    "args": {
      "External id": 2487836,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 4635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657704657.223, "dur": 22.053,
    "args": {
      "External id": 2487837,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 4636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657704687.864, "dur": 46.179,
    "args": {
      "External id": 2487838,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 4637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 1336759, "tid": 1381173,
    "ts": 1514657704762.406, "dur": 26.974,
    "args": {
      "External id": 2487839,"kernel_hash": "cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/g4/cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [65536, 2048], []], "Ev Idx": 4638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657704806.914, "dur": 21.448,
    "args": {
      "External id": 2487840,"kernel_hash": "c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/4n/c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657704845.980, "dur": 15.868,
    "args": {
      "External id": 2487841,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 4640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657704877.974, "dur": 13.087,
    "args": {
      "External id": 2487842,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 4641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 1336759, "tid": 1381173,
    "ts": 1514657704905.966, "dur": 14.608,
    "args": {
      "External id": 2487843,"kernel_hash": "cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u7/cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [8388608, 2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [16, 4096, 2048], []], "Ev Idx": 4642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705029.488, "dur": 17.526,
    "args": {
      "External id": 2487844,"Record function id": 0, "Ev Idx": 4643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705033.750, "dur": 11.990,
    "args": {
      "External id": 2487845,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705038.132, "dur": 6.208,
    "args": {
      "External id": 2487846,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705039.852, "dur": 4.236,
    "args": {
      "External id": 2487847,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705052.097, "dur": 6.424,
    "args": {
      "External id": 2487848,"Record function id": 0, "Ev Idx": 4647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705053.516, "dur": 4.584,
    "args": {
      "External id": 2487849,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705054.465, "dur": 3.169,
    "args": {
      "External id": 2487850,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705054.981, "dur": 2.590,
    "args": {
      "External id": 2487851,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705061.723, "dur": 4.691,
    "args": {
      "External id": 2487852,"Record function id": 0, "Ev Idx": 4651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705062.954, "dur": 2.991,
    "args": {
      "External id": 2487853,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705063.853, "dur": 1.656,
    "args": {
      "External id": 2487854,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705064.634, "dur": 0.759,
    "args": {
      "External id": 2487855,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705069.621, "dur": 4.554,
    "args": {
      "External id": 2487856,"Record function id": 0, "Ev Idx": 4655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705070.974, "dur": 2.801,
    "args": {
      "External id": 2487857,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 4656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705071.708, "dur": 1.542,
    "args": {
      "External id": 2487858,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 4657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705072.184, "dur": 0.981,
    "args": {
      "External id": 2487859,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 4658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705077.175, "dur": 4.047,
    "args": {
      "External id": 2487860,"Record function id": 0, "Ev Idx": 4659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705078.247, "dur": 2.593,
    "args": {
      "External id": 2487861,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705079.119, "dur": 1.284,
    "args": {
      "External id": 2487862,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705079.461, "dur": 0.875,
    "args": {
      "External id": 2487863,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705084.367, "dur": 3.735,
    "args": {
      "External id": 2487864,"Record function id": 0, "Ev Idx": 4663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705085.770, "dur": 1.920,
    "args": {
      "External id": 2487865,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705086.266, "dur": 0.979,
    "args": {
      "External id": 2487866,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705086.605, "dur": 0.577,
    "args": {
      "External id": 2487867,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705091.291, "dur": 3.330,
    "args": {
      "External id": 2487868,"Record function id": 0, "Ev Idx": 4667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705092.363, "dur": 1.860,
    "args": {
      "External id": 2487869,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705092.897, "dur": 0.919,
    "args": {
      "External id": 2487870,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705093.194, "dur": 0.556,
    "args": {
      "External id": 2487871,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705097.596, "dur": 4.916,
    "args": {
      "External id": 2487872,"Record function id": 0, "Ev Idx": 4671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705098.505, "dur": 3.592,
    "args": {
      "External id": 2487873,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705099.019, "dur": 2.664,
    "args": {
      "External id": 2487874,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705100.941, "dur": 0.680,
    "args": {
      "External id": 2487875,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705105.567, "dur": 5.486,
    "args": {
      "External id": 2487876,"Record function id": 0, "Ev Idx": 4675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705106.492, "dur": 4.153,
    "args": {
      "External id": 2487877,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705107.012, "dur": 3.217,
    "args": {
      "External id": 2487878,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705107.287, "dur": 2.879,
    "args": {
      "External id": 2487879,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705114.429, "dur": 35340.913,
    "args": {
      "External id": 2487880,"Record function id": 0, "Sequence number": 24740435, "Fwd thread id": 1, "Ev Idx": 4679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705115.959, "dur": 35331.615,
    "args": {
      "External id": 2487881,"Sequence number": 24740435, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 4680
    }
  },
  {
    "ph": "f", "id": 213, "pid": 1336759, "tid": 1381173, "ts": 1514657705115.959,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.7)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705144.598, "dur": 39.792,
    "args": {
      "External id": 2487882,"Record function id": 0, "Ev Idx": 4681
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.7)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705192.035, "dur": 62.876,
    "args": {
      "External id": 2487883,"Record function id": 0, "Ev Idx": 4682
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.7)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705260.515, "dur": 35179.250,
    "args": {
      "External id": 2487884,"Record function id": 0, "Ev Idx": 4683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705352.802, "dur": 6.408,
    "args": {
      "External id": 2487885,"Record function id": 0, "Concrete Inputs": ["[51384320]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705368.013, "dur": 4.674,
    "args": {
      "External id": 2487886,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 4685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705386.005, "dur": 34242.885,
    "args": {
      "External id": 2487887,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 4686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705399.056, "dur": 34220.808,
    "args": {
      "External id": 2487888,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 4687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705447.547, "dur": 13.620,
    "args": {
      "External id": 2487889,"Record function id": 0, "Concrete Inputs": ["[12602]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705467.238, "dur": 34114.961,
    "args": {
      "External id": 2487890,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], [], []], "Ev Idx": 4689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705469.794, "dur": 34111.720,
    "args": {
      "External id": 2487891,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], []], "Ev Idx": 4690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705474.198, "dur": 5.411,
    "args": {
      "External id": 2487892,"Record function id": 0, "Concrete Inputs": ["[12602]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657705481.358, "dur": 34096.501,
    "args": {
      "External id": 2487893,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[12602], [12602], []], "Ev Idx": 4692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657739716.274, "dur": 8.580,
    "args": {
      "External id": 2487894,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[51384320], [], [], [], [], []], "Ev Idx": 4693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657739718.985, "dur": 5.553,
    "args": {
      "External id": 2487895,"Record function id": 0, "Concrete Inputs": ["[6423040]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657739754.175, "dur": 397.153,
    "args": {
      "External id": 2487896,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[6423040], [51384320], [], [], [], []], "Ev Idx": 4695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657739781.458, "dur": 364.230,
    "args": {
      "External id": 2487897,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 6423040, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[51384320], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 4696, "In msg nelems": 51384320
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657739792.085, "dur": 347.332,
    "args": {
      "External id": 2487898,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[51384320]], "Ev Idx": 4697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740175.152, "dur": 1.972,
    "args": {
      "External id": 2487899,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 4698, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740236.833, "dur": 6.488,
    "args": {
      "External id": 2487900,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740286.733, "dur": 1.442,
    "args": {
      "External id": 2487901,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "256"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740305.395, "dur": 1.176,
    "args": {
      "External id": 2487902,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740317.629, "dur": 1.200,
    "args": {
      "External id": 2487903,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1048832"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740329.935, "dur": 1.018,
    "args": {
      "External id": 2487904,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1573120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740340.961, "dur": 1.014,
    "args": {
      "External id": 2487905,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740354.453, "dur": 0.697,
    "args": {
      "External id": 2487906,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "2097664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740366.439, "dur": 0.989,
    "args": {
      "External id": 2487907,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "3539456"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740380.110, "dur": 0.708,
    "args": {
      "External id": 2487908,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "4981248"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740468.788, "dur": 2680.941,
    "args": {
      "External id": 2487909,"Record function id": 0, "Ev Idx": 4708
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.6)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740486.721, "dur": 1024.100,
    "args": {
      "External id": 2487910,"Record function id": 0, "Ev Idx": 4709
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.6)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740500.962, "dur": 335.245,
    "args": {
      "External id": 2487911,"Record function id": 0, "Ev Idx": 4710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740592.936, "dur": 4.706,
    "args": {
      "External id": 2487912,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 4711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740600.975, "dur": 1.118,
    "args": {
      "External id": 2487913,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 4712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740604.004, "dur": 2.773,
    "args": {
      "External id": 2487914,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 4713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740608.673, "dur": 0.863,
    "args": {
      "External id": 2487915,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 4714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740611.148, "dur": 0.805,
    "args": {
      "External id": 2487916,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 4715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740614.662, "dur": 1.086,
    "args": {
      "External id": 2487917,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 4716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740617.322, "dur": 0.854,
    "args": {
      "External id": 2487918,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 4717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740619.651, "dur": 0.998,
    "args": {
      "External id": 2487919,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 4718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740622.181, "dur": 1.147,
    "args": {
      "External id": 2487920,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 4719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740625.645, "dur": 0.721,
    "args": {
      "External id": 2487921,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 4720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740646.024, "dur": 159.940,
    "args": {
      "External id": 2487922,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 4721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740662.292, "dur": 139.375,
    "args": {
      "External id": 2487923,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 4722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740690.678, "dur": 13.368,
    "args": {
      "External id": 2487924,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740706.806, "dur": 66.585,
    "args": {
      "External id": 2487925,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 4724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740709.243, "dur": 63.791,
    "args": {
      "External id": 2487926,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 4725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740716.605, "dur": 5.551,
    "args": {
      "External id": 2487927,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740723.989, "dur": 48.232,
    "args": {
      "External id": 2487928,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 4727
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.5", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740914.934, "dur": 588.576,
    "args": {
      "External id": 2487929,"Record function id": 0, "Ev Idx": 4728
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.5)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657740931.562, "dur": 560.317,
    "args": {
      "External id": 2487930,"Record function id": 0, "Ev Idx": 4729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741028.649, "dur": 6.415,
    "args": {
      "External id": 2487931,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741051.186, "dur": 34.645,
    "args": {
      "External id": 2487932,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 4731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741059.875, "dur": 1.657,
    "args": {
      "External id": 2487933,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741063.868, "dur": 0.400,
    "args": {
      "External id": 2487934,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741066.416, "dur": 0.689,
    "args": {
      "External id": 2487935,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741068.586, "dur": 0.631,
    "args": {
      "External id": 2487936,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741070.692, "dur": 1.397,
    "args": {
      "External id": 2487937,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741073.512, "dur": 2.667,
    "args": {
      "External id": 2487938,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741077.524, "dur": 0.475,
    "args": {
      "External id": 2487939,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741079.638, "dur": 0.465,
    "args": {
      "External id": 2487940,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741081.605, "dur": 0.400,
    "args": {
      "External id": 2487941,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741095.842, "dur": 36.765,
    "args": {
      "External id": 2487942,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 4741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741164.065, "dur": 107.826,
    "args": {
      "External id": 2487943,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 4742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741173.656, "dur": 3.332,
    "args": {
      "External id": 2487944,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741182.015, "dur": 9.500,
    "args": {
      "External id": 2487945,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 4744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741185.659, "dur": 5.448,
    "args": {
      "External id": 2487946,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 4745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741189.306, "dur": 0.661,
    "args": {
      "External id": 2487947,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 4746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741198.424, "dur": 27.039,
    "args": {
      "External id": 2487948,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 4747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741200.499, "dur": 0.603,
    "args": {
      "External id": 2487949,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741203.111, "dur": 0.309,
    "args": {
      "External id": 2487950,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741208.315, "dur": 2.490,
    "args": {
      "External id": 2487951,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741212.165, "dur": 0.428,
    "args": {
      "External id": 2487952,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741213.845, "dur": 0.498,
    "args": {
      "External id": 2487953,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741215.769, "dur": 0.496,
    "args": {
      "External id": 2487954,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741217.535, "dur": 0.477,
    "args": {
      "External id": 2487955,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741219.498, "dur": 0.567,
    "args": {
      "External id": 2487956,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741221.434, "dur": 0.349,
    "args": {
      "External id": 2487957,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741243.723, "dur": 21.071,
    "args": {
      "External id": 2487958,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 4757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741316.428, "dur": 110.041,
    "args": {
      "External id": 2487959,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 4758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741344.872, "dur": 78.099,
    "args": {
      "External id": 2487960,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 4759, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741354.385, "dur": 64.525,
    "args": {
      "External id": 2487961,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 4760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741440.735, "dur": 1.793,
    "args": {
      "External id": 2487962,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 4761, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741517.264, "dur": 1614.490,
    "args": {
      "External id": 2487963,"Sequence number": 24740434, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 4762
    }
  },
  {
    "ph": "f", "id": 214, "pid": 1336759, "tid": 1381173, "ts": 1514657741517.264,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741642.438, "dur": 101.852,
    "args": {
      "External id": 2487964,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [5632, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 4763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741784.545, "dur": 36.197,
    "args": {
      "External id": 2487965,"kernel_hash": "cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/m5/cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [65536, 5632], [16, 4096, 5632], [16, 4096, 5632], [16, 4096, 5632], []], "Ev Idx": 4764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741835.316, "dur": 47.076,
    "args": {
      "External id": 2487966,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 2048], [0, 5632, 1], [11534336, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632], [1, 2048, 5632]], "Ev Idx": 4765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741891.712, "dur": 31.443,
    "args": {
      "External id": 2487967,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 4766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657741932.320, "dur": 45.035,
    "args": {
      "External id": 2487968,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 4767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657742026.131, "dur": 33.138,
    "args": {
      "External id": 2487969,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 4768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657742067.687, "dur": 41.474,
    "args": {
      "External id": 2487970,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 4769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657742134.421, "dur": 22.440,
    "args": {
      "External id": 2487971,"kernel_hash": "c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/47/c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 4770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657742177.569, "dur": 26.792,
    "args": {
      "External id": 2487972,"kernel_hash": "cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/xr/cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657742222.493, "dur": 17.321,
    "args": {
      "External id": 2487973,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 4772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657742254.468, "dur": 14.567,
    "args": {
      "External id": 2487974,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 4773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657742277.293, "dur": 27.119,
    "args": {
      "External id": 2487975,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 4774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657742307.804, "dur": 32.549,
    "args": {
      "External id": 2487976,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 4775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657742373.227, "dur": 183.604,
    "args": {
      "External id": 2487977,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 4776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657742450.346, "dur": 5.695,
    "args": {
      "External id": 2487978,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657742457.833, "dur": 2.137,
    "args": {
      "External id": 2487979,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657742594.429, "dur": 23.616,
    "args": {
      "External id": 2487980,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657742630.851, "dur": 13.834,
    "args": {
      "External id": 2487981,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657742653.364, "dur": 40.662,
    "args": {
      "External id": 2487982,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 4781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657742699.429, "dur": 33.462,
    "args": {
      "External id": 2487983,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 4782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657742740.718, "dur": 23.197,
    "args": {
      "External id": 2487984,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 4783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657742768.348, "dur": 29.218,
    "args": {
      "External id": 2487985,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 4784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657742806.363, "dur": 18.686,
    "args": {
      "External id": 2487986,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 4785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657742832.236, "dur": 30.005,
    "args": {
      "External id": 2487987,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 4786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 1336759, "tid": 1381173,
    "ts": 1514657742881.328, "dur": 19.798,
    "args": {
      "External id": 2487988,"kernel_hash": "cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/g4/cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [65536, 2048], []], "Ev Idx": 4787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657742920.118, "dur": 37.165,
    "args": {
      "External id": 2487989,"kernel_hash": "c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/4n/c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657742977.097, "dur": 51.815,
    "args": {
      "External id": 2487990,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 4789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743057.068, "dur": 15.239,
    "args": {
      "External id": 2487991,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 4790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743085.255, "dur": 19.091,
    "args": {
      "External id": 2487992,"kernel_hash": "cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u7/cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [8388608, 2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [16, 4096, 2048], []], "Ev Idx": 4791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743174.462, "dur": 14.268,
    "args": {
      "External id": 2487993,"Record function id": 0, "Ev Idx": 4792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743177.432, "dur": 10.300,
    "args": {
      "External id": 2487994,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743181.620, "dur": 5.132,
    "args": {
      "External id": 2487995,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743182.801, "dur": 3.840,
    "args": {
      "External id": 2487996,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743192.533, "dur": 4.868,
    "args": {
      "External id": 2487997,"Record function id": 0, "Ev Idx": 4796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743193.953, "dur": 3.017,
    "args": {
      "External id": 2487998,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743194.735, "dur": 1.746,
    "args": {
      "External id": 2487999,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743195.248, "dur": 1.120,
    "args": {
      "External id": 2488000,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743200.623, "dur": 4.071,
    "args": {
      "External id": 2488001,"Record function id": 0, "Ev Idx": 4800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743201.636, "dur": 2.647,
    "args": {
      "External id": 2488002,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743202.347, "dur": 1.499,
    "args": {
      "External id": 2488003,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743202.898, "dur": 0.863,
    "args": {
      "External id": 2488004,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743207.821, "dur": 7.238,
    "args": {
      "External id": 2488005,"Record function id": 0, "Ev Idx": 4804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743212.253, "dur": 2.337,
    "args": {
      "External id": 2488006,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 4805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743212.882, "dur": 1.310,
    "args": {
      "External id": 2488007,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 4806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743213.344, "dur": 0.780,
    "args": {
      "External id": 2488008,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 4807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743218.027, "dur": 5.907,
    "args": {
      "External id": 2488009,"Record function id": 0, "Ev Idx": 4808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743218.973, "dur": 4.546,
    "args": {
      "External id": 2488010,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743219.676, "dur": 3.424,
    "args": {
      "External id": 2488011,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743220.082, "dur": 2.929,
    "args": {
      "External id": 2488012,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743226.937, "dur": 5.390,
    "args": {
      "External id": 2488013,"Record function id": 0, "Ev Idx": 4812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743228.137, "dur": 3.769,
    "args": {
      "External id": 2488014,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743228.750, "dur": 2.750,
    "args": {
      "External id": 2488015,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743230.661, "dur": 0.752,
    "args": {
      "External id": 2488016,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743235.655, "dur": 3.702,
    "args": {
      "External id": 2488017,"Record function id": 0, "Ev Idx": 4816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743236.667, "dur": 2.283,
    "args": {
      "External id": 2488018,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743237.318, "dur": 1.165,
    "args": {
      "External id": 2488019,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743237.726, "dur": 0.689,
    "args": {
      "External id": 2488020,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743242.344, "dur": 3.799,
    "args": {
      "External id": 2488021,"Record function id": 0, "Ev Idx": 4820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743243.259, "dur": 2.436,
    "args": {
      "External id": 2488022,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743243.894, "dur": 1.386,
    "args": {
      "External id": 2488023,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743244.461, "dur": 0.753,
    "args": {
      "External id": 2488024,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743249.117, "dur": 3.791,
    "args": {
      "External id": 2488025,"Record function id": 0, "Ev Idx": 4824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743250.335, "dur": 2.184,
    "args": {
      "External id": 2488026,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743250.836, "dur": 1.269,
    "args": {
      "External id": 2488027,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743251.380, "dur": 0.664,
    "args": {
      "External id": 2488028,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743256.354, "dur": 35899.202,
    "args": {
      "External id": 2488029,"Record function id": 0, "Sequence number": 24740433, "Fwd thread id": 1, "Ev Idx": 4828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743257.461, "dur": 35889.804,
    "args": {
      "External id": 2488030,"Sequence number": 24740433, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 4829
    }
  },
  {
    "ph": "f", "id": 215, "pid": 1336759, "tid": 1381173, "ts": 1514657743257.461,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.6)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743284.185, "dur": 34.839,
    "args": {
      "External id": 2488031,"Record function id": 0, "Ev Idx": 4830
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.6)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743325.855, "dur": 58.243,
    "args": {
      "External id": 2488032,"Record function id": 0, "Ev Idx": 4831
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.6)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743389.358, "dur": 35750.668,
    "args": {
      "External id": 2488033,"Record function id": 0, "Ev Idx": 4832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743472.864, "dur": 6.404,
    "args": {
      "External id": 2488034,"Record function id": 0, "Concrete Inputs": ["[51384320]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743488.981, "dur": 4.638,
    "args": {
      "External id": 2488035,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 4834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743507.802, "dur": 34822.703,
    "args": {
      "External id": 2488036,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 4835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743524.265, "dur": 34796.730,
    "args": {
      "External id": 2488037,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 4836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743594.553, "dur": 14.279,
    "args": {
      "External id": 2488038,"Record function id": 0, "Concrete Inputs": ["[12602]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743615.192, "dur": 34667.857,
    "args": {
      "External id": 2488039,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], [], []], "Ev Idx": 4838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743617.651, "dur": 34664.623,
    "args": {
      "External id": 2488040,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], []], "Ev Idx": 4839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743621.785, "dur": 7.798,
    "args": {
      "External id": 2488041,"Record function id": 0, "Concrete Inputs": ["[12602]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657743631.148, "dur": 34647.603,
    "args": {
      "External id": 2488042,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[12602], [12602], []], "Ev Idx": 4841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657778416.716, "dur": 8.200,
    "args": {
      "External id": 2488043,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[51384320], [], [], [], [], []], "Ev Idx": 4842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657778419.611, "dur": 4.955,
    "args": {
      "External id": 2488044,"Record function id": 0, "Concrete Inputs": ["[6423040]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657778455.766, "dur": 361.329,
    "args": {
      "External id": 2488045,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[6423040], [51384320], [], [], [], []], "Ev Idx": 4844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657778483.749, "dur": 327.980,
    "args": {
      "External id": 2488046,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 6423040, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[51384320], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 4845, "In msg nelems": 51384320
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657778494.930, "dur": 310.912,
    "args": {
      "External id": 2488047,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[51384320]], "Ev Idx": 4846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657778839.947, "dur": 2.251,
    "args": {
      "External id": 2488048,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 4847, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657778897.887, "dur": 6.570,
    "args": {
      "External id": 2488049,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657778947.389, "dur": 1.557,
    "args": {
      "External id": 2488050,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "256"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657778963.837, "dur": 1.267,
    "args": {
      "External id": 2488051,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657778975.473, "dur": 3.198,
    "args": {
      "External id": 2488052,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1048832"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779028.338, "dur": 1.731,
    "args": {
      "External id": 2488053,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1573120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779042.001, "dur": 0.789,
    "args": {
      "External id": 2488054,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779053.383, "dur": 1.193,
    "args": {
      "External id": 2488055,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "2097664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779065.834, "dur": 2.906,
    "args": {
      "External id": 2488056,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "3539456"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779079.579, "dur": 0.995,
    "args": {
      "External id": 2488057,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "4981248"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779170.976, "dur": 2612.883,
    "args": {
      "External id": 2488058,"Record function id": 0, "Ev Idx": 4857
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.5)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779189.736, "dur": 986.538,
    "args": {
      "External id": 2488059,"Record function id": 0, "Ev Idx": 4858
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.5)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779203.502, "dur": 295.534,
    "args": {
      "External id": 2488060,"Record function id": 0, "Ev Idx": 4859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779277.075, "dur": 3.988,
    "args": {
      "External id": 2488061,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 4860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779284.157, "dur": 0.998,
    "args": {
      "External id": 2488062,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 4861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779286.821, "dur": 0.822,
    "args": {
      "External id": 2488063,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 4862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779289.275, "dur": 0.627,
    "args": {
      "External id": 2488064,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 4863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779292.061, "dur": 1.096,
    "args": {
      "External id": 2488065,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 4864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779294.694, "dur": 2.873,
    "args": {
      "External id": 2488066,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 4865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779299.228, "dur": 0.677,
    "args": {
      "External id": 2488067,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 4866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779301.265, "dur": 0.793,
    "args": {
      "External id": 2488068,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 4867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779304.172, "dur": 1.129,
    "args": {
      "External id": 2488069,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 4868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779306.928, "dur": 0.641,
    "args": {
      "External id": 2488070,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 4869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779325.731, "dur": 147.985,
    "args": {
      "External id": 2488071,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 4870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779341.629, "dur": 127.894,
    "args": {
      "External id": 2488072,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 4871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779363.675, "dur": 12.124,
    "args": {
      "External id": 2488073,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779379.003, "dur": 64.101,
    "args": {
      "External id": 2488074,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 4873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779381.461, "dur": 61.203,
    "args": {
      "External id": 2488075,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 4874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779385.068, "dur": 5.860,
    "args": {
      "External id": 2488076,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779393.578, "dur": 48.251,
    "args": {
      "External id": 2488077,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 4876
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.4", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779593.475, "dur": 575.187,
    "args": {
      "External id": 2488078,"Record function id": 0, "Ev Idx": 4877
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.4)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779609.273, "dur": 548.181,
    "args": {
      "External id": 2488079,"Record function id": 0, "Ev Idx": 4878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779663.832, "dur": 5.401,
    "args": {
      "External id": 2488080,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779687.563, "dur": 31.585,
    "args": {
      "External id": 2488081,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 4880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779692.912, "dur": 3.191,
    "args": {
      "External id": 2488082,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779698.723, "dur": 0.575,
    "args": {
      "External id": 2488083,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779700.579, "dur": 1.451,
    "args": {
      "External id": 2488084,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779703.812, "dur": 0.306,
    "args": {
      "External id": 2488085,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779705.599, "dur": 0.380,
    "args": {
      "External id": 2488086,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779708.895, "dur": 0.325,
    "args": {
      "External id": 2488087,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779710.478, "dur": 0.349,
    "args": {
      "External id": 2488088,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779712.086, "dur": 0.445,
    "args": {
      "External id": 2488089,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779714.138, "dur": 2.159,
    "args": {
      "External id": 2488090,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779729.523, "dur": 33.113,
    "args": {
      "External id": 2488091,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 4890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779792.113, "dur": 98.800,
    "args": {
      "External id": 2488092,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 4891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779801.501, "dur": 3.291,
    "args": {
      "External id": 2488093,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779809.809, "dur": 9.615,
    "args": {
      "External id": 2488094,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 4893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779813.933, "dur": 5.065,
    "args": {
      "External id": 2488095,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 4894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779817.332, "dur": 0.533,
    "args": {
      "External id": 2488096,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 4895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779825.947, "dur": 25.729,
    "args": {
      "External id": 2488097,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 4896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779828.204, "dur": 0.540,
    "args": {
      "External id": 2488098,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779830.513, "dur": 0.650,
    "args": {
      "External id": 2488099,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779833.029, "dur": 0.338,
    "args": {
      "External id": 2488100,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779834.864, "dur": 0.466,
    "args": {
      "External id": 2488101,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779836.613, "dur": 0.420,
    "args": {
      "External id": 2488102,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779838.420, "dur": 2.470,
    "args": {
      "External id": 2488103,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779842.207, "dur": 0.296,
    "args": {
      "External id": 2488104,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779844.180, "dur": 0.398,
    "args": {
      "External id": 2488105,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779845.933, "dur": 0.378,
    "args": {
      "External id": 2488106,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779862.648, "dur": 20.833,
    "args": {
      "External id": 2488107,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 4906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779932.277, "dur": 152.697,
    "args": {
      "External id": 2488108,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 4907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779959.175, "dur": 121.887,
    "args": {
      "External id": 2488109,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 4908, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657779967.606, "dur": 108.991,
    "args": {
      "External id": 2488110,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 4909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657780103.361, "dur": 2.068,
    "args": {
      "External id": 2488111,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 4910, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657780183.411, "dur": 1582.261,
    "args": {
      "External id": 2488112,"Sequence number": 24740432, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 4911
    }
  },
  {
    "ph": "f", "id": 216, "pid": 1336759, "tid": 1381173, "ts": 1514657780183.411,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657780287.339, "dur": 99.900,
    "args": {
      "External id": 2488113,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [5632, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 4912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657780428.772, "dur": 35.994,
    "args": {
      "External id": 2488114,"kernel_hash": "cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/m5/cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [65536, 5632], [16, 4096, 5632], [16, 4096, 5632], [16, 4096, 5632], []], "Ev Idx": 4913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657780479.854, "dur": 47.189,
    "args": {
      "External id": 2488115,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 2048], [0, 5632, 1], [11534336, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632], [1, 2048, 5632]], "Ev Idx": 4914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657780558.083, "dur": 36.538,
    "args": {
      "External id": 2488116,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 4915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657780601.101, "dur": 44.516,
    "args": {
      "External id": 2488117,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 4916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657780654.474, "dur": 27.134,
    "args": {
      "External id": 2488118,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 4917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657780689.515, "dur": 41.293,
    "args": {
      "External id": 2488119,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 4918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657780756.769, "dur": 26.039,
    "args": {
      "External id": 2488120,"kernel_hash": "c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/47/c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 4919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657780800.500, "dur": 28.195,
    "args": {
      "External id": 2488121,"kernel_hash": "cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/xr/cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657780852.318, "dur": 19.335,
    "args": {
      "External id": 2488122,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 4921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657780886.593, "dur": 14.569,
    "args": {
      "External id": 2488123,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 4922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657780910.100, "dur": 26.393,
    "args": {
      "External id": 2488124,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 4923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657780939.611, "dur": 31.008,
    "args": {
      "External id": 2488125,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 4924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781035.422, "dur": 174.149,
    "args": {
      "External id": 2488126,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 4925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781113.020, "dur": 8.139,
    "args": {
      "External id": 2488127,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781123.096, "dur": 2.609,
    "args": {
      "External id": 2488128,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781243.205, "dur": 23.219,
    "args": {
      "External id": 2488129,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781284.096, "dur": 13.777,
    "args": {
      "External id": 2488130,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781305.360, "dur": 37.766,
    "args": {
      "External id": 2488131,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 4930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781349.017, "dur": 34.292,
    "args": {
      "External id": 2488132,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 4931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781390.704, "dur": 19.819,
    "args": {
      "External id": 2488133,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 4932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781416.966, "dur": 29.664,
    "args": {
      "External id": 2488134,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 4933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781452.363, "dur": 19.111,
    "args": {
      "External id": 2488135,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 4934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781477.264, "dur": 29.142,
    "args": {
      "External id": 2488136,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 4935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781527.229, "dur": 40.865,
    "args": {
      "External id": 2488137,"kernel_hash": "cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/g4/cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [65536, 2048], []], "Ev Idx": 4936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781589.708, "dur": 25.356,
    "args": {
      "External id": 2488138,"kernel_hash": "c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/4n/c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 4937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781630.237, "dur": 31.755,
    "args": {
      "External id": 2488139,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 4938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781687.088, "dur": 15.133,
    "args": {
      "External id": 2488140,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 4939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781717.563, "dur": 16.341,
    "args": {
      "External id": 2488141,"kernel_hash": "cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u7/cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [8388608, 2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [16, 4096, 2048], []], "Ev Idx": 4940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781805.712, "dur": 17.842,
    "args": {
      "External id": 2488142,"Record function id": 0, "Ev Idx": 4941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781808.602, "dur": 13.954,
    "args": {
      "External id": 2488143,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781816.285, "dur": 5.408,
    "args": {
      "External id": 2488144,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781817.681, "dur": 3.893,
    "args": {
      "External id": 2488145,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781827.191, "dur": 4.512,
    "args": {
      "External id": 2488146,"Record function id": 0, "Ev Idx": 4945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781828.474, "dur": 2.806,
    "args": {
      "External id": 2488147,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781829.171, "dur": 1.633,
    "args": {
      "External id": 2488148,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781829.734, "dur": 0.957,
    "args": {
      "External id": 2488149,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781834.930, "dur": 4.310,
    "args": {
      "External id": 2488150,"Record function id": 0, "Ev Idx": 4949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781835.895, "dur": 2.928,
    "args": {
      "External id": 2488151,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781836.887, "dur": 1.473,
    "args": {
      "External id": 2488152,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781837.600, "dur": 0.650,
    "args": {
      "External id": 2488153,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 4952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781842.473, "dur": 3.940,
    "args": {
      "External id": 2488154,"Record function id": 0, "Ev Idx": 4953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781843.697, "dur": 2.326,
    "args": {
      "External id": 2488155,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 4954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781844.369, "dur": 1.251,
    "args": {
      "External id": 2488156,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 4955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781844.998, "dur": 0.540,
    "args": {
      "External id": 2488157,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 4956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781849.512, "dur": 5.907,
    "args": {
      "External id": 2488158,"Record function id": 0, "Ev Idx": 4957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781850.583, "dur": 4.403,
    "args": {
      "External id": 2488159,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781851.502, "dur": 3.012,
    "args": {
      "External id": 2488160,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781851.968, "dur": 2.484,
    "args": {
      "External id": 2488161,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781858.462, "dur": 3.941,
    "args": {
      "External id": 2488162,"Record function id": 0, "Ev Idx": 4961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781859.578, "dur": 2.429,
    "args": {
      "External id": 2488163,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781860.276, "dur": 1.338,
    "args": {
      "External id": 2488164,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781860.795, "dur": 0.757,
    "args": {
      "External id": 2488165,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781865.527, "dur": 3.413,
    "args": {
      "External id": 2488166,"Record function id": 0, "Ev Idx": 4965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781866.510, "dur": 2.004,
    "args": {
      "External id": 2488167,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781867.112, "dur": 0.964,
    "args": {
      "External id": 2488168,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781867.389, "dur": 0.627,
    "args": {
      "External id": 2488169,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781875.509, "dur": 3.472,
    "args": {
      "External id": 2488170,"Record function id": 0, "Ev Idx": 4969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781876.436, "dur": 2.160,
    "args": {
      "External id": 2488171,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781876.963, "dur": 1.226,
    "args": {
      "External id": 2488172,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781877.265, "dur": 0.859,
    "args": {
      "External id": 2488173,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 4972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781881.972, "dur": 4.558,
    "args": {
      "External id": 2488174,"Record function id": 0, "Ev Idx": 4973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781882.944, "dur": 3.187,
    "args": {
      "External id": 2488175,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781883.411, "dur": 2.307,
    "args": {
      "External id": 2488176,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781885.063, "dur": 0.541,
    "args": {
      "External id": 2488177,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 4976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781889.911, "dur": 38928.340,
    "args": {
      "External id": 2488178,"Record function id": 0, "Sequence number": 24740431, "Fwd thread id": 1, "Ev Idx": 4977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781891.411, "dur": 38918.094,
    "args": {
      "External id": 2488179,"Sequence number": 24740431, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 4978
    }
  },
  {
    "ph": "f", "id": 217, "pid": 1336759, "tid": 1381173, "ts": 1514657781891.411,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.5)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781918.869, "dur": 40.505,
    "args": {
      "External id": 2488180,"Record function id": 0, "Ev Idx": 4979
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.5)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657781966.557, "dur": 105.806,
    "args": {
      "External id": 2488181,"Record function id": 0, "Ev Idx": 4980
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.5)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657782079.821, "dur": 38721.726,
    "args": {
      "External id": 2488182,"Record function id": 0, "Ev Idx": 4981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657782173.523, "dur": 6.616,
    "args": {
      "External id": 2488183,"Record function id": 0, "Concrete Inputs": ["[51384320]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657782190.296, "dur": 5.110,
    "args": {
      "External id": 2488184,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 4983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657782208.889, "dur": 37726.990,
    "args": {
      "External id": 2488185,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 4984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657782222.180, "dur": 37704.046,
    "args": {
      "External id": 2488186,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 4985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657782271.594, "dur": 13.425,
    "args": {
      "External id": 2488187,"Record function id": 0, "Concrete Inputs": ["[12602]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657782291.323, "dur": 37598.117,
    "args": {
      "External id": 2488188,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], [], []], "Ev Idx": 4987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657782293.915, "dur": 37594.620,
    "args": {
      "External id": 2488189,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], []], "Ev Idx": 4988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657782298.100, "dur": 4.894,
    "args": {
      "External id": 2488190,"Record function id": 0, "Concrete Inputs": ["[12602]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657782304.918, "dur": 37580.282,
    "args": {
      "External id": 2488191,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[12602], [12602], []], "Ev Idx": 4990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657820063.494, "dur": 9.369,
    "args": {
      "External id": 2488192,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[51384320], [], [], [], [], []], "Ev Idx": 4991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657820066.449, "dur": 5.891,
    "args": {
      "External id": 2488193,"Record function id": 0, "Concrete Inputs": ["[6423040]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 4992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657820100.856, "dur": 392.672,
    "args": {
      "External id": 2488194,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[6423040], [51384320], [], [], [], []], "Ev Idx": 4993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657820131.891, "dur": 356.692,
    "args": {
      "External id": 2488195,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 6423040, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[51384320], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 4994, "In msg nelems": 51384320
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657820143.035, "dur": 340.503,
    "args": {
      "External id": 2488196,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[51384320]], "Ev Idx": 4995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657820514.487, "dur": 2.458,
    "args": {
      "External id": 2488197,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 4996, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657820592.948, "dur": 8.795,
    "args": {
      "External id": 2488198,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657820645.571, "dur": 1.567,
    "args": {
      "External id": 2488199,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "256"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657820662.718, "dur": 1.405,
    "args": {
      "External id": 2488200,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 4999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657820676.780, "dur": 0.866,
    "args": {
      "External id": 2488201,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1048832"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657820689.901, "dur": 2.513,
    "args": {
      "External id": 2488202,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1573120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657820702.600, "dur": 0.900,
    "args": {
      "External id": 2488203,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657820714.590, "dur": 1.084,
    "args": {
      "External id": 2488204,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "2097664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657820726.863, "dur": 1.123,
    "args": {
      "External id": 2488205,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "3539456"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657820737.653, "dur": 2.793,
    "args": {
      "External id": 2488206,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "4981248"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657820833.754, "dur": 2671.380,
    "args": {
      "External id": 2488207,"Record function id": 0, "Ev Idx": 5006
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.4)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657820851.746, "dur": 1014.211,
    "args": {
      "External id": 2488208,"Record function id": 0, "Ev Idx": 5007
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.4)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657820866.765, "dur": 357.510,
    "args": {
      "External id": 2488209,"Record function id": 0, "Ev Idx": 5008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657820940.600, "dur": 4.005,
    "args": {
      "External id": 2488210,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 5009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657820947.406, "dur": 1.031,
    "args": {
      "External id": 2488211,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 5010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657820951.686, "dur": 0.911,
    "args": {
      "External id": 2488212,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657820954.581, "dur": 0.958,
    "args": {
      "External id": 2488213,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657820957.579, "dur": 0.911,
    "args": {
      "External id": 2488214,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657820960.040, "dur": 0.870,
    "args": {
      "External id": 2488215,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657820964.512, "dur": 1.054,
    "args": {
      "External id": 2488216,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 5015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657820967.008, "dur": 2.613,
    "args": {
      "External id": 2488217,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 5016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657820971.034, "dur": 1.034,
    "args": {
      "External id": 2488218,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 5017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657820973.626, "dur": 0.973,
    "args": {
      "External id": 2488219,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 5018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821030.251, "dur": 163.709,
    "args": {
      "External id": 2488220,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 5019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821047.357, "dur": 138.991,
    "args": {
      "External id": 2488221,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 5020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821070.892, "dur": 14.258,
    "args": {
      "External id": 2488222,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821088.533, "dur": 70.119,
    "args": {
      "External id": 2488223,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 5022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821090.938, "dur": 67.407,
    "args": {
      "External id": 2488224,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 5023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821094.852, "dur": 6.825,
    "args": {
      "External id": 2488225,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821103.388, "dur": 54.325,
    "args": {
      "External id": 2488226,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 5025
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821303.745, "dur": 555.246,
    "args": {
      "External id": 2488227,"Record function id": 0, "Ev Idx": 5026
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.3)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821320.905, "dur": 525.833,
    "args": {
      "External id": 2488228,"Record function id": 0, "Ev Idx": 5027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821377.926, "dur": 4.730,
    "args": {
      "External id": 2488229,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821400.969, "dur": 29.110,
    "args": {
      "External id": 2488230,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 5029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821405.943, "dur": 2.915,
    "args": {
      "External id": 2488231,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821410.752, "dur": 0.560,
    "args": {
      "External id": 2488232,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821412.908, "dur": 2.093,
    "args": {
      "External id": 2488233,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821417.850, "dur": 0.316,
    "args": {
      "External id": 2488234,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821419.369, "dur": 0.238,
    "args": {
      "External id": 2488235,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821420.783, "dur": 0.282,
    "args": {
      "External id": 2488236,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821422.986, "dur": 0.340,
    "args": {
      "External id": 2488237,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821424.586, "dur": 0.401,
    "args": {
      "External id": 2488238,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821426.607, "dur": 0.612,
    "args": {
      "External id": 2488239,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821439.577, "dur": 31.869,
    "args": {
      "External id": 2488240,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 5039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821499.938, "dur": 123.722,
    "args": {
      "External id": 2488241,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 5040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821508.826, "dur": 3.268,
    "args": {
      "External id": 2488242,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821516.754, "dur": 28.447,
    "args": {
      "External id": 2488243,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 5042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821520.581, "dur": 23.762,
    "args": {
      "External id": 2488244,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 5043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821524.345, "dur": 2.335,
    "args": {
      "External id": 2488245,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 5044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821554.366, "dur": 25.425,
    "args": {
      "External id": 2488246,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 5045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821556.998, "dur": 0.481,
    "args": {
      "External id": 2488247,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821559.499, "dur": 0.352,
    "args": {
      "External id": 2488248,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821561.109, "dur": 0.792,
    "args": {
      "External id": 2488249,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821563.266, "dur": 0.293,
    "args": {
      "External id": 2488250,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821564.823, "dur": 0.313,
    "args": {
      "External id": 2488251,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821566.885, "dur": 0.386,
    "args": {
      "External id": 2488252,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821568.424, "dur": 0.300,
    "args": {
      "External id": 2488253,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821570.153, "dur": 2.532,
    "args": {
      "External id": 2488254,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821574.171, "dur": 0.494,
    "args": {
      "External id": 2488255,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821593.189, "dur": 22.808,
    "args": {
      "External id": 2488256,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 5055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821668.293, "dur": 111.063,
    "args": {
      "External id": 2488257,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 5056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821694.977, "dur": 80.925,
    "args": {
      "External id": 2488258,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 5057, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821704.310, "dur": 67.198,
    "args": {
      "External id": 2488259,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 5058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821794.973, "dur": 1.748,
    "args": {
      "External id": 2488260,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 5059, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821872.373, "dur": 1615.477,
    "args": {
      "External id": 2488261,"Sequence number": 24740430, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 5060
    }
  },
  {
    "ph": "f", "id": 218, "pid": 1336759, "tid": 1381173, "ts": 1514657821872.373,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657821976.303, "dur": 144.354,
    "args": {
      "External id": 2488262,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [5632, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 5061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657822166.663, "dur": 42.088,
    "args": {
      "External id": 2488263,"kernel_hash": "cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/m5/cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [65536, 5632], [16, 4096, 5632], [16, 4096, 5632], [16, 4096, 5632], []], "Ev Idx": 5062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657822227.542, "dur": 47.994,
    "args": {
      "External id": 2488264,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 2048], [0, 5632, 1], [11534336, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632], [1, 2048, 5632]], "Ev Idx": 5063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657822285.161, "dur": 31.254,
    "args": {
      "External id": 2488265,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 5064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657822323.041, "dur": 44.245,
    "args": {
      "External id": 2488266,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 5065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657822373.550, "dur": 26.761,
    "args": {
      "External id": 2488267,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 5066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657822409.488, "dur": 42.459,
    "args": {
      "External id": 2488268,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 5067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657822473.483, "dur": 26.106,
    "args": {
      "External id": 2488269,"kernel_hash": "c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/47/c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 5068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657822518.299, "dur": 45.856,
    "args": {
      "External id": 2488270,"kernel_hash": "cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/xr/cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657822588.720, "dur": 18.874,
    "args": {
      "External id": 2488271,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 5070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657822624.810, "dur": 14.563,
    "args": {
      "External id": 2488272,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 5071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657822648.898, "dur": 33.686,
    "args": {
      "External id": 2488273,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 5072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657822685.612, "dur": 31.468,
    "args": {
      "External id": 2488274,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 5073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657822743.834, "dur": 165.062,
    "args": {
      "External id": 2488275,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 5074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657822819.556, "dur": 6.183,
    "args": {
      "External id": 2488276,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657822827.677, "dur": 2.584,
    "args": {
      "External id": 2488277,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657822940.584, "dur": 24.408,
    "args": {
      "External id": 2488278,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657822976.320, "dur": 55.752,
    "args": {
      "External id": 2488279,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823044.391, "dur": 40.509,
    "args": {
      "External id": 2488280,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 5079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823092.614, "dur": 36.451,
    "args": {
      "External id": 2488281,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 5080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823136.136, "dur": 20.359,
    "args": {
      "External id": 2488282,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 5081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823160.740, "dur": 29.429,
    "args": {
      "External id": 2488283,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 5082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823195.726, "dur": 19.068,
    "args": {
      "External id": 2488284,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 5083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823222.862, "dur": 28.542,
    "args": {
      "External id": 2488285,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 5084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823285.677, "dur": 30.222,
    "args": {
      "External id": 2488286,"kernel_hash": "cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/g4/cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [65536, 2048], []], "Ev Idx": 5085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823336.390, "dur": 26.329,
    "args": {
      "External id": 2488287,"kernel_hash": "c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/4n/c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5086
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823382.449, "dur": 15.134,
    "args": {
      "External id": 2488288,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 5087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823416.991, "dur": 13.446,
    "args": {
      "External id": 2488289,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 5088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823442.103, "dur": 15.743,
    "args": {
      "External id": 2488290,"kernel_hash": "cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u7/cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [8388608, 2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [16, 4096, 2048], []], "Ev Idx": 5089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823527.364, "dur": 35.943,
    "args": {
      "External id": 2488291,"Record function id": 0, "Ev Idx": 5090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823550.972, "dur": 11.145,
    "args": {
      "External id": 2488292,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 5091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823554.988, "dur": 5.808,
    "args": {
      "External id": 2488293,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 5092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823556.494, "dur": 4.018,
    "args": {
      "External id": 2488294,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 5093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823568.965, "dur": 7.413,
    "args": {
      "External id": 2488295,"Record function id": 0, "Ev Idx": 5094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823570.720, "dur": 5.173,
    "args": {
      "External id": 2488296,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823571.602, "dur": 3.754,
    "args": {
      "External id": 2488297,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823572.113, "dur": 3.147,
    "args": {
      "External id": 2488298,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823579.664, "dur": 4.168,
    "args": {
      "External id": 2488299,"Record function id": 0, "Ev Idx": 5098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823580.911, "dur": 2.517,
    "args": {
      "External id": 2488300,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 5099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823581.458, "dur": 1.551,
    "args": {
      "External id": 2488301,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 5100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823582.086, "dur": 0.847,
    "args": {
      "External id": 2488302,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 5101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823590.557, "dur": 3.635,
    "args": {
      "External id": 2488303,"Record function id": 0, "Ev Idx": 5102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823591.723, "dur": 2.084,
    "args": {
      "External id": 2488304,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 5103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823592.271, "dur": 1.111,
    "args": {
      "External id": 2488305,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 5104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823592.552, "dur": 0.764,
    "args": {
      "External id": 2488306,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 5105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823597.221, "dur": 3.617,
    "args": {
      "External id": 2488307,"Record function id": 0, "Ev Idx": 5106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823598.369, "dur": 2.088,
    "args": {
      "External id": 2488308,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 5107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823598.919, "dur": 1.134,
    "args": {
      "External id": 2488309,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 5108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823599.333, "dur": 0.654,
    "args": {
      "External id": 2488310,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 5109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823603.905, "dur": 5.296,
    "args": {
      "External id": 2488311,"Record function id": 0, "Ev Idx": 5110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823604.971, "dur": 3.805,
    "args": {
      "External id": 2488312,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823605.694, "dur": 2.678,
    "args": {
      "External id": 2488313,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823607.547, "dur": 0.760,
    "args": {
      "External id": 2488314,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823612.332, "dur": 3.759,
    "args": {
      "External id": 2488315,"Record function id": 0, "Ev Idx": 5114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823613.311, "dur": 2.398,
    "args": {
      "External id": 2488316,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823614.094, "dur": 1.033,
    "args": {
      "External id": 2488317,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823614.361, "dur": 0.697,
    "args": {
      "External id": 2488318,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823619.133, "dur": 31.492,
    "args": {
      "External id": 2488319,"Record function id": 0, "Ev Idx": 5118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823620.031, "dur": 30.150,
    "args": {
      "External id": 2488320,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 5119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823648.508, "dur": 1.208,
    "args": {
      "External id": 2488321,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 5120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823648.979, "dur": 0.674,
    "args": {
      "External id": 2488322,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 5121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823653.729, "dur": 6.587,
    "args": {
      "External id": 2488323,"Record function id": 0, "Ev Idx": 5122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823654.913, "dur": 4.978,
    "args": {
      "External id": 2488324,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823655.436, "dur": 4.030,
    "args": {
      "External id": 2488325,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823657.198, "dur": 2.171,
    "args": {
      "External id": 2488326,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823663.970, "dur": 35757.659,
    "args": {
      "External id": 2488327,"Record function id": 0, "Sequence number": 24740429, "Fwd thread id": 1, "Ev Idx": 5126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823665.629, "dur": 35747.662,
    "args": {
      "External id": 2488328,"Sequence number": 24740429, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 5127
    }
  },
  {
    "ph": "f", "id": 219, "pid": 1336759, "tid": 1381173, "ts": 1514657823665.629,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.4)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823695.397, "dur": 37.902,
    "args": {
      "External id": 2488329,"Record function id": 0, "Ev Idx": 5128
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.4)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823740.482, "dur": 62.203,
    "args": {
      "External id": 2488330,"Record function id": 0, "Ev Idx": 5129
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.4)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823808.329, "dur": 35597.201,
    "args": {
      "External id": 2488331,"Record function id": 0, "Ev Idx": 5130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823898.177, "dur": 6.610,
    "args": {
      "External id": 2488332,"Record function id": 0, "Concrete Inputs": ["[51384320]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823914.054, "dur": 4.609,
    "args": {
      "External id": 2488333,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 5132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823932.949, "dur": 34677.280,
    "args": {
      "External id": 2488334,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 5133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657823945.317, "dur": 34655.144,
    "args": {
      "External id": 2488335,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 5134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657824030.836, "dur": 19.161,
    "args": {
      "External id": 2488336,"Record function id": 0, "Concrete Inputs": ["[12602]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657824058.208, "dur": 34503.191,
    "args": {
      "External id": 2488337,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], [], []], "Ev Idx": 5136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657824060.592, "dur": 34499.890,
    "args": {
      "External id": 2488338,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], []], "Ev Idx": 5137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657824064.620, "dur": 6.334,
    "args": {
      "External id": 2488339,"Record function id": 0, "Concrete Inputs": ["[12602]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657824072.866, "dur": 34483.210,
    "args": {
      "External id": 2488340,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[12602], [12602], []], "Ev Idx": 5139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657858705.892, "dur": 10.989,
    "args": {
      "External id": 2488341,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[51384320], [], [], [], [], []], "Ev Idx": 5140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657858709.040, "dur": 7.400,
    "args": {
      "External id": 2488342,"Record function id": 0, "Concrete Inputs": ["[6423040]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657858745.460, "dur": 368.875,
    "args": {
      "External id": 2488343,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[6423040], [51384320], [], [], [], []], "Ev Idx": 5142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657858774.248, "dur": 334.645,
    "args": {
      "External id": 2488344,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 6423040, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[51384320], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 5143, "In msg nelems": 51384320
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657858784.851, "dur": 318.223,
    "args": {
      "External id": 2488345,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[51384320]], "Ev Idx": 5144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859136.640, "dur": 2.339,
    "args": {
      "External id": 2488346,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 5145, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859198.384, "dur": 6.540,
    "args": {
      "External id": 2488347,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859248.288, "dur": 1.510,
    "args": {
      "External id": 2488348,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "256"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859264.579, "dur": 1.284,
    "args": {
      "External id": 2488349,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859279.558, "dur": 0.906,
    "args": {
      "External id": 2488350,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1048832"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859291.798, "dur": 1.087,
    "args": {
      "External id": 2488351,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1573120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859303.044, "dur": 0.977,
    "args": {
      "External id": 2488352,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859315.232, "dur": 0.979,
    "args": {
      "External id": 2488353,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "2097664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859328.563, "dur": 1.296,
    "args": {
      "External id": 2488354,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "3539456"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859342.099, "dur": 0.769,
    "args": {
      "External id": 2488355,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "4981248"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859435.104, "dur": 2656.321,
    "args": {
      "External id": 2488356,"Record function id": 0, "Ev Idx": 5155
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.3)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859453.307, "dur": 997.228,
    "args": {
      "External id": 2488357,"Record function id": 0, "Ev Idx": 5156
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.3)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859468.930, "dur": 321.541,
    "args": {
      "External id": 2488358,"Record function id": 0, "Ev Idx": 5157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859561.863, "dur": 4.952,
    "args": {
      "External id": 2488359,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 5158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859570.513, "dur": 0.841,
    "args": {
      "External id": 2488360,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 5159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859574.227, "dur": 0.945,
    "args": {
      "External id": 2488361,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859576.483, "dur": 0.836,
    "args": {
      "External id": 2488362,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859578.711, "dur": 2.649,
    "args": {
      "External id": 2488363,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859582.701, "dur": 0.673,
    "args": {
      "External id": 2488364,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859585.910, "dur": 0.827,
    "args": {
      "External id": 2488365,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 5164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859588.294, "dur": 0.685,
    "args": {
      "External id": 2488366,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 5165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859590.466, "dur": 0.898,
    "args": {
      "External id": 2488367,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 5166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859592.751, "dur": 1.176,
    "args": {
      "External id": 2488368,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 5167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859613.226, "dur": 148.360,
    "args": {
      "External id": 2488369,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 5168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859628.136, "dur": 128.844,
    "args": {
      "External id": 2488370,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 5169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859649.558, "dur": 12.220,
    "args": {
      "External id": 2488371,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859664.478, "dur": 65.357,
    "args": {
      "External id": 2488372,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 5171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859666.890, "dur": 62.650,
    "args": {
      "External id": 2488373,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 5172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859670.601, "dur": 6.346,
    "args": {
      "External id": 2488374,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859678.576, "dur": 50.315,
    "args": {
      "External id": 2488375,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 5174
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859871.117, "dur": 572.435,
    "args": {
      "External id": 2488376,"Record function id": 0, "Ev Idx": 5175
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.2)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859886.643, "dur": 545.126,
    "args": {
      "External id": 2488377,"Record function id": 0, "Ev Idx": 5176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859947.337, "dur": 6.882,
    "args": {
      "External id": 2488378,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859972.031, "dur": 63.203,
    "args": {
      "External id": 2488379,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 5178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859976.827, "dur": 1.676,
    "args": {
      "External id": 2488380,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657859980.220, "dur": 35.833,
    "args": {
      "External id": 2488381,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860019.253, "dur": 0.512,
    "args": {
      "External id": 2488382,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860021.458, "dur": 0.449,
    "args": {
      "External id": 2488383,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860022.821, "dur": 0.393,
    "args": {
      "External id": 2488384,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860024.107, "dur": 0.921,
    "args": {
      "External id": 2488385,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860026.216, "dur": 0.547,
    "args": {
      "External id": 2488386,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860027.894, "dur": 2.025,
    "args": {
      "External id": 2488387,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860031.560, "dur": 0.440,
    "args": {
      "External id": 2488388,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860046.282, "dur": 37.960,
    "args": {
      "External id": 2488389,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 5188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860116.595, "dur": 98.688,
    "args": {
      "External id": 2488390,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 5189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860129.584, "dur": 4.296,
    "args": {
      "External id": 2488391,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860138.637, "dur": 9.446,
    "args": {
      "External id": 2488392,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 5191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860142.777, "dur": 4.859,
    "args": {
      "External id": 2488393,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 5192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860145.916, "dur": 0.575,
    "args": {
      "External id": 2488394,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 5193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860155.176, "dur": 20.831,
    "args": {
      "External id": 2488395,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 5194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860156.713, "dur": 0.429,
    "args": {
      "External id": 2488396,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860158.353, "dur": 0.462,
    "args": {
      "External id": 2488397,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860160.125, "dur": 0.283,
    "args": {
      "External id": 2488398,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860161.301, "dur": 0.294,
    "args": {
      "External id": 2488399,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860162.714, "dur": 2.256,
    "args": {
      "External id": 2488400,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860166.132, "dur": 0.472,
    "args": {
      "External id": 2488401,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860167.548, "dur": 0.316,
    "args": {
      "External id": 2488402,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860168.716, "dur": 0.604,
    "args": {
      "External id": 2488403,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860170.345, "dur": 0.591,
    "args": {
      "External id": 2488404,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860187.003, "dur": 21.207,
    "args": {
      "External id": 2488405,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 5204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860259.487, "dur": 108.010,
    "args": {
      "External id": 2488406,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 5205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860284.506, "dur": 79.578,
    "args": {
      "External id": 2488407,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 5206, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860293.307, "dur": 66.946,
    "args": {
      "External id": 2488408,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 5207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860382.143, "dur": 1.948,
    "args": {
      "External id": 2488409,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 5208, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860457.519, "dur": 1615.040,
    "args": {
      "External id": 2488410,"Sequence number": 24740428, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 5209
    }
  },
  {
    "ph": "f", "id": 220, "pid": 1336759, "tid": 1381173, "ts": 1514657860457.519,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860579.271, "dur": 106.962,
    "args": {
      "External id": 2488411,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [5632, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 5210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860730.451, "dur": 37.861,
    "args": {
      "External id": 2488412,"kernel_hash": "cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/m5/cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [65536, 5632], [16, 4096, 5632], [16, 4096, 5632], [16, 4096, 5632], []], "Ev Idx": 5211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860784.477, "dur": 47.949,
    "args": {
      "External id": 2488413,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 2048], [0, 5632, 1], [11534336, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632], [1, 2048, 5632]], "Ev Idx": 5212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860841.533, "dur": 31.504,
    "args": {
      "External id": 2488414,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 5213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860880.158, "dur": 44.429,
    "args": {
      "External id": 2488415,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 5214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860934.267, "dur": 28.439,
    "args": {
      "External id": 2488416,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 5215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657860970.570, "dur": 84.189,
    "args": {
      "External id": 2488417,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 5216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657861083.201, "dur": 23.146,
    "args": {
      "External id": 2488418,"kernel_hash": "c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/47/c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 5217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657861124.691, "dur": 27.958,
    "args": {
      "External id": 2488419,"kernel_hash": "cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/xr/cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657861173.348, "dur": 18.271,
    "args": {
      "External id": 2488420,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 5219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657861209.791, "dur": 14.810,
    "args": {
      "External id": 2488421,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 5220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657861234.550, "dur": 33.228,
    "args": {
      "External id": 2488422,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 5221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657861270.991, "dur": 31.706,
    "args": {
      "External id": 2488423,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 5222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657861330.346, "dur": 158.807,
    "args": {
      "External id": 2488424,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 5223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657861403.696, "dur": 5.767,
    "args": {
      "External id": 2488425,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657861411.238, "dur": 2.427,
    "args": {
      "External id": 2488426,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657861523.662, "dur": 40.315,
    "args": {
      "External id": 2488427,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657861581.152, "dur": 14.849,
    "args": {
      "External id": 2488428,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657861606.538, "dur": 35.792,
    "args": {
      "External id": 2488429,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 5228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657861647.928, "dur": 37.892,
    "args": {
      "External id": 2488430,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 5229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657861691.974, "dur": 22.877,
    "args": {
      "External id": 2488431,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 5230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657861719.651, "dur": 30.266,
    "args": {
      "External id": 2488432,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 5231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657861757.167, "dur": 22.156,
    "args": {
      "External id": 2488433,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 5232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657861786.172, "dur": 28.840,
    "args": {
      "External id": 2488434,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 5233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 1336759, "tid": 1381173,
    "ts": 1514657861832.950, "dur": 20.267,
    "args": {
      "External id": 2488435,"kernel_hash": "cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/g4/cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [65536, 2048], []], "Ev Idx": 5234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657861884.134, "dur": 27.336,
    "args": {
      "External id": 2488436,"kernel_hash": "c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/4n/c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657861926.501, "dur": 15.318,
    "args": {
      "External id": 2488437,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 5236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657861957.201, "dur": 15.707,
    "args": {
      "External id": 2488438,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 5237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862020.628, "dur": 22.268,
    "args": {
      "External id": 2488439,"kernel_hash": "cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u7/cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [8388608, 2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [16, 4096, 2048], []], "Ev Idx": 5238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862113.106, "dur": 16.150,
    "args": {
      "External id": 2488440,"Record function id": 0, "Ev Idx": 5239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862116.130, "dur": 12.234,
    "args": {
      "External id": 2488441,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 5240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862120.137, "dur": 7.094,
    "args": {
      "External id": 2488442,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 5241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862121.596, "dur": 5.532,
    "args": {
      "External id": 2488443,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 5242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862132.954, "dur": 4.283,
    "args": {
      "External id": 2488444,"Record function id": 0, "Ev Idx": 5243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862134.158, "dur": 2.624,
    "args": {
      "External id": 2488445,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862134.790, "dur": 1.474,
    "args": {
      "External id": 2488446,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862135.276, "dur": 0.866,
    "args": {
      "External id": 2488447,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862140.467, "dur": 4.341,
    "args": {
      "External id": 2488448,"Record function id": 0, "Ev Idx": 5247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862141.599, "dur": 2.765,
    "args": {
      "External id": 2488449,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 5248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862142.167, "dur": 1.756,
    "args": {
      "External id": 2488450,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 5249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862142.853, "dur": 0.936,
    "args": {
      "External id": 2488451,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 5250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862147.922, "dur": 3.525,
    "args": {
      "External id": 2488452,"Record function id": 0, "Ev Idx": 5251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862149.091, "dur": 1.937,
    "args": {
      "External id": 2488453,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 5252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862149.555, "dur": 1.050,
    "args": {
      "External id": 2488454,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 5253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862149.840, "dur": 0.657,
    "args": {
      "External id": 2488455,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 5254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862154.478, "dur": 3.261,
    "args": {
      "External id": 2488456,"Record function id": 0, "Ev Idx": 5255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862155.405, "dur": 1.937,
    "args": {
      "External id": 2488457,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 5256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862155.894, "dur": 0.980,
    "args": {
      "External id": 2488458,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 5257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862156.134, "dur": 0.676,
    "args": {
      "External id": 2488459,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 5258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862160.938, "dur": 3.334,
    "args": {
      "External id": 2488460,"Record function id": 0, "Ev Idx": 5259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862161.936, "dur": 1.912,
    "args": {
      "External id": 2488461,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862162.425, "dur": 1.003,
    "args": {
      "External id": 2488462,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862162.781, "dur": 0.580,
    "args": {
      "External id": 2488463,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862167.431, "dur": 4.393,
    "args": {
      "External id": 2488464,"Record function id": 0, "Ev Idx": 5263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862168.379, "dur": 3.048,
    "args": {
      "External id": 2488465,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862168.854, "dur": 2.121,
    "args": {
      "External id": 2488466,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862170.283, "dur": 0.600,
    "args": {
      "External id": 2488467,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862177.759, "dur": 9.945,
    "args": {
      "External id": 2488468,"Record function id": 0, "Ev Idx": 5267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862178.674, "dur": 8.593,
    "args": {
      "External id": 2488469,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 5268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862184.006, "dur": 2.764,
    "args": {
      "External id": 2488470,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 5269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862184.302, "dur": 2.405,
    "args": {
      "External id": 2488471,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 5270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862190.813, "dur": 3.493,
    "args": {
      "External id": 2488472,"Record function id": 0, "Ev Idx": 5271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862191.829, "dur": 2.071,
    "args": {
      "External id": 2488473,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862192.295, "dur": 1.160,
    "args": {
      "External id": 2488474,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862192.546, "dur": 0.848,
    "args": {
      "External id": 2488475,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862197.633, "dur": 36776.367,
    "args": {
      "External id": 2488476,"Record function id": 0, "Sequence number": 24740427, "Fwd thread id": 1, "Ev Idx": 5275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862199.389, "dur": 36766.442,
    "args": {
      "External id": 2488477,"Sequence number": 24740427, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 5276
    }
  },
  {
    "ph": "f", "id": 221, "pid": 1336759, "tid": 1381173, "ts": 1514657862199.389,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.3)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862227.175, "dur": 38.607,
    "args": {
      "External id": 2488478,"Record function id": 0, "Ev Idx": 5277
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.3)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862272.910, "dur": 63.107,
    "args": {
      "External id": 2488479,"Record function id": 0, "Ev Idx": 5278
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.3)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862342.026, "dur": 36616.915,
    "args": {
      "External id": 2488480,"Record function id": 0, "Ev Idx": 5279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862431.931, "dur": 6.783,
    "args": {
      "External id": 2488481,"Record function id": 0, "Concrete Inputs": ["[51384320]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862447.550, "dur": 4.654,
    "args": {
      "External id": 2488482,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 5281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862465.233, "dur": 35723.808,
    "args": {
      "External id": 2488483,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 5282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862477.613, "dur": 35701.582,
    "args": {
      "External id": 2488484,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 5283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862545.791, "dur": 14.537,
    "args": {
      "External id": 2488485,"Record function id": 0, "Concrete Inputs": ["[12602]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862566.821, "dur": 35572.556,
    "args": {
      "External id": 2488486,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], [], []], "Ev Idx": 5285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862569.407, "dur": 35569.057,
    "args": {
      "External id": 2488487,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], []], "Ev Idx": 5286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862573.124, "dur": 10.007,
    "args": {
      "External id": 2488488,"Record function id": 0, "Concrete Inputs": ["[12602]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657862585.882, "dur": 35549.162,
    "args": {
      "External id": 2488489,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[12602], [12602], []], "Ev Idx": 5288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657898281.627, "dur": 9.016,
    "args": {
      "External id": 2488490,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[51384320], [], [], [], [], []], "Ev Idx": 5289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657898284.491, "dur": 5.756,
    "args": {
      "External id": 2488491,"Record function id": 0, "Concrete Inputs": ["[6423040]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657898318.744, "dur": 361.105,
    "args": {
      "External id": 2488492,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[6423040], [51384320], [], [], [], []], "Ev Idx": 5291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657898345.847, "dur": 328.517,
    "args": {
      "External id": 2488493,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 6423040, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[51384320], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 5292, "In msg nelems": 51384320
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657898356.720, "dur": 312.216,
    "args": {
      "External id": 2488494,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[51384320]], "Ev Idx": 5293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657898700.844, "dur": 2.267,
    "args": {
      "External id": 2488495,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 5294, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657898758.215, "dur": 6.719,
    "args": {
      "External id": 2488496,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657898807.576, "dur": 1.903,
    "args": {
      "External id": 2488497,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "256"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657898823.969, "dur": 3.198,
    "args": {
      "External id": 2488498,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657898838.508, "dur": 1.208,
    "args": {
      "External id": 2488499,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1048832"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657898851.256, "dur": 0.699,
    "args": {
      "External id": 2488500,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1573120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657898862.448, "dur": 0.698,
    "args": {
      "External id": 2488501,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657898873.867, "dur": 2.674,
    "args": {
      "External id": 2488502,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "2097664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657898890.474, "dur": 1.113,
    "args": {
      "External id": 2488503,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "3539456"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657898901.886, "dur": 0.731,
    "args": {
      "External id": 2488504,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "4981248"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899021.774, "dur": 2637.009,
    "args": {
      "External id": 2488505,"Record function id": 0, "Ev Idx": 5304
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.2)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899042.365, "dur": 1000.191,
    "args": {
      "External id": 2488506,"Record function id": 0, "Ev Idx": 5305
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.2)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899056.830, "dur": 311.174,
    "args": {
      "External id": 2488507,"Record function id": 0, "Ev Idx": 5306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899137.848, "dur": 5.103,
    "args": {
      "External id": 2488508,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 5307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899146.269, "dur": 1.241,
    "args": {
      "External id": 2488509,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 5308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899149.119, "dur": 0.995,
    "args": {
      "External id": 2488510,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899151.753, "dur": 2.825,
    "args": {
      "External id": 2488511,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899157.200, "dur": 0.910,
    "args": {
      "External id": 2488512,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899159.385, "dur": 1.111,
    "args": {
      "External id": 2488513,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899162.106, "dur": 0.794,
    "args": {
      "External id": 2488514,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 5313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899164.366, "dur": 0.918,
    "args": {
      "External id": 2488515,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 5314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899168.541, "dur": 1.420,
    "args": {
      "External id": 2488516,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 5315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899171.314, "dur": 1.016,
    "args": {
      "External id": 2488517,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 5316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899189.848, "dur": 149.384,
    "args": {
      "External id": 2488518,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 5317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899205.870, "dur": 129.253,
    "args": {
      "External id": 2488519,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 5318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899225.098, "dur": 12.065,
    "args": {
      "External id": 2488520,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899239.958, "dur": 67.616,
    "args": {
      "External id": 2488521,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 5320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899243.803, "dur": 63.489,
    "args": {
      "External id": 2488522,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 5321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899247.465, "dur": 8.317,
    "args": {
      "External id": 2488523,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899257.673, "dur": 49.179,
    "args": {
      "External id": 2488524,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 5323
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899447.451, "dur": 587.891,
    "args": {
      "External id": 2488525,"Record function id": 0, "Ev Idx": 5324
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.1)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899461.818, "dur": 557.387,
    "args": {
      "External id": 2488526,"Record function id": 0, "Ev Idx": 5325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899518.066, "dur": 4.697,
    "args": {
      "External id": 2488527,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899564.474, "dur": 26.729,
    "args": {
      "External id": 2488528,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 5327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899569.234, "dur": 2.020,
    "args": {
      "External id": 2488529,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899572.876, "dur": 0.473,
    "args": {
      "External id": 2488530,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899574.483, "dur": 0.524,
    "args": {
      "External id": 2488531,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899576.271, "dur": 0.599,
    "args": {
      "External id": 2488532,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899578.606, "dur": 0.577,
    "args": {
      "External id": 2488533,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899580.278, "dur": 0.428,
    "args": {
      "External id": 2488534,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899582.203, "dur": 2.263,
    "args": {
      "External id": 2488535,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899586.028, "dur": 0.454,
    "args": {
      "External id": 2488536,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899587.747, "dur": 0.574,
    "args": {
      "External id": 2488537,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899601.661, "dur": 34.046,
    "args": {
      "External id": 2488538,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 5337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899666.581, "dur": 99.736,
    "args": {
      "External id": 2488539,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 5338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899676.900, "dur": 4.208,
    "args": {
      "External id": 2488540,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899685.975, "dur": 10.150,
    "args": {
      "External id": 2488541,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 5340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899690.321, "dur": 5.401,
    "args": {
      "External id": 2488542,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 5341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899693.858, "dur": 0.616,
    "args": {
      "External id": 2488543,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 5342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899702.991, "dur": 26.066,
    "args": {
      "External id": 2488544,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 5343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899704.796, "dur": 0.566,
    "args": {
      "External id": 2488545,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899706.708, "dur": 0.461,
    "args": {
      "External id": 2488546,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899708.349, "dur": 0.503,
    "args": {
      "External id": 2488547,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899710.291, "dur": 2.488,
    "args": {
      "External id": 2488548,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899714.052, "dur": 0.261,
    "args": {
      "External id": 2488549,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899715.514, "dur": 0.588,
    "args": {
      "External id": 2488550,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899719.648, "dur": 0.438,
    "args": {
      "External id": 2488551,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899721.291, "dur": 0.469,
    "args": {
      "External id": 2488552,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899722.970, "dur": 0.417,
    "args": {
      "External id": 2488553,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899739.298, "dur": 19.851,
    "args": {
      "External id": 2488554,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 5353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899809.506, "dur": 107.991,
    "args": {
      "External id": 2488555,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 5354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899834.516, "dur": 79.716,
    "args": {
      "External id": 2488556,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 5355, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899843.685, "dur": 65.785,
    "args": {
      "External id": 2488557,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 5356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657899933.714, "dur": 2.065,
    "args": {
      "External id": 2488558,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 5357, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657900050.754, "dur": 1590.189,
    "args": {
      "External id": 2488559,"Sequence number": 24740426, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 5358
    }
  },
  {
    "ph": "f", "id": 222, "pid": 1336759, "tid": 1381173, "ts": 1514657900050.754,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657900156.898, "dur": 107.483,
    "args": {
      "External id": 2488560,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [5632, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 5359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657900309.594, "dur": 37.297,
    "args": {
      "External id": 2488561,"kernel_hash": "cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/m5/cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [65536, 5632], [16, 4096, 5632], [16, 4096, 5632], [16, 4096, 5632], []], "Ev Idx": 5360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657900363.405, "dur": 46.214,
    "args": {
      "External id": 2488562,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 2048], [0, 5632, 1], [11534336, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632], [1, 2048, 5632]], "Ev Idx": 5361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657900419.111, "dur": 31.253,
    "args": {
      "External id": 2488563,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 5362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657900456.969, "dur": 43.228,
    "args": {
      "External id": 2488564,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 5363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657900508.648, "dur": 43.227,
    "args": {
      "External id": 2488565,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 5364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657900563.010, "dur": 44.452,
    "args": {
      "External id": 2488566,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 5365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657900631.796, "dur": 24.670,
    "args": {
      "External id": 2488567,"kernel_hash": "c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/47/c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 5366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657900676.099, "dur": 28.930,
    "args": {
      "External id": 2488568,"kernel_hash": "cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/xr/cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657900727.697, "dur": 18.106,
    "args": {
      "External id": 2488569,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 5368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657900759.575, "dur": 14.693,
    "args": {
      "External id": 2488570,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 5369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657900785.299, "dur": 28.042,
    "args": {
      "External id": 2488571,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 5370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657900816.974, "dur": 33.011,
    "args": {
      "External id": 2488572,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 5371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657900876.683, "dur": 201.437,
    "args": {
      "External id": 2488573,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 5372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657900948.319, "dur": 5.462,
    "args": {
      "External id": 2488574,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657900955.753, "dur": 2.053,
    "args": {
      "External id": 2488575,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901111.665, "dur": 23.934,
    "args": {
      "External id": 2488576,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901146.338, "dur": 14.807,
    "args": {
      "External id": 2488577,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901172.342, "dur": 38.663,
    "args": {
      "External id": 2488578,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 5377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901217.791, "dur": 38.153,
    "args": {
      "External id": 2488579,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 5378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901263.059, "dur": 23.713,
    "args": {
      "External id": 2488580,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 5379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901292.351, "dur": 29.708,
    "args": {
      "External id": 2488581,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 5380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901330.436, "dur": 22.710,
    "args": {
      "External id": 2488582,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 5381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901361.772, "dur": 29.818,
    "args": {
      "External id": 2488583,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 5382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901409.795, "dur": 19.417,
    "args": {
      "External id": 2488584,"kernel_hash": "cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/g4/cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [65536, 2048], []], "Ev Idx": 5383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901463.974, "dur": 30.581,
    "args": {
      "External id": 2488585,"kernel_hash": "c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/4n/c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901511.293, "dur": 36.072,
    "args": {
      "External id": 2488586,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 5385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901565.948, "dur": 15.599,
    "args": {
      "External id": 2488587,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 5386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901594.054, "dur": 15.939,
    "args": {
      "External id": 2488588,"kernel_hash": "cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u7/cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [8388608, 2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [16, 4096, 2048], []], "Ev Idx": 5387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901680.847, "dur": 18.766,
    "args": {
      "External id": 2488589,"Record function id": 0, "Ev Idx": 5388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901683.837, "dur": 14.915,
    "args": {
      "External id": 2488590,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 5389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901691.681, "dur": 6.130,
    "args": {
      "External id": 2488591,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 5390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901693.528, "dur": 4.095,
    "args": {
      "External id": 2488592,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 5391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901703.462, "dur": 4.893,
    "args": {
      "External id": 2488593,"Record function id": 0, "Ev Idx": 5392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901704.908, "dur": 3.007,
    "args": {
      "External id": 2488594,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901705.758, "dur": 1.656,
    "args": {
      "External id": 2488595,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901706.261, "dur": 1.076,
    "args": {
      "External id": 2488596,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901711.515, "dur": 3.998,
    "args": {
      "External id": 2488597,"Record function id": 0, "Ev Idx": 5396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901712.652, "dur": 2.424,
    "args": {
      "External id": 2488598,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 5397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901713.324, "dur": 1.344,
    "args": {
      "External id": 2488599,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 5398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901713.790, "dur": 0.782,
    "args": {
      "External id": 2488600,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 5399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901718.636, "dur": 6.052,
    "args": {
      "External id": 2488601,"Record function id": 0, "Ev Idx": 5400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901720.008, "dur": 4.272,
    "args": {
      "External id": 2488602,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 5401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901720.541, "dur": 3.297,
    "args": {
      "External id": 2488603,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 5402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901720.979, "dur": 2.792,
    "args": {
      "External id": 2488604,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 5403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901727.823, "dur": 3.856,
    "args": {
      "External id": 2488605,"Record function id": 0, "Ev Idx": 5404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901729.105, "dur": 2.138,
    "args": {
      "External id": 2488606,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 5405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901729.552, "dur": 1.291,
    "args": {
      "External id": 2488607,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 5406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901729.803, "dur": 0.948,
    "args": {
      "External id": 2488608,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 5407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901734.691, "dur": 3.610,
    "args": {
      "External id": 2488609,"Record function id": 0, "Ev Idx": 5408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901735.760, "dur": 2.095,
    "args": {
      "External id": 2488610,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901736.236, "dur": 1.162,
    "args": {
      "External id": 2488611,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901736.700, "dur": 0.578,
    "args": {
      "External id": 2488612,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901741.500, "dur": 3.580,
    "args": {
      "External id": 2488613,"Record function id": 0, "Ev Idx": 5412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901742.644, "dur": 1.966,
    "args": {
      "External id": 2488614,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901743.112, "dur": 1.079,
    "args": {
      "External id": 2488615,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901743.374, "dur": 0.687,
    "args": {
      "External id": 2488616,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901748.107, "dur": 3.655,
    "args": {
      "External id": 2488617,"Record function id": 0, "Ev Idx": 5416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901749.116, "dur": 2.228,
    "args": {
      "External id": 2488618,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 5417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901749.598, "dur": 1.307,
    "args": {
      "External id": 2488619,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 5418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901750.058, "dur": 0.740,
    "args": {
      "External id": 2488620,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 5419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901754.707, "dur": 3.854,
    "args": {
      "External id": 2488621,"Record function id": 0, "Ev Idx": 5420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901755.622, "dur": 2.522,
    "args": {
      "External id": 2488622,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901756.227, "dur": 1.475,
    "args": {
      "External id": 2488623,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901756.840, "dur": 0.748,
    "args": {
      "External id": 2488624,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901762.099, "dur": 36469.281,
    "args": {
      "External id": 2488625,"Record function id": 0, "Sequence number": 24740425, "Fwd thread id": 1, "Ev Idx": 5424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901763.432, "dur": 36459.490,
    "args": {
      "External id": 2488626,"Sequence number": 24740425, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 5425
    }
  },
  {
    "ph": "f", "id": 223, "pid": 1336759, "tid": 1381173, "ts": 1514657901763.432,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.2)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901789.573, "dur": 36.306,
    "args": {
      "External id": 2488627,"Record function id": 0, "Ev Idx": 5426
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.2)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901834.021, "dur": 63.453,
    "args": {
      "External id": 2488628,"Record function id": 0, "Ev Idx": 5427
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.2)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657901903.221, "dur": 36311.361,
    "args": {
      "External id": 2488629,"Record function id": 0, "Ev Idx": 5428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657902033.226, "dur": 7.659,
    "args": {
      "External id": 2488630,"Record function id": 0, "Concrete Inputs": ["[51384320]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657902054.554, "dur": 5.054,
    "args": {
      "External id": 2488631,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 5430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657902075.105, "dur": 35269.306,
    "args": {
      "External id": 2488632,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 5431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657902087.782, "dur": 35247.439,
    "args": {
      "External id": 2488633,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 5432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657902132.760, "dur": 13.512,
    "args": {
      "External id": 2488634,"Record function id": 0, "Concrete Inputs": ["[12602]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657902152.235, "dur": 35146.995,
    "args": {
      "External id": 2488635,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], [], []], "Ev Idx": 5434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657902154.684, "dur": 35143.842,
    "args": {
      "External id": 2488636,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], []], "Ev Idx": 5435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657902158.452, "dur": 7.832,
    "args": {
      "External id": 2488637,"Record function id": 0, "Concrete Inputs": ["[12602]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657902167.978, "dur": 35127.059,
    "args": {
      "External id": 2488638,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[12602], [12602], []], "Ev Idx": 5437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657937433.521, "dur": 9.860,
    "args": {
      "External id": 2488639,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[51384320], [], [], [], [], []], "Ev Idx": 5438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657937436.397, "dur": 6.583,
    "args": {
      "External id": 2488640,"Record function id": 0, "Concrete Inputs": ["[6423040]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657937472.232, "dur": 406.230,
    "args": {
      "External id": 2488641,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[6423040], [51384320], [], [], [], []], "Ev Idx": 5440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657937502.520, "dur": 370.789,
    "args": {
      "External id": 2488642,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 6423040, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[51384320], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 5441, "In msg nelems": 51384320
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657937513.616, "dur": 353.942,
    "args": {
      "External id": 2488643,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[51384320]], "Ev Idx": 5442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657937899.462, "dur": 2.380,
    "args": {
      "External id": 2488644,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 5443, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657937961.596, "dur": 6.481,
    "args": {
      "External id": 2488645,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938058.343, "dur": 2.219,
    "args": {
      "External id": 2488646,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "256"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938076.975, "dur": 1.093,
    "args": {
      "External id": 2488647,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938090.491, "dur": 2.815,
    "args": {
      "External id": 2488648,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1048832"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938104.607, "dur": 0.941,
    "args": {
      "External id": 2488649,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1573120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938115.317, "dur": 0.934,
    "args": {
      "External id": 2488650,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938126.741, "dur": 1.104,
    "args": {
      "External id": 2488651,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "2097664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938138.043, "dur": 2.878,
    "args": {
      "External id": 2488652,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "3539456"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938150.563, "dur": 0.852,
    "args": {
      "External id": 2488653,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "4981248"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938247.546, "dur": 2662.951,
    "args": {
      "External id": 2488654,"Record function id": 0, "Ev Idx": 5453
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.1)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938267.489, "dur": 1012.064,
    "args": {
      "External id": 2488655,"Record function id": 0, "Ev Idx": 5454
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.1)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938283.097, "dur": 329.001,
    "args": {
      "External id": 2488656,"Record function id": 0, "Ev Idx": 5455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938365.522, "dur": 4.062,
    "args": {
      "External id": 2488657,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 5456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938372.601, "dur": 1.039,
    "args": {
      "External id": 2488658,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 5457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938375.243, "dur": 0.775,
    "args": {
      "External id": 2488659,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938377.773, "dur": 0.861,
    "args": {
      "External id": 2488660,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938379.907, "dur": 0.865,
    "args": {
      "External id": 2488661,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938381.942, "dur": 0.704,
    "args": {
      "External id": 2488662,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938384.262, "dur": 3.048,
    "args": {
      "External id": 2488663,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 5462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938389.111, "dur": 0.779,
    "args": {
      "External id": 2488664,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 5463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938391.442, "dur": 1.491,
    "args": {
      "External id": 2488665,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 5464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938394.790, "dur": 1.051,
    "args": {
      "External id": 2488666,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 5465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938412.895, "dur": 167.697,
    "args": {
      "External id": 2488667,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 5466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938428.570, "dur": 146.797,
    "args": {
      "External id": 2488668,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 5467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938446.433, "dur": 12.917,
    "args": {
      "External id": 2488669,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938462.654, "dur": 83.210,
    "args": {
      "External id": 2488670,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 5469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938466.211, "dur": 79.147,
    "args": {
      "External id": 2488671,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 5470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938470.172, "dur": 5.656,
    "args": {
      "External id": 2488672,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938477.442, "dur": 66.960,
    "args": {
      "External id": 2488673,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 5472
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::backward_prefetch for model.layers.0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938702.507, "dur": 569.788,
    "args": {
      "External id": 2488674,"Record function id": 0, "Ev Idx": 5473
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.0)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938721.360, "dur": 539.174,
    "args": {
      "External id": 2488675,"Record function id": 0, "Ev Idx": 5474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938788.128, "dur": 5.581,
    "args": {
      "External id": 2488676,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938808.606, "dur": 26.252,
    "args": {
      "External id": 2488677,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 5476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938813.349, "dur": 1.483,
    "args": {
      "External id": 2488678,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938816.882, "dur": 0.510,
    "args": {
      "External id": 2488679,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938819.274, "dur": 2.287,
    "args": {
      "External id": 2488680,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938822.745, "dur": 0.336,
    "args": {
      "External id": 2488681,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938824.954, "dur": 0.348,
    "args": {
      "External id": 2488682,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938826.408, "dur": 0.551,
    "args": {
      "External id": 2488683,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938827.949, "dur": 0.312,
    "args": {
      "External id": 2488684,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938829.482, "dur": 0.553,
    "args": {
      "External id": 2488685,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938831.668, "dur": 0.388,
    "args": {
      "External id": 2488686,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938845.171, "dur": 29.695,
    "args": {
      "External id": 2488687,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 5486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938903.226, "dur": 135.286,
    "args": {
      "External id": 2488688,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 5487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938913.359, "dur": 3.217,
    "args": {
      "External id": 2488689,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938921.191, "dur": 11.261,
    "args": {
      "External id": 2488690,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 5489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938925.043, "dur": 6.980,
    "args": {
      "External id": 2488691,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 5490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938928.515, "dur": 2.389,
    "args": {
      "External id": 2488692,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 5491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938938.754, "dur": 21.848,
    "args": {
      "External id": 2488693,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 5492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938940.421, "dur": 0.386,
    "args": {
      "External id": 2488694,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938942.258, "dur": 0.452,
    "args": {
      "External id": 2488695,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938944.350, "dur": 0.348,
    "args": {
      "External id": 2488696,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938946.072, "dur": 0.585,
    "args": {
      "External id": 2488697,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938948.060, "dur": 0.305,
    "args": {
      "External id": 2488698,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938949.279, "dur": 0.464,
    "args": {
      "External id": 2488699,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938951.000, "dur": 0.297,
    "args": {
      "External id": 2488700,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938952.875, "dur": 2.088,
    "args": {
      "External id": 2488701,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938955.992, "dur": 0.550,
    "args": {
      "External id": 2488702,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657938971.134, "dur": 58.200,
    "args": {
      "External id": 2488703,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 5502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657939085.566, "dur": 107.278,
    "args": {
      "External id": 2488704,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 5503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657939108.603, "dur": 80.836,
    "args": {
      "External id": 2488705,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 5504, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657939118.018, "dur": 67.527,
    "args": {
      "External id": 2488706,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 5505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657939209.323, "dur": 1.875,
    "args": {
      "External id": 2488707,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 5506, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657939286.669, "dur": 1605.661,
    "args": {
      "External id": 2488708,"Sequence number": 24740424, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 5507
    }
  },
  {
    "ph": "f", "id": 224, "pid": 1336759, "tid": 1381173, "ts": 1514657939286.669,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657939392.288, "dur": 100.877,
    "args": {
      "External id": 2488709,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [5632, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 5508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657939548.941, "dur": 43.098,
    "args": {
      "External id": 2488710,"kernel_hash": "cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/m5/cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [65536, 5632], [16, 4096, 5632], [16, 4096, 5632], [16, 4096, 5632], []], "Ev Idx": 5509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657939610.664, "dur": 55.465,
    "args": {
      "External id": 2488711,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 2048], [0, 5632, 1], [11534336, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632], [1, 2048, 5632]], "Ev Idx": 5510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657939677.834, "dur": 32.191,
    "args": {
      "External id": 2488712,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 5511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657939716.322, "dur": 44.774,
    "args": {
      "External id": 2488713,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 5512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657939768.031, "dur": 27.026,
    "args": {
      "External id": 2488714,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 5513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657939802.230, "dur": 40.735,
    "args": {
      "External id": 2488715,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 5514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657939868.564, "dur": 22.406,
    "args": {
      "External id": 2488716,"kernel_hash": "c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/47/c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 5515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657939909.798, "dur": 28.859,
    "args": {
      "External id": 2488717,"kernel_hash": "cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/xr/cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657939959.378, "dur": 18.460,
    "args": {
      "External id": 2488718,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 5517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940032.132, "dur": 19.179,
    "args": {
      "External id": 2488719,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 5518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940061.731, "dur": 33.242,
    "args": {
      "External id": 2488720,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 5519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940098.680, "dur": 33.764,
    "args": {
      "External id": 2488721,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 5520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940162.575, "dur": 165.484,
    "args": {
      "External id": 2488722,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 5521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940234.411, "dur": 8.830,
    "args": {
      "External id": 2488723,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940245.086, "dur": 2.079,
    "args": {
      "External id": 2488724,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940362.038, "dur": 23.814,
    "args": {
      "External id": 2488725,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940396.782, "dur": 18.755,
    "args": {
      "External id": 2488726,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940423.764, "dur": 34.442,
    "args": {
      "External id": 2488727,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 5526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940464.735, "dur": 37.945,
    "args": {
      "External id": 2488728,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 5527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940510.759, "dur": 41.708,
    "args": {
      "External id": 2488729,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 5528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940560.592, "dur": 36.924,
    "args": {
      "External id": 2488730,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 5529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940604.704, "dur": 19.277,
    "args": {
      "External id": 2488731,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 5530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940631.109, "dur": 41.493,
    "args": {
      "External id": 2488732,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 5531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940697.259, "dur": 25.853,
    "args": {
      "External id": 2488733,"kernel_hash": "cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/g4/cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [65536, 2048], []], "Ev Idx": 5532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940745.591, "dur": 23.302,
    "args": {
      "External id": 2488734,"kernel_hash": "c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/4n/c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940786.587, "dur": 16.464,
    "args": {
      "External id": 2488735,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 5534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940817.894, "dur": 13.670,
    "args": {
      "External id": 2488736,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 5535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940847.857, "dur": 16.158,
    "args": {
      "External id": 2488737,"kernel_hash": "cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u7/cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [8388608, 2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [16, 4096, 2048], []], "Ev Idx": 5536
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940931.941, "dur": 16.640,
    "args": {
      "External id": 2488738,"Record function id": 0, "Ev Idx": 5537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940935.445, "dur": 12.233,
    "args": {
      "External id": 2488739,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 5538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940939.510, "dur": 7.001,
    "args": {
      "External id": 2488740,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 5539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940941.085, "dur": 5.269,
    "args": {
      "External id": 2488741,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 5540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940952.539, "dur": 5.145,
    "args": {
      "External id": 2488742,"Record function id": 0, "Ev Idx": 5541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940953.764, "dur": 3.464,
    "args": {
      "External id": 2488743,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940954.628, "dur": 2.142,
    "args": {
      "External id": 2488744,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940955.643, "dur": 1.063,
    "args": {
      "External id": 2488745,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940960.977, "dur": 4.060,
    "args": {
      "External id": 2488746,"Record function id": 0, "Ev Idx": 5545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940962.364, "dur": 2.284,
    "args": {
      "External id": 2488747,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 5546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940962.847, "dur": 1.395,
    "args": {
      "External id": 2488748,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 5547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940963.342, "dur": 0.796,
    "args": {
      "External id": 2488749,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 5548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940968.178, "dur": 3.462,
    "args": {
      "External id": 2488750,"Record function id": 0, "Ev Idx": 5549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940969.224, "dur": 2.033,
    "args": {
      "External id": 2488751,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 5550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940969.767, "dur": 1.083,
    "args": {
      "External id": 2488752,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 5551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940970.081, "dur": 0.707,
    "args": {
      "External id": 2488753,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 5552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940974.599, "dur": 3.880,
    "args": {
      "External id": 2488754,"Record function id": 0, "Ev Idx": 5553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940975.724, "dur": 2.346,
    "args": {
      "External id": 2488755,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 5554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940976.358, "dur": 1.211,
    "args": {
      "External id": 2488756,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 5555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657940976.804, "dur": 0.699,
    "args": {
      "External id": 2488757,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 5556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657941032.562, "dur": 8.219,
    "args": {
      "External id": 2488758,"Record function id": 0, "Ev Idx": 5557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657941034.912, "dur": 5.058,
    "args": {
      "External id": 2488759,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657941036.347, "dur": 2.654,
    "args": {
      "External id": 2488760,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657941037.404, "dur": 1.338,
    "args": {
      "External id": 2488761,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657941044.851, "dur": 3.615,
    "args": {
      "External id": 2488762,"Record function id": 0, "Ev Idx": 5561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657941045.940, "dur": 2.097,
    "args": {
      "External id": 2488763,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657941046.415, "dur": 1.179,
    "args": {
      "External id": 2488764,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657941046.807, "dur": 0.691,
    "args": {
      "External id": 2488765,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657941051.551, "dur": 5.522,
    "args": {
      "External id": 2488766,"Record function id": 0, "Ev Idx": 5565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657941052.900, "dur": 3.723,
    "args": {
      "External id": 2488767,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 5566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657941053.406, "dur": 2.719,
    "args": {
      "External id": 2488768,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 5567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657941054.056, "dur": 1.982,
    "args": {
      "External id": 2488769,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 5568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657941060.079, "dur": 3.911,
    "args": {
      "External id": 2488770,"Record function id": 0, "Ev Idx": 5569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657941061.079, "dur": 2.515,
    "args": {
      "External id": 2488771,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657941061.585, "dur": 1.587,
    "args": {
      "External id": 2488772,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657941062.108, "dur": 0.951,
    "args": {
      "External id": 2488773,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657941067.271, "dur": 37082.933,
    "args": {
      "External id": 2488774,"Record function id": 0, "Sequence number": 24740423, "Fwd thread id": 1, "Ev Idx": 5573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657941068.739, "dur": 37073.231,
    "args": {
      "External id": 2488775,"Sequence number": 24740423, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 5574
    }
  },
  {
    "ph": "f", "id": 225, "pid": 1336759, "tid": 1381173, "ts": 1514657941068.739,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.1)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657941099.159, "dur": 38.685,
    "args": {
      "External id": 2488776,"Record function id": 0, "Ev Idx": 5575
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.1)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657941146.208, "dur": 67.162,
    "args": {
      "External id": 2488777,"Record function id": 0, "Ev Idx": 5576
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.1)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657941218.905, "dur": 36914.776,
    "args": {
      "External id": 2488778,"Record function id": 0, "Ev Idx": 5577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657941312.385, "dur": 6.571,
    "args": {
      "External id": 2488779,"Record function id": 0, "Concrete Inputs": ["[51384320]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657941328.325, "dur": 5.023,
    "args": {
      "External id": 2488780,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 5579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657941348.098, "dur": 35978.835,
    "args": {
      "External id": 2488781,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 5580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657941360.818, "dur": 35956.778,
    "args": {
      "External id": 2488782,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 5581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657941400.968, "dur": 11.906,
    "args": {
      "External id": 2488783,"Record function id": 0, "Concrete Inputs": ["[12602]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657941419.007, "dur": 35861.393,
    "args": {
      "External id": 2488784,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], [], []], "Ev Idx": 5583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657941421.680, "dur": 35857.956,
    "args": {
      "External id": 2488785,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], []], "Ev Idx": 5584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657941425.818, "dur": 4.817,
    "args": {
      "External id": 2488786,"Record function id": 0, "Concrete Inputs": ["[12602]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657941432.468, "dur": 35843.725,
    "args": {
      "External id": 2488787,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[12602], [12602], []], "Ev Idx": 5586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657977415.105, "dur": 10.142,
    "args": {
      "External id": 2488788,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[51384320], [], [], [], [], []], "Ev Idx": 5587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657977419.270, "dur": 5.649,
    "args": {
      "External id": 2488789,"Record function id": 0, "Concrete Inputs": ["[6423040]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657977453.010, "dur": 353.287,
    "args": {
      "External id": 2488790,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[6423040], [51384320], [], [], [], []], "Ev Idx": 5589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657977479.480, "dur": 321.611,
    "args": {
      "External id": 2488791,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 6423040, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[51384320], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 5590, "In msg nelems": 51384320
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514657977490.596, "dur": 305.342,
    "args": {
      "External id": 2488792,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[51384320]], "Ev Idx": 5591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514657977826.567, "dur": 2.309,
    "args": {
      "External id": 2488793,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 5592, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657977886.126, "dur": 6.330,
    "args": {
      "External id": 2488794,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657977935.295, "dur": 1.656,
    "args": {
      "External id": 2488795,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "256"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657977951.769, "dur": 2.291,
    "args": {
      "External id": 2488796,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657977965.105, "dur": 0.727,
    "args": {
      "External id": 2488797,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1048832"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657977976.240, "dur": 0.951,
    "args": {
      "External id": 2488798,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1573120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657978034.408, "dur": 1.596,
    "args": {
      "External id": 2488799,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657978047.484, "dur": 2.124,
    "args": {
      "External id": 2488800,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "2097664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657978060.963, "dur": 1.289,
    "args": {
      "External id": 2488801,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "3539456"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657978071.204, "dur": 0.909,
    "args": {
      "External id": 2488802,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "4981248"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657978165.128, "dur": 2089.124,
    "args": {
      "External id": 2488803,"Record function id": 0, "Ev Idx": 5602
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_backward (model.layers.0)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657978184.112, "dur": 444.066,
    "args": {
      "External id": 2488804,"Record function id": 0, "Ev Idx": 5603
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.0)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657978199.627, "dur": 309.844,
    "args": {
      "External id": 2488805,"Record function id": 0, "Ev Idx": 5604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657978286.580, "dur": 4.250,
    "args": {
      "External id": 2488806,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 5605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657978294.109, "dur": 0.852,
    "args": {
      "External id": 2488807,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 5606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657978296.642, "dur": 0.940,
    "args": {
      "External id": 2488808,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657978299.159, "dur": 0.868,
    "args": {
      "External id": 2488809,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657978301.981, "dur": 2.599,
    "args": {
      "External id": 2488810,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657978307.047, "dur": 0.895,
    "args": {
      "External id": 2488811,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657978309.746, "dur": 0.683,
    "args": {
      "External id": 2488812,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 5611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657978312.011, "dur": 0.731,
    "args": {
      "External id": 2488813,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 5612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657978314.019, "dur": 1.022,
    "args": {
      "External id": 2488814,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 5613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657978317.549, "dur": 0.655,
    "args": {
      "External id": 2488815,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 5614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657978336.045, "dur": 146.112,
    "args": {
      "External id": 2488816,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 5615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657978353.034, "dur": 124.754,
    "args": {
      "External id": 2488817,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 5616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657978370.012, "dur": 12.211,
    "args": {
      "External id": 2488818,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657978385.296, "dur": 66.062,
    "args": {
      "External id": 2488819,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 5618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657978387.930, "dur": 63.062,
    "args": {
      "External id": 2488820,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 5619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657978391.785, "dur": 6.223,
    "args": {
      "External id": 2488821,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657978399.713, "dur": 50.786,
    "args": {
      "External id": 2488822,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 5621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657978636.636, "dur": 1597.527,
    "args": {
      "External id": 2488823,"Sequence number": 24740422, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 5622
    }
  },
  {
    "ph": "f", "id": 226, "pid": 1336759, "tid": 1381173, "ts": 1514657978636.636,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657978742.104, "dur": 102.859,
    "args": {
      "External id": 2488824,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [5632, 1], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 5623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_add_mul_rsub_silu_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657978886.071, "dur": 38.994,
    "args": {
      "External id": 2488825,"kernel_hash": "cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/m5/cm5os3r6hv7ysipdpiairo6lyepmpaszxgfbktm3djkrvim3pzmx.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [65536, 5632], [16, 4096, 5632], [16, 4096, 5632], [16, 4096, 5632], []], "Ev Idx": 5624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::bmm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657978941.481, "dur": 87.248,
    "args": {
      "External id": 2488826,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[0, 1, 2048], [0, 5632, 1], [11534336, 5632, 1]], "Input Dims": [[1, 2048, 65536], [1, 65536, 5632], [1, 2048, 5632]], "Ev Idx": 5625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657979043.030, "dur": 36.529,
    "args": {
      "External id": 2488827,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 5626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657979088.254, "dur": 45.536,
    "args": {
      "External id": 2488828,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 5627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657979141.601, "dur": 30.822,
    "args": {
      "External id": 2488829,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 5632], [2048, 1], [2048, 1]], "Input Dims": [[5632, 65536], [65536, 2048], [5632, 2048]], "Ev Idx": 5628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657979179.222, "dur": 42.473,
    "args": {
      "External id": 2488830,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 5629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657979245.740, "dur": 22.340,
    "args": {
      "External id": 2488831,"kernel_hash": "c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/47/c47x727mhg3ctiwzlmxagokbourh3jzbxm2pk4af4zajlvil6mbi.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 5630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514657979285.043, "dur": 28.435,
    "args": {
      "External id": 2488832,"kernel_hash": "cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "True", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/xr/cxrz26z2x5pfml2f2w22jccdkyygktk5hbwlaypdvvhkn4fqlm5d.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657979353.619, "dur": 20.066,
    "args": {
      "External id": 2488833,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 5632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657979387.649, "dur": 14.588,
    "args": {
      "External id": 2488834,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 5633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657979411.619, "dur": 27.668,
    "args": {
      "External id": 2488835,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 5634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657979442.497, "dur": 30.899,
    "args": {
      "External id": 2488836,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 5635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_backward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657979500.990, "dur": 194.575,
    "args": {
      "External id": 2488837,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar", "long int"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [131072, 4096, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [1]], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 32, 4096], [16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [2]], "Ev Idx": 5636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657979595.976, "dur": 6.775,
    "args": {
      "External id": 2488838,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657979604.754, "dur": 2.946,
    "args": {
      "External id": 2488839,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657979728.919, "dur": 25.029,
    "args": {
      "External id": 2488840,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514657979764.536, "dur": 14.998,
    "args": {
      "External id": 2488841,"kernel_hash": "c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "True"], "kernel_file": "/tmp/torchinductor_cvm/6o/c6otstk2ryxkcc2uxmrjmab44n43krf6sqcczamirphtzygmmb65.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657979787.583, "dur": 37.009,
    "args": {
      "External id": 2488842,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 5641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657979834.424, "dur": 34.464,
    "args": {
      "External id": 2488843,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 5642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657979875.510, "dur": 23.625,
    "args": {
      "External id": 2488844,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 5643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657979905.463, "dur": 29.903,
    "args": {
      "External id": 2488845,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 5644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657979941.647, "dur": 22.082,
    "args": {
      "External id": 2488846,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 2048], [2048, 1], [2048, 1]], "Input Dims": [[2048, 65536], [65536, 2048], [2048, 2048]], "Ev Idx": 5645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1381173,
    "ts": 1514657979970.604, "dur": 71.729,
    "args": {
      "External id": 2488847,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 5646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_4", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980062.407, "dur": 22.347,
    "args": {
      "External id": 2488848,"kernel_hash": "cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/g4/cg4ukr25hvc6fhe6tywzzdige4qswrvbmpfxtd3w73o2rbqqnxdc.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [65536, 2048], []], "Ev Idx": 5647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_bwd_kernel_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980101.301, "dur": 23.509,
    "args": {
      "External id": 2488849,"kernel_hash": "c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "65536", "2048", "1", "497", "132", "True", "2048", "False", "False", "True", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/4n/c4n3ou2ncjbisanhyo3hqtp5omwwxz2moubkkc2nbkm4ze4hfkis.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [2048], [65536, 2048], [65536, 2048], [132, 2048], [65536], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 5648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_red_fused_sum_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980138.334, "dur": 16.476,
    "args": {
      "External id": 2488850,"kernel_hash": "co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n", "grid": "grid(4096,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "4096", "66"], "kernel_file": "/tmp/torchinductor_cvm/o6/co6fecdb2uywcxfuxcz5nj3kv5higyqx2aburlbzi3bvynrhar7n.py", "kernel_backend": "triton", "Input type": ["float", "float", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [4096, 1, 2048], [], []], "Input Dims": [[132, 2048], [1, 2048, 2], [], []], "Ev Idx": 5649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_per_fused__to_copy_sum_3", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980167.592, "dur": 13.249,
    "args": {
      "External id": 2488851,"kernel_hash": "c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty", "grid": "grid(2048,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "2048", "2"], "kernel_file": "/tmp/torchinductor_cvm/35/c35ivrp4vzea2wu6fb3b63szlroo6fc5n7miteort3yhwxwmfsty.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 2048], [2048, 1], [], []], "Input Dims": [[1, 2048, 2], [1, 2048], [], []], "Ev Idx": 5650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_5", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980191.466, "dur": 15.509,
    "args": {
      "External id": 2488852,"kernel_hash": "cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u7/cu7quhizahhnunbe25uov4qu4xq6kfyaol7a4xbozhkfrtszumjz.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [8388608, 2048, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], [16, 4096, 2048], []], "Ev Idx": 5651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980276.280, "dur": 18.991,
    "args": {
      "External id": 2488853,"Record function id": 0, "Ev Idx": 5652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980279.626, "dur": 14.737,
    "args": {
      "External id": 2488854,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 5653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980286.660, "dur": 6.751,
    "args": {
      "External id": 2488855,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 5654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980288.280, "dur": 4.989,
    "args": {
      "External id": 2488856,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 5655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980299.118, "dur": 4.759,
    "args": {
      "External id": 2488857,"Record function id": 0, "Ev Idx": 5656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980300.560, "dur": 2.903,
    "args": {
      "External id": 2488858,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980301.279, "dur": 1.593,
    "args": {
      "External id": 2488859,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980301.868, "dur": 0.905,
    "args": {
      "External id": 2488860,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980307.091, "dur": 4.312,
    "args": {
      "External id": 2488861,"Record function id": 0, "Ev Idx": 5660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980308.587, "dur": 2.422,
    "args": {
      "External id": 2488862,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 5661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980309.108, "dur": 1.386,
    "args": {
      "External id": 2488863,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 5662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980309.413, "dur": 0.987,
    "args": {
      "External id": 2488864,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[2048]], "Ev Idx": 5663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980314.538, "dur": 3.687,
    "args": {
      "External id": 2488865,"Record function id": 0, "Ev Idx": 5664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980315.954, "dur": 1.876,
    "args": {
      "External id": 2488866,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 5665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980316.458, "dur": 0.975,
    "args": {
      "External id": 2488867,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 5666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980316.721, "dur": 0.617,
    "args": {
      "External id": 2488868,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 5667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980321.279, "dur": 3.677,
    "args": {
      "External id": 2488869,"Record function id": 0, "Ev Idx": 5668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980322.502, "dur": 2.039,
    "args": {
      "External id": 2488870,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 5669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980322.987, "dur": 0.987,
    "args": {
      "External id": 2488871,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 5670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980323.370, "dur": 0.536,
    "args": {
      "External id": 2488872,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 5671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980328.052, "dur": 3.906,
    "args": {
      "External id": 2488873,"Record function id": 0, "Ev Idx": 5672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980329.342, "dur": 2.199,
    "args": {
      "External id": 2488874,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980330.011, "dur": 1.129,
    "args": {
      "External id": 2488875,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980330.423, "dur": 0.596,
    "args": {
      "External id": 2488876,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980335.267, "dur": 3.633,
    "args": {
      "External id": 2488877,"Record function id": 0, "Ev Idx": 5676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980336.409, "dur": 2.069,
    "args": {
      "External id": 2488878,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980336.920, "dur": 1.125,
    "args": {
      "External id": 2488879,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980337.233, "dur": 0.719,
    "args": {
      "External id": 2488880,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980345.366, "dur": 5.303,
    "args": {
      "External id": 2488881,"Record function id": 0, "Ev Idx": 5680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980346.655, "dur": 3.603,
    "args": {
      "External id": 2488882,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 5681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980347.153, "dur": 2.666,
    "args": {
      "External id": 2488883,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 5682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980347.605, "dur": 2.122,
    "args": {
      "External id": 2488884,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 5683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980353.749, "dur": 3.762,
    "args": {
      "External id": 2488885,"Record function id": 0, "Ev Idx": 5684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980354.984, "dur": 2.121,
    "args": {
      "External id": 2488886,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980355.419, "dur": 1.256,
    "args": {
      "External id": 2488887,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980355.900, "dur": 0.662,
    "args": {
      "External id": 2488888,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 5687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980361.016, "dur": 37134.144,
    "args": {
      "External id": 2488889,"Record function id": 0, "Sequence number": 24740421, "Fwd thread id": 1, "Ev Idx": 5688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980362.296, "dur": 37124.393,
    "args": {
      "External id": 2488890,"Sequence number": 24740421, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 5689
    }
  },
  {
    "ph": "f", "id": 227, "pid": 1336759, "tid": 1381173, "ts": 1514657980362.296,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate (model.layers.0)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980387.968, "dur": 36.077,
    "args": {
      "External id": 2488891,"Record function id": 0, "Ev Idx": 5690
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard (model.layers.0)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980430.930, "dur": 65.916,
    "args": {
      "External id": 2488892,"Record function id": 0, "Ev Idx": 5691
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce (model.layers.0)", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980502.302, "dur": 36977.105,
    "args": {
      "External id": 2488893,"Record function id": 0, "Ev Idx": 5692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980620.963, "dur": 8.569,
    "args": {
      "External id": 2488894,"Record function id": 0, "Concrete Inputs": ["[51384320]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980640.516, "dur": 5.276,
    "args": {
      "External id": 2488895,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 5694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980660.452, "dur": 36051.221,
    "args": {
      "External id": 2488896,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 5695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980673.803, "dur": 36029.753,
    "args": {
      "External id": 2488897,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[[1], [2048, 1], [2048, 1], [2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], [], [], [6423040, 1]], "Input Dims": [[[2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], [], [], [8, 6423040]], "Ev Idx": 5696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980721.720, "dur": 13.795,
    "args": {
      "External id": 2488898,"Record function id": 0, "Concrete Inputs": ["[12602]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980745.644, "dur": 35920.364,
    "args": {
      "External id": 2488899,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], [], []], "Ev Idx": 5698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980748.015, "dur": 35917.315,
    "args": {
      "External id": 2488900,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[12602], [], [], [], [], [], []], "Ev Idx": 5699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980751.815, "dur": 5.175,
    "args": {
      "External id": 2488901,"Record function id": 0, "Concrete Inputs": ["[12602]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514657980758.704, "dur": 35903.595,
    "args": {
      "External id": 2488902,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[12602], [12602], []], "Ev Idx": 5701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514658016792.919, "dur": 8.032,
    "args": {
      "External id": 2488903,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[51384320], [], [], [], [], []], "Ev Idx": 5702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514658016795.403, "dur": 5.229,
    "args": {
      "External id": 2488904,"Record function id": 0, "Concrete Inputs": ["[6423040]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514658016829.104, "dur": 372.600,
    "args": {
      "External id": 2488905,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[6423040], [51384320], [], [], [], []], "Ev Idx": 5704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514658016854.564, "dur": 341.932,
    "args": {
      "External id": 2488906,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 6423040, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[51384320], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 5705, "In msg nelems": 51384320
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514658016865.151, "dur": 325.248,
    "args": {
      "External id": 2488907,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[51384320]], "Ev Idx": 5706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514658017221.575, "dur": 2.302,
    "args": {
      "External id": 2488908,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 5707, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658017280.839, "dur": 6.275,
    "args": {
      "External id": 2488909,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658017329.613, "dur": 1.580,
    "args": {
      "External id": 2488910,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "256"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658017346.788, "dur": 2.902,
    "args": {
      "External id": 2488911,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "524544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658017361.148, "dur": 0.902,
    "args": {
      "External id": 2488912,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1048832"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658017373.474, "dur": 0.870,
    "args": {
      "External id": 2488913,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "1573120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658017383.709, "dur": 0.808,
    "args": {
      "External id": 2488914,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658017395.025, "dur": 2.163,
    "args": {
      "External id": 2488915,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "2097664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658017407.636, "dur": 0.891,
    "args": {
      "External id": 2488916,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "3539456"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658017418.035, "dur": 0.967,
    "args": {
      "External id": 2488917,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "4981248"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514658017509.433, "dur": 276.561,
    "args": {
      "External id": 2488918,"Record function id": 0, "Sequence number": 24740420, "Fwd thread id": 1, "Ev Idx": 5717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunctionBackward", "pid": 1336759, "tid": 1381173,
    "ts": 1514658017512.142, "dur": 264.830,
    "args": {
      "External id": 2488919,"Sequence number": 24740420, "Fwd thread id": 1, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 5718
    }
  },
  {
    "ph": "f", "id": 228, "pid": 1336759, "tid": 1381173, "ts": 1514658017512.142,
    "cat": "fwdbwd", "name": "fwdbwd", "bp": "e"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_embedding_dense_backward_0", "pid": 1336759, "tid": 1381173,
    "ts": 1514658017643.021, "dur": 41.580,
    "args": {
      "External id": 2488920,"kernel_hash": "cqy2wwurxdpkuy27hkscegzor44hvby6k37yecnbppfr7quxx5hx", "grid": "grid(65536000,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "65536000"], "kernel_file": "/tmp/torchinductor_cvm/qy/cqy2wwurxdpkuy27hkscegzor44hvby6k37yecnbppfr7quxx5hx.py", "kernel_backend": "triton", "Input type": ["float", "Scalar"], "Input Strides": [[2048, 1], []], "Input Dims": [[32000, 2048], []], "Ev Idx": 5719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_embedding_dense_backward_1", "pid": 1336759, "tid": 1381173,
    "ts": 1514658017698.928, "dur": 26.599,
    "args": {
      "External id": 2488921,"kernel_hash": "cdyaclk3pd7t6z6zuqzswmelsxchlewwniqxs4434hgglihydkpz", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/dy/cdyaclk3pd7t6z6zuqzswmelsxchlewwniqxs4434hgglihydkpz.py", "kernel_backend": "triton", "Input type": ["long int", "c10::BFloat16", "float", "Scalar"], "Input Strides": [[4096, 1], [8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096], [16, 4096, 2048], [32000, 2048], []], "Ev Idx": 5720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_embedding_dense_backward_2", "pid": 1336759, "tid": 1381173,
    "ts": 1514658017742.003, "dur": 19.867,
    "args": {
      "External id": 2488922,"kernel_hash": "cvyxdor6u3emsju5qmfjq37syr77ijvjswkrldbg6t62lvcr3cem", "grid": "grid(65536000,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "65536000"], "kernel_file": "/tmp/torchinductor_cvm/vy/cvyxdor6u3emsju5qmfjq37syr77ijvjswkrldbg6t62lvcr3cem.py", "kernel_backend": "triton", "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 5721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "autograd::engine::evaluate_function: torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514658017797.707, "dur": 13.793,
    "args": {
      "External id": 2488923,"Record function id": 0, "Ev Idx": 5722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "torch::autograd::AccumulateGrad", "pid": 1336759, "tid": 1381173,
    "ts": 1514658017800.600, "dur": 10.036,
    "args": {
      "External id": 2488924,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 5723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514658017804.065, "dur": 5.603,
    "args": {
      "External id": 2488925,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 5724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1381173,
    "ts": 1514658017805.847, "dur": 3.644,
    "args": {
      "External id": 2488926,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 5725
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::root_post_backward_callback", "pid": 1336759, "tid": 1381173,
    "ts": 1514658017831.162, "dur": 9089.875,
    "args": {
      "External id": 2488927,"Record function id": 0, "Ev Idx": 5726
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_accumulate", "pid": 1336759, "tid": 1381173,
    "ts": 1514658017849.467, "dur": 35.512,
    "args": {
      "External id": 2488928,"Record function id": 0, "Ev Idx": 5727
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reshard", "pid": 1336759, "tid": 1381173,
    "ts": 1514658017891.259, "dur": 271.702,
    "args": {
      "External id": 2488929,"Record function id": 0, "Ev Idx": 5728
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_backward_reduce", "pid": 1336759, "tid": 1381173,
    "ts": 1514658018171.780, "dur": 8490.103,
    "args": {
      "External id": 2488930,"Record function id": 0, "Ev Idx": 5729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514658018293.549, "dur": 7.026,
    "args": {
      "External id": 2488931,"Record function id": 0, "Concrete Inputs": ["[336611328]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1381173,
    "ts": 1514658018311.392, "dur": 5.058,
    "args": {
      "External id": 2488932,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[336611328], []], "Ev Idx": 5731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514658018341.595, "dur": 7039.799,
    "args": {
      "External id": 2488933,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[], [], [], [42076416, 1]], "Input Dims": [[], [], [], [8, 42076416]], "Ev Idx": 5732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_chunk_cat", "pid": 1336759, "tid": 1381173,
    "ts": 1514658018360.523, "dur": 7008.078,
    "args": {
      "External id": 2488934,"Record function id": 0, "Concrete Inputs": ["", "0", "8", ""], "Input type": ["TensorList", "Scalar", "Scalar", "float"], "Input Strides": [[], [], [], [42076416, 1]], "Input Dims": [[], [], [], [8, 42076416]], "Ev Idx": 5733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514658018651.220, "dur": 21.209,
    "args": {
      "External id": 2488935,"Record function id": 0, "Concrete Inputs": ["[82421]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1381173,
    "ts": 1514658018741.270, "dur": 6576.458,
    "args": {
      "External id": 2488936,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[82421], [], [], [], [], [], [], []], "Ev Idx": 5735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1381173,
    "ts": 1514658018744.930, "dur": 6571.653,
    "args": {
      "External id": 2488937,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[82421], [], [], [], [], [], []], "Ev Idx": 5736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658018749.626, "dur": 13.982,
    "args": {
      "External id": 2488938,"Record function id": 0, "Concrete Inputs": ["[82421]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1381173,
    "ts": 1514658018765.676, "dur": 6545.476,
    "args": {
      "External id": 2488939,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[82421], [82421], []], "Ev Idx": 5738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514658025510.465, "dur": 10.890,
    "args": {
      "External id": 2488940,"Record function id": 0, "Concrete Inputs": ["", "[42076416]", "", "", "", "False"], "Input type": ["float", "ScalarList", "", "", "", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[336611328], [], [], [], [], []], "Ev Idx": 5739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1381173,
    "ts": 1514658025513.612, "dur": 7.330,
    "args": {
      "External id": 2488941,"Record function id": 0, "Concrete Inputs": ["[42076416]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_reduce_scatter_base_", "pid": 1336759, "tid": 1381173,
    "ts": 1514658025564.340, "dur": 462.343,
    "args": {
      "External id": 2488942,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "-1"], "Input type": ["float", "float", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], [], []], "Input Dims": [[42076416], [336611328], [], [], [], []], "Ev Idx": 5741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514658025590.534, "dur": 430.992,
    "args": {
      "External id": 2488943,"Record function id": 0, "Collective name": "_reduce_scatter_base", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 42076416, "Process Group Name": "0", "Input type": ["float", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[336611328], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 5742, "In msg nelems": 336611328
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_reduce_scatter_base", "pid": 1336759, "tid": 1381173,
    "ts": 1514658025606.707, "dur": 407.226,
    "args": {
      "External id": 2488944,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[336611328]], "Ev Idx": 5743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026046.632, "dur": 2.753,
    "args": {
      "External id": 2488945,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 5744, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026109.121, "dur": 7.233,
    "args": {
      "External id": 2488946,"Record function id": 0, "Concrete Inputs": ["", "[4000, 2048]", "[2048, 1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026159.011, "dur": 1.251,
    "args": {
      "External id": 2488947,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "8192000"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026174.774, "dur": 1.372,
    "args": {
      "External id": 2488948,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "8192256"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026187.764, "dur": 1.934,
    "args": {
      "External id": 2488949,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "8716544"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026200.573, "dur": 1.026,
    "args": {
      "External id": 2488950,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "9240832"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026210.920, "dur": 1.291,
    "args": {
      "External id": 2488951,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "9765120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026224.599, "dur": 0.994,
    "args": {
      "External id": 2488952,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "10289408"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026236.503, "dur": 2.388,
    "args": {
      "External id": 2488953,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "10289664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026249.548, "dur": 1.037,
    "args": {
      "External id": 2488954,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "11731456"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026259.964, "dur": 1.369,
    "args": {
      "External id": 2488955,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "13173248"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026270.779, "dur": 1.498,
    "args": {
      "External id": 2488956,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "14615040"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026281.281, "dur": 2.827,
    "args": {
      "External id": 2488957,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "14615296"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026293.195, "dur": 1.354,
    "args": {
      "External id": 2488958,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "15139584"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026304.404, "dur": 0.936,
    "args": {
      "External id": 2488959,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "15663872"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026314.623, "dur": 1.054,
    "args": {
      "External id": 2488960,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "16188160"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026324.655, "dur": 2.867,
    "args": {
      "External id": 2488961,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "16712448"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026337.274, "dur": 1.161,
    "args": {
      "External id": 2488962,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "16712704"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026347.571, "dur": 1.240,
    "args": {
      "External id": 2488963,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "18154496"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026358.079, "dur": 1.138,
    "args": {
      "External id": 2488964,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "19596288"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026368.564, "dur": 3.272,
    "args": {
      "External id": 2488965,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "21038080"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026380.985, "dur": 1.529,
    "args": {
      "External id": 2488966,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "21038336"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026391.556, "dur": 1.179,
    "args": {
      "External id": 2488967,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "21562624"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026402.445, "dur": 1.483,
    "args": {
      "External id": 2488968,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "22086912"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026412.574, "dur": 2.421,
    "args": {
      "External id": 2488969,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "22611200"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026425.357, "dur": 1.231,
    "args": {
      "External id": 2488970,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "23135488"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026435.713, "dur": 1.496,
    "args": {
      "External id": 2488971,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "23135744"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026446.641, "dur": 1.340,
    "args": {
      "External id": 2488972,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "24577536"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026457.678, "dur": 2.300,
    "args": {
      "External id": 2488973,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "26019328"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026469.088, "dur": 1.226,
    "args": {
      "External id": 2488974,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "27461120"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026479.697, "dur": 1.178,
    "args": {
      "External id": 2488975,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "27461376"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026490.570, "dur": 1.012,
    "args": {
      "External id": 2488976,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "27985664"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026500.780, "dur": 2.130,
    "args": {
      "External id": 2488977,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "28509952"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026511.880, "dur": 1.100,
    "args": {
      "External id": 2488978,"Record function id": 0, "Concrete Inputs": ["", "[256, 2048]", "[2048, 1]", "29034240"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026521.723, "dur": 1.572,
    "args": {
      "External id": 2488979,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "29558528"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026549.886, "dur": 2.061,
    "args": {
      "External id": 2488980,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "29558784"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026563.202, "dur": 2.153,
    "args": {
      "External id": 2488981,"Record function id": 0, "Concrete Inputs": ["", "[704, 2048]", "[2048, 1]", "31000576"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026575.302, "dur": 1.445,
    "args": {
      "External id": 2488982,"Record function id": 0, "Concrete Inputs": ["", "[256, 5632]", "[5632, 1]", "32442368"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026586.398, "dur": 1.612,
    "args": {
      "External id": 2488983,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "33884160"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1381173,
    "ts": 1514658026597.851, "dur": 1.135,
    "args": {
      "External id": 2488984,"Record function id": 0, "Concrete Inputs": ["", "[4000, 2048]", "[2048, 1]", "33884416"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5783
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "ProfilerStep#102911", "pid": 1336759, "tid": 1336759,
    "ts": 1514655936770.986, "dur": 2107992.644,
    "args": {
      "External id": 2478081,"Record function id": 0, "Ev Idx": 5784
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "Optimizer.zero_grad#AdamW.zero_grad", "pid": 1336759, "tid": 1336759,
    "ts": 1514655936802.432, "dur": 553.818,
    "args": {
      "External id": 2478082,"Record function id": 0, "Ev Idx": 5785
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "enumerate(DataLoader)#_StatefulMultiProcessingDataLoaderIter.__next__", "pid": 1336759, "tid": 1336759,
    "ts": 1514655937397.093, "dur": 1955.070,
    "args": {
      "External id": 2478083,"Record function id": 0, "Ev Idx": 5786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514655938389.648, "dur": 7.281,
    "args": {
      "External id": 2478084,"Record function id": 0, "Concrete Inputs": ["[0]", "4", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::set_", "pid": 1336759, "tid": 1336759,
    "ts": 1514655938418.202, "dur": 7.245,
    "args": {
      "External id": 2478085,"Sequence number": 24740420, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "0", "[16, 8192]", "[8192, 1]"], "Input type": ["long int", "", "Scalar", "ScalarList", "ScalarList"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[0], [], [], [], []], "Ev Idx": 5788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514655938824.134, "dur": 2.275,
    "args": {
      "External id": 2478086,"Record function id": 0, "Concrete Inputs": ["[0]", "4", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::set_", "pid": 1336759, "tid": 1336759,
    "ts": 1514655938833.088, "dur": 2.334,
    "args": {
      "External id": 2478087,"Sequence number": 24740420, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "0", "[16, 8192]", "[8192, 1]"], "Input type": ["long int", "", "Scalar", "ScalarList", "ScalarList"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[0], [], [], [], []], "Ev Idx": 5790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514655939252.744, "dur": 2.126,
    "args": {
      "External id": 2478088,"Record function id": 0, "Concrete Inputs": ["[0]", "4", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::set_", "pid": 1336759, "tid": 1336759,
    "ts": 1514655939260.911, "dur": 2.405,
    "args": {
      "External id": 2478089,"Sequence number": 24740420, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "0", "[16, 8192]", "[8192, 1]"], "Input type": ["long int", "", "Scalar", "ScalarList", "ScalarList"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[0], [], [], [], []], "Ev Idx": 5792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514655940101.542, "dur": 14.073,
    "args": {
      "External id": 2478090,"Sequence number": 24740420, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], [], [], []], "Input Dims": [[16, 8192], [], [], [], []], "Ev Idx": 5793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655940109.279, "dur": 2.649,
    "args": {
      "External id": 2478091,"Record function id": 0, "Concrete Inputs": ["", "[16, 8192]", "[8192, 1]", "0"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1], [], [], []], "Input Dims": [[16, 8192], [], [], []], "Ev Idx": 5794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514655940117.229, "dur": 4.153,
    "args": {
      "External id": 2478092,"Sequence number": 24740420, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "4096", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], [], [], []], "Input Dims": [[16, 8192], [], [], [], []], "Ev Idx": 5795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655940119.197, "dur": 1.105,
    "args": {
      "External id": 2478093,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096]", "[8192, 1]", "0"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1], [], [], []], "Input Dims": [[16, 8192], [], [], []], "Ev Idx": 5796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514655940146.644, "dur": 551.550,
    "args": {
      "External id": 2478094,"Sequence number": 24740420, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "4", "0", "", "", "False", "False", ""], "Input type": ["long int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], [], [], [], []], "Input Dims": [[16, 4096], [], [], [], [], [], [], []], "Ev Idx": 5797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514655940154.398, "dur": 543.112,
    "args": {
      "External id": 2478095,"Sequence number": 24740420, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "4", "0", "", "", "False", ""], "Input type": ["long int", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], [], [], []], "Input Dims": [[16, 4096], [], [], [], [], [], []], "Ev Idx": 5798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655940161.858, "dur": 9.707,
    "args": {
      "External id": 2478096,"Record function id": 0, "Concrete Inputs": ["[16, 4096]", "[4096, 1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514655940174.300, "dur": 521.313,
    "args": {
      "External id": 2478097,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[4096, 1], [8192, 1], []], "Input Dims": [[16, 4096], [16, 4096], []], "Ev Idx": 5800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514655940184.026, "dur": 0.312,
    "args": {
      "External id": 2478098,"Record function id": 0, "Concrete Inputs": ["", "4", "False", "False", ""], "Input type": ["long int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], []], "Input Dims": [[16, 4096], [], [], [], []], "Ev Idx": 5801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514655940187.805, "dur": 8.792,
    "args": {
      "External id": 2478099,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["long int", "long int"], "Input Strides": [[8192, 1], [4096, 1]], "Input Dims": [[16, 4096], [16, 4096]], "Ev Idx": 5802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 1336759, "tid": 1336759,
    "ts": 1514655940192.239, "dur": 4.205,
    "args": {
      "External id": 2478100,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096]", "False"], "Input type": ["long int", "ScalarList", "Scalar"], "Input Strides": [[8192, 1], [], []], "Input Dims": [[16, 4096], [], []], "Ev Idx": 5803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655940195.655, "dur": 0.536,
    "args": {
      "External id": 2478101,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096]", "[8192, 1]", ""], "Input type": ["long int", "ScalarList", "ScalarList", ""], "Input Strides": [[8192, 1], [], [], []], "Input Dims": [[16, 4096], [], [], []], "Ev Idx": 5804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 1336759, "tid": 1336759,
    "ts": 1514655940199.149, "dur": 247.417,
    "args": {
      "External id": 2478102,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["long int", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[16, 4096], []], "Ev Idx": 5805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1336759, "tid": 1336759,
    "ts": 1514655940202.168, "dur": 244.044,
    "args": {
      "External id": 2478103,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["long int", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[16, 4096], []], "Ev Idx": 5806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514655940204.643, "dur": 14.967,
    "args": {
      "External id": 2478104,"Record function id": 0, "Concrete Inputs": ["", "4", "0", "", "", "0"], "Input type": ["long int", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[8192, 1], [], [], [], [], []], "Input Dims": [[16, 4096], [], [], [], [], []], "Ev Idx": 5807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514655940206.843, "dur": 12.223,
    "args": {
      "External id": 2478105,"Record function id": 0, "Concrete Inputs": ["[16, 4096]", "4", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514655940220.503, "dur": 224.929,
    "args": {
      "External id": 2478106,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[4096, 1], [8192, 1], []], "Input Dims": [[16, 4096], [16, 4096], []], "Ev Idx": 5809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514655940448.734, "dur": 242.913,
    "args": {
      "External id": 2478107,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[4096, 1], [4096, 1], []], "Input Dims": [[16, 4096], [16, 4096], []], "Ev Idx": 5810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514655940717.300, "dur": 695.217,
    "args": {
      "External id": 2478108,"Sequence number": 24740420, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "4", "0", "", "", "False", "False", ""], "Input type": ["long int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], [], [], [], []], "Input Dims": [[16, 8192], [], [], [], [], [], [], []], "Ev Idx": 5811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514655940719.270, "dur": 692.654,
    "args": {
      "External id": 2478109,"Sequence number": 24740420, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "4", "0", "", "", "False", ""], "Input type": ["long int", "Scalar", "Scalar", "", "", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], [], [], []], "Input Dims": [[16, 8192], [], [], [], [], [], []], "Ev Idx": 5812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655940726.331, "dur": 8.947,
    "args": {
      "External id": 2478110,"Record function id": 0, "Concrete Inputs": ["[16, 8192]", "[8192, 1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514655940736.552, "dur": 670.804,
    "args": {
      "External id": 2478111,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[8192, 1], [8192, 1], []], "Input Dims": [[16, 8192], [16, 8192], []], "Ev Idx": 5814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::arange", "pid": 1336759, "tid": 1336759,
    "ts": 1514655941445.274, "dur": 52.906,
    "args": {
      "External id": 2478112,"Record function id": 0, "Concrete Inputs": ["0", "4096", "", "", "", "False"], "Input type": ["Scalar", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514655941451.523, "dur": 5.913,
    "args": {
      "External id": 2478113,"Record function id": 0, "Concrete Inputs": ["[0]", "4", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::arange", "pid": 1336759, "tid": 1336759,
    "ts": 1514655941460.403, "dur": 37.461,
    "args": {
      "External id": 2478114,"Record function id": 0, "Concrete Inputs": ["0", "4096", "1", ""], "Input type": ["Scalar", "Scalar", "Scalar", "long int"], "Input Strides": [[], [], [], [1]], "Input Dims": [[], [], [], [0]], "Ev Idx": 5817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514655941466.114, "dur": 6.373,
    "args": {
      "External id": 2478115,"Record function id": 0, "Concrete Inputs": ["", "[4096]", ""], "Input type": ["long int", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 5818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::repeat", "pid": 1336759, "tid": 1336759,
    "ts": 1514655941510.347, "dur": 88.926,
    "args": {
      "External id": 2478116,"Sequence number": 24740420, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 1]"], "Input type": ["long int", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4096], []], "Ev Idx": 5819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 1336759, "tid": 1336759,
    "ts": 1514655941515.615, "dur": 6.736,
    "args": {
      "External id": 2478117,"Record function id": 0, "Concrete Inputs": ["", "[1, 4096]", "False"], "Input type": ["long int", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[4096], [], []], "Ev Idx": 5820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655941520.455, "dur": 1.598,
    "args": {
      "External id": 2478118,"Record function id": 0, "Concrete Inputs": ["", "[1, 4096]", "[4096, 1]", ""], "Input type": ["long int", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[4096], [], [], []], "Ev Idx": 5821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514655941523.524, "dur": 21.786,
    "args": {
      "External id": 2478119,"Record function id": 0, "Concrete Inputs": ["[16, 4096]", "4", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 1336759, "tid": 1336759,
    "ts": 1514655941548.373, "dur": 3.080,
    "args": {
      "External id": 2478120,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Strides": [[4096, 1]], "Input Dims": [[16, 4096]], "Ev Idx": 5823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unfold", "pid": 1336759, "tid": 1336759,
    "ts": 1514655941556.119, "dur": 5.345,
    "args": {
      "External id": 2478121,"Record function id": 0, "Concrete Inputs": ["", "0", "1", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4096, 1], [], [], []], "Input Dims": [[16, 4096], [], [], []], "Ev Idx": 5824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655941560.511, "dur": 0.794,
    "args": {
      "External id": 2478122,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 1]", "[4096, 1, 4096]", ""], "Input type": ["long int", "ScalarList", "ScalarList", ""], "Input Strides": [[4096, 1], [], [], []], "Input Dims": [[16, 4096], [], [], []], "Ev Idx": 5825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unfold", "pid": 1336759, "tid": 1336759,
    "ts": 1514655941562.329, "dur": 4.433,
    "args": {
      "External id": 2478123,"Record function id": 0, "Concrete Inputs": ["", "1", "4096", "4096"], "Input type": ["long int", "Scalar", "Scalar", "Scalar"], "Input Strides": [[4096, 1, 4096], [], [], []], "Input Dims": [[16, 4096, 1], [], [], []], "Ev Idx": 5826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655941565.716, "dur": 0.954,
    "args": {
      "External id": 2478124,"Record function id": 0, "Concrete Inputs": ["", "[16, 1, 1, 4096]", "[4096, 4096, 4096, 1]", ""], "Input type": ["long int", "ScalarList", "ScalarList", ""], "Input Strides": [[4096, 1, 4096], [], [], []], "Input Dims": [[16, 4096, 1], [], [], []], "Ev Idx": 5827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514655941568.433, "dur": 3.535,
    "args": {
      "External id": 2478125,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["long int", "long int"], "Input Strides": [[4096, 1], [4096, 4096, 4096, 1]], "Input Dims": [[1, 4096], [16, 1, 1, 4096]], "Ev Idx": 5828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::expand", "pid": 1336759, "tid": 1336759,
    "ts": 1514655941569.318, "dur": 2.551,
    "args": {
      "External id": 2478126,"Record function id": 0, "Concrete Inputs": ["", "[16, 1, 1, 4096]", "False"], "Input type": ["long int", "ScalarList", "Scalar"], "Input Strides": [[4096, 1], [], []], "Input Dims": [[1, 4096], [], []], "Ev Idx": 5829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655941571.056, "dur": 0.680,
    "args": {
      "External id": 2478127,"Record function id": 0, "Concrete Inputs": ["", "[16, 1, 1, 4096]", "[0, 4096, 4096, 1]", ""], "Input type": ["long int", "ScalarList", "ScalarList", ""], "Input Strides": [[4096, 1], [], [], []], "Input Dims": [[1, 4096], [], [], []], "Ev Idx": 5830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514655941572.997, "dur": 25.675,
    "args": {
      "External id": 2478128,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[4096, 4096, 4096, 1], [0, 4096, 4096, 1], []], "Input Dims": [[16, 1, 1, 4096], [16, 1, 1, 4096], []], "Ev Idx": 5831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514655941607.888, "dur": 31.992,
    "args": {
      "External id": 2478129,"Sequence number": 24740420, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "False", "False", ""], "Input type": ["long int", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[4096, 1], [], [], [], []], "Input Dims": [[16, 4096], [], [], [], []], "Ev Idx": 5832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514655941609.373, "dur": 30.284,
    "args": {
      "External id": 2478130,"Sequence number": 24740420, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "3", "", "", "", "False", ""], "Input type": ["long int", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[4096, 1], [], [], [], [], [], []], "Input Dims": [[16, 4096], [], [], [], [], [], []], "Ev Idx": 5833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655941616.579, "dur": 3.375,
    "args": {
      "External id": 2478131,"Record function id": 0, "Concrete Inputs": ["[16, 4096]", "[4096, 1]", "3", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514655941620.726, "dur": 18.497,
    "args": {
      "External id": 2478132,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["int", "long int", "Scalar"], "Input Strides": [[4096, 1], [4096, 1], []], "Input Dims": [[16, 4096], [16, 4096], []], "Ev Idx": 5835
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::root_pre_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514655941749.102, "dur": 141.386,
    "args": {
      "External id": 2478133,"Record function id": 0, "Ev Idx": 5836
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::inputs_to_device", "pid": 1336759, "tid": 1336759,
    "ts": 1514655941826.294, "dur": 53.529,
    "args": {
      "External id": 2478134,"Record function id": 0, "Ev Idx": 5837
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336759, "tid": 1336759,
    "ts": 1514655941897.788, "dur": 42.759,
    "args": {
      "External id": 2478135,"Record function id": 0, "Ev Idx": 5838
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514655941948.252, "dur": 10305.546,
    "args": {
      "External id": 2478136,"Record function id": 0, "Ev Idx": 5839
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather", "pid": 1336759, "tid": 1336759,
    "ts": 1514655941956.971, "dur": 1256.032,
    "args": {
      "External id": 2478137,"Record function id": 0, "Ev Idx": 5840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942091.218, "dur": 7.412,
    "args": {
      "External id": 2478138,"Record function id": 0, "Concrete Inputs": ["[42076416]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942114.738, "dur": 112.302,
    "args": {
      "External id": 2478139,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["c10::BFloat16", "", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[42076416], [], []], "Ev Idx": 5842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942119.278, "dur": 1.735,
    "args": {
      "External id": 2478140,"Record function id": 0, "Concrete Inputs": ["", "[8192000]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942126.318, "dur": 0.505,
    "args": {
      "External id": 2478141,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "8192000"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942127.459, "dur": 0.533,
    "args": {
      "External id": 2478142,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "8192256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942128.754, "dur": 1.689,
    "args": {
      "External id": 2478143,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "8716544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942134.549, "dur": 0.637,
    "args": {
      "External id": 2478144,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "9240832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942136.192, "dur": 0.264,
    "args": {
      "External id": 2478145,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "9765120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942137.301, "dur": 1.820,
    "args": {
      "External id": 2478146,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "10289408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942141.642, "dur": 0.181,
    "args": {
      "External id": 2478147,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "10289664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942142.291, "dur": 0.156,
    "args": {
      "External id": 2478148,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "11731456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942147.122, "dur": 0.310,
    "args": {
      "External id": 2478149,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "13173248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942148.028, "dur": 0.497,
    "args": {
      "External id": 2478150,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "14615040"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942149.572, "dur": 1.792,
    "args": {
      "External id": 2478151,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "14615296"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942154.552, "dur": 0.379,
    "args": {
      "External id": 2478152,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "15139584"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942156.045, "dur": 0.160,
    "args": {
      "External id": 2478153,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "15663872"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942156.854, "dur": 1.954,
    "args": {
      "External id": 2478154,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "16188160"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942161.115, "dur": 0.312,
    "args": {
      "External id": 2478155,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "16712448"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942161.910, "dur": 0.181,
    "args": {
      "External id": 2478156,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "16712704"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942166.172, "dur": 0.173,
    "args": {
      "External id": 2478157,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "18154496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942167.048, "dur": 0.389,
    "args": {
      "External id": 2478158,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "19596288"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942168.289, "dur": 1.264,
    "args": {
      "External id": 2478159,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "21038080"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942172.087, "dur": 0.524,
    "args": {
      "External id": 2478160,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "21038336"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942173.974, "dur": 0.159,
    "args": {
      "External id": 2478161,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "21562624"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942174.849, "dur": 1.794,
    "args": {
      "External id": 2478162,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "22086912"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942178.851, "dur": 0.158,
    "args": {
      "External id": 2478163,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "22611200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942179.689, "dur": 0.153,
    "args": {
      "External id": 2478164,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "23135488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942184.440, "dur": 0.152,
    "args": {
      "External id": 2478165,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "23135744"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942185.240, "dur": 0.527,
    "args": {
      "External id": 2478166,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "24577536"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942186.237, "dur": 1.140,
    "args": {
      "External id": 2478167,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "26019328"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942190.622, "dur": 0.383,
    "args": {
      "External id": 2478168,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "27461120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942191.896, "dur": 0.157,
    "args": {
      "External id": 2478169,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "27461376"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942192.685, "dur": 1.938,
    "args": {
      "External id": 2478170,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "27985664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942196.972, "dur": 0.267,
    "args": {
      "External id": 2478171,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "28509952"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942197.896, "dur": 0.382,
    "args": {
      "External id": 2478172,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "29034240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942202.470, "dur": 0.270,
    "args": {
      "External id": 2478173,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "29558528"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942203.365, "dur": 0.393,
    "args": {
      "External id": 2478174,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "29558784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942204.460, "dur": 1.722,
    "args": {
      "External id": 2478175,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "31000576"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942208.629, "dur": 0.370,
    "args": {
      "External id": 2478176,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "32442368"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942210.016, "dur": 0.156,
    "args": {
      "External id": 2478177,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "33884160"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942210.827, "dur": 1.884,
    "args": {
      "External id": 2478178,"Record function id": 0, "Concrete Inputs": ["", "[8192000]", "[1]", "33884416"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942247.481, "dur": 81.124,
    "args": {
      "External id": 2478179,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 5882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942388.097, "dur": 257.787,
    "args": {
      "External id": 2478180,"Record function id": 0, "Concrete Inputs": ["", "", "42076416", "8", "6", "15", ""], "Input type": ["TensorList", "", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], []], "Ev Idx": 5883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942403.472, "dur": 4.120,
    "args": {
      "External id": 2478181,"Record function id": 0, "Concrete Inputs": ["[336611328]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942413.792, "dur": 11.628,
    "args": {
      "External id": 2478182,"Record function id": 0, "Concrete Inputs": ["", "0", "252458496", "42076416"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[336611328], [], [], []], "Ev Idx": 5885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942418.114, "dur": 6.862,
    "args": {
      "External id": 2478183,"Record function id": 0, "Concrete Inputs": ["", "0", "252458496", "294534912", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[336611328], [], [], [], []], "Ev Idx": 5886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942422.260, "dur": 0.588,
    "args": {
      "External id": 2478184,"Record function id": 0, "Concrete Inputs": ["", "[42076416]", "[1]", "252458496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[336611328], [], [], []], "Ev Idx": 5887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942433.287, "dur": 114.523,
    "args": {
      "External id": 2478185,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["c10::BFloat16", "", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[42076416], [], []], "Ev Idx": 5888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942434.945, "dur": 0.242,
    "args": {
      "External id": 2478186,"Record function id": 0, "Concrete Inputs": ["", "[8192000]", "[1]", "252458496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942436.657, "dur": 0.477,
    "args": {
      "External id": 2478187,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "260650496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942439.288, "dur": 1.433,
    "args": {
      "External id": 2478188,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "260650752"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942441.821, "dur": 0.586,
    "args": {
      "External id": 2478189,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "261175040"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942444.539, "dur": 0.195,
    "args": {
      "External id": 2478190,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "261699328"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942445.505, "dur": 0.284,
    "args": {
      "External id": 2478191,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "262223616"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942446.505, "dur": 0.198,
    "args": {
      "External id": 2478192,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "262747904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942449.168, "dur": 0.237,
    "args": {
      "External id": 2478193,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "262748160"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942450.098, "dur": 0.597,
    "args": {
      "External id": 2478194,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "264189952"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942451.589, "dur": 0.470,
    "args": {
      "External id": 2478195,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "265631744"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942454.356, "dur": 1.058,
    "args": {
      "External id": 2478196,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "267073536"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942456.060, "dur": 0.515,
    "args": {
      "External id": 2478197,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "267073792"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942458.735, "dur": 1.871,
    "args": {
      "External id": 2478198,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "267598080"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942461.562, "dur": 0.168,
    "args": {
      "External id": 2478199,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "268122368"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942462.440, "dur": 0.467,
    "args": {
      "External id": 2478200,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "268646656"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942466.963, "dur": 0.248,
    "args": {
      "External id": 2478201,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "269170944"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942467.866, "dur": 0.426,
    "args": {
      "External id": 2478202,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "269171200"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942469.284, "dur": 0.347,
    "args": {
      "External id": 2478203,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "270612992"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942473.266, "dur": 1.403,
    "args": {
      "External id": 2478204,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "272054784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942475.610, "dur": 0.366,
    "args": {
      "External id": 2478205,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "273496576"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942478.036, "dur": 1.897,
    "args": {
      "External id": 2478206,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "273496832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942480.880, "dur": 0.476,
    "args": {
      "External id": 2478207,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "274021120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942481.962, "dur": 0.527,
    "args": {
      "External id": 2478208,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "274545408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942486.473, "dur": 0.144,
    "args": {
      "External id": 2478209,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "275069696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942487.071, "dur": 0.360,
    "args": {
      "External id": 2478210,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "275593984"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942488.040, "dur": 0.146,
    "args": {
      "External id": 2478211,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "275594240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942492.435, "dur": 1.003,
    "args": {
      "External id": 2478212,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "277036032"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942494.277, "dur": 0.516,
    "args": {
      "External id": 2478213,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "278477824"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942496.743, "dur": 2.028,
    "args": {
      "External id": 2478214,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "279919616"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942499.429, "dur": 0.474,
    "args": {
      "External id": 2478215,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "279919872"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942500.577, "dur": 0.462,
    "args": {
      "External id": 2478216,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "280444160"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942505.418, "dur": 0.246,
    "args": {
      "External id": 2478217,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "280968448"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942506.390, "dur": 0.152,
    "args": {
      "External id": 2478218,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "281492736"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942507.159, "dur": 0.173,
    "args": {
      "External id": 2478219,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "282017024"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942511.089, "dur": 1.222,
    "args": {
      "External id": 2478220,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "282017280"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942512.980, "dur": 0.151,
    "args": {
      "External id": 2478221,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "283459072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942515.742, "dur": 2.023,
    "args": {
      "External id": 2478222,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "284900864"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942518.651, "dur": 0.355,
    "args": {
      "External id": 2478223,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "286342656"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942519.615, "dur": 0.161,
    "args": {
      "External id": 2478224,"Record function id": 0, "Concrete Inputs": ["", "[8192000]", "[1]", "286342912"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[42076416], [], [], []], "Ev Idx": 5927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942570.047, "dur": 60.562,
    "args": {
      "External id": 2478225,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 5928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942702.487, "dur": 402.375,
    "args": {
      "External id": 2478226,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[336611328], [42076416], [], [], []], "Ev Idx": 5929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942740.984, "dur": 358.021,
    "args": {
      "External id": 2478227,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 336611328, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[42076416], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 5930, "In msg nelems": 42076416
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1336759,
    "ts": 1514655942751.446, "dur": 341.215,
    "args": {
      "External id": 2478228,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[42076416]], "Ev Idx": 5931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943130.880, "dur": 2.466,
    "args": {
      "External id": 2478229,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 5932, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943228.653, "dur": 8860.512,
    "args": {
      "External id": 2478230,"Record function id": 0, "Ev Idx": 5933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943392.554, "dur": 6.753,
    "args": {
      "External id": 2478231,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[336611328], []], "Ev Idx": 5934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943402.847, "dur": 1.199,
    "args": {
      "External id": 2478232,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[65536000], []], "Ev Idx": 5935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943405.840, "dur": 0.730,
    "args": {
      "External id": 2478233,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 5936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943408.427, "dur": 1.903,
    "args": {
      "External id": 2478234,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943411.584, "dur": 0.841,
    "args": {
      "External id": 2478235,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943415.655, "dur": 0.702,
    "args": {
      "External id": 2478236,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943417.890, "dur": 0.765,
    "args": {
      "External id": 2478237,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943419.923, "dur": 1.628,
    "args": {
      "External id": 2478238,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 5941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943424.886, "dur": 0.644,
    "args": {
      "External id": 2478239,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 5942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943428.160, "dur": 0.401,
    "args": {
      "External id": 2478240,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 5943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943430.094, "dur": 0.535,
    "args": {
      "External id": 2478241,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 5944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943431.859, "dur": 1.556,
    "args": {
      "External id": 2478242,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 5945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943435.151, "dur": 0.494,
    "args": {
      "External id": 2478243,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943438.577, "dur": 0.604,
    "args": {
      "External id": 2478244,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943440.372, "dur": 0.657,
    "args": {
      "External id": 2478245,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943442.217, "dur": 2.006,
    "args": {
      "External id": 2478246,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943447.176, "dur": 0.805,
    "args": {
      "External id": 2478247,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 5950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943451.001, "dur": 1.077,
    "args": {
      "External id": 2478248,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 5951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943453.608, "dur": 0.663,
    "args": {
      "External id": 2478249,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 5952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943455.587, "dur": 1.700,
    "args": {
      "External id": 2478250,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 5953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943459.285, "dur": 0.452,
    "args": {
      "External id": 2478251,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 5954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943462.634, "dur": 0.720,
    "args": {
      "External id": 2478252,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943464.770, "dur": 0.645,
    "args": {
      "External id": 2478253,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943466.780, "dur": 1.485,
    "args": {
      "External id": 2478254,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943478.901, "dur": 0.706,
    "args": {
      "External id": 2478255,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943482.440, "dur": 0.738,
    "args": {
      "External id": 2478256,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 5959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943484.579, "dur": 0.535,
    "args": {
      "External id": 2478257,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 5960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943486.363, "dur": 1.668,
    "args": {
      "External id": 2478258,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 5961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943489.496, "dur": 0.738,
    "args": {
      "External id": 2478259,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 5962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943493.124, "dur": 0.710,
    "args": {
      "External id": 2478260,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 5963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943495.077, "dur": 0.720,
    "args": {
      "External id": 2478261,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943497.007, "dur": 1.574,
    "args": {
      "External id": 2478262,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943501.582, "dur": 0.615,
    "args": {
      "External id": 2478263,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943504.740, "dur": 0.649,
    "args": {
      "External id": 2478264,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 5967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943506.729, "dur": 0.410,
    "args": {
      "External id": 2478265,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 5968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943508.447, "dur": 1.692,
    "args": {
      "External id": 2478266,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 5969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943511.966, "dur": 0.713,
    "args": {
      "External id": 2478267,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 5970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943515.504, "dur": 0.862,
    "args": {
      "External id": 2478268,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 5971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943517.794, "dur": 0.822,
    "args": {
      "External id": 2478269,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 5972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943519.754, "dur": 1.754,
    "args": {
      "External id": 2478270,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[65536000], []], "Ev Idx": 5973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943563.030, "dur": 8474.540,
    "args": {
      "External id": 2478271,"Record function id": 0, "Concrete Inputs": ["", "", "1", ""], "Input type": ["c10::BFloat16", "", "Scalar", "TensorList"], "Input Strides": [[42076416, 1], [], [], []], "Input Dims": [[8, 42076416], [], [], []], "Ev Idx": 5974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943586.276, "dur": 8441.899,
    "args": {
      "External id": 2478272,"Record function id": 0, "Concrete Inputs": ["", "", "1", ""], "Input type": ["c10::BFloat16", "", "Scalar", "TensorList"], "Input Strides": [[42076416, 1], [], [], []], "Input Dims": [[8, 42076416], [], [], []], "Ev Idx": 5975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943610.768, "dur": 5.195,
    "args": {
      "External id": 2478273,"Record function id": 0, "Concrete Inputs": ["[4290]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943622.364, "dur": 8340.600,
    "args": {
      "External id": 2478274,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[4290], [], [], [], [], [], [], []], "Ev Idx": 5977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943625.215, "dur": 8337.219,
    "args": {
      "External id": 2478275,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[4290], [], [], [], [], [], []], "Ev Idx": 5978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943631.586, "dur": 5.657,
    "args": {
      "External id": 2478276,"Record function id": 0, "Concrete Inputs": ["[4290]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514655943638.989, "dur": 8320.232,
    "args": {
      "External id": 2478277,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[4290], [4290], []], "Ev Idx": 5980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952328.593, "dur": 32.635,
    "args": {
      "External id": 2478278,"Record function id": 0, "Ev Idx": 5981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 0/0", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952362.595, "dur": 213.714,
    "args": {
      "External id": 2478279,"Record function id": 0, "Ev Idx": 5982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952400.579, "dur": 165.510,
    "args": {
      "External id": 2478280,"Sequence number": 24740420, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "long int"], "Input Strides": [[2048, 1], [4096, 1]], "Input Dims": [[32000, 2048], [16, 4096]], "Ev Idx": 5983
    }
  },
  {
    "ph": "s", "id": 228, "pid": 1336759, "tid": 1336759, "ts": 1514655952400.579,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_embedding_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952466.305, "dur": 48.871,
    "args": {
      "External id": 2478281,"kernel_hash": "czer4rmela5ly34fgupsreshd4qo5eamimqpvrklpagr2lqevpwb", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/ze/czer4rmela5ly34fgupsreshd4qo5eamimqpvrklpagr2lqevpwb.py", "kernel_backend": "triton", "Input type": ["long int", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[4096, 1], [2048, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096], [32000, 2048], [16, 4096, 2048], []], "Ev Idx": 5984
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952638.522, "dur": 55.423,
    "args": {
      "External id": 2478282,"Record function id": 0, "Ev Idx": 5985
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.0)", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952704.404, "dur": 7995.720,
    "args": {
      "External id": 2478283,"Record function id": 0, "Ev Idx": 5986
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.0)", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952712.686, "dur": 804.683,
    "args": {
      "External id": 2478284,"Record function id": 0, "Ev Idx": 5987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952782.467, "dur": 10.489,
    "args": {
      "External id": 2478285,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 5988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952805.971, "dur": 36.758,
    "args": {
      "External id": 2478286,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 5989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952814.292, "dur": 2.318,
    "args": {
      "External id": 2478287,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952818.203, "dur": 0.490,
    "args": {
      "External id": 2478288,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952819.688, "dur": 1.866,
    "args": {
      "External id": 2478289,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952824.098, "dur": 0.529,
    "args": {
      "External id": 2478290,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952825.558, "dur": 0.455,
    "args": {
      "External id": 2478291,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952829.871, "dur": 0.179,
    "args": {
      "External id": 2478292,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952831.060, "dur": 0.419,
    "args": {
      "External id": 2478293,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952832.083, "dur": 1.220,
    "args": {
      "External id": 2478294,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952836.625, "dur": 0.476,
    "args": {
      "External id": 2478295,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 5998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952852.128, "dur": 37.332,
    "args": {
      "External id": 2478296,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 5999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952920.899, "dur": 182.196,
    "args": {
      "External id": 2478297,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 6000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952931.053, "dur": 3.520,
    "args": {
      "External id": 2478298,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952939.288, "dur": 11.259,
    "args": {
      "External id": 2478299,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 6002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952943.139, "dur": 6.987,
    "args": {
      "External id": 2478300,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 6003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952948.361, "dur": 0.618,
    "args": {
      "External id": 2478301,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 6004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952956.837, "dur": 93.242,
    "args": {
      "External id": 2478302,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 6005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952958.585, "dur": 0.464,
    "args": {
      "External id": 2478303,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952962.278, "dur": 0.555,
    "args": {
      "External id": 2478304,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952965.404, "dur": 0.299,
    "args": {
      "External id": 2478305,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952968.211, "dur": 0.179,
    "args": {
      "External id": 2478306,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952969.313, "dur": 2.818,
    "args": {
      "External id": 2478307,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952972.964, "dur": 0.192,
    "args": {
      "External id": 2478308,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952975.665, "dur": 0.347,
    "args": {
      "External id": 2478309,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952978.484, "dur": 0.193,
    "args": {
      "External id": 2478310,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655952979.553, "dur": 0.400,
    "args": {
      "External id": 2478311,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514655953066.020, "dur": 28.046,
    "args": {
      "External id": 2478312,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 6015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1336759,
    "ts": 1514655953156.076, "dur": 278.515,
    "args": {
      "External id": 2478313,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 6016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514655953183.233, "dur": 247.104,
    "args": {
      "External id": 2478314,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 6017, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1336759,
    "ts": 1514655953193.235, "dur": 231.967,
    "args": {
      "External id": 2478315,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 6018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514655953453.258, "dur": 1.910,
    "args": {
      "External id": 2478316,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 6019, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.0)", "pid": 1336759, "tid": 1336759,
    "ts": 1514655953555.960, "dur": 6921.690,
    "args": {
      "External id": 2478317,"Record function id": 0, "Ev Idx": 6020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655953646.465, "dur": 6.359,
    "args": {
      "External id": 2478318,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 6021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655953656.180, "dur": 1.087,
    "args": {
      "External id": 2478319,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 6022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655953658.793, "dur": 0.987,
    "args": {
      "External id": 2478320,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655953662.640, "dur": 1.512,
    "args": {
      "External id": 2478321,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655953665.355, "dur": 0.935,
    "args": {
      "External id": 2478322,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655953667.513, "dur": 0.777,
    "args": {
      "External id": 2478323,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655953669.609, "dur": 0.636,
    "args": {
      "External id": 2478324,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 6027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655953674.925, "dur": 1.620,
    "args": {
      "External id": 2478325,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655953677.757, "dur": 1.014,
    "args": {
      "External id": 2478326,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655953679.844, "dur": 0.691,
    "args": {
      "External id": 2478327,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514655953696.898, "dur": 6745.328,
    "args": {
      "External id": 2478328,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 6031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514655953712.455, "dur": 6723.912,
    "args": {
      "External id": 2478329,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 6032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514655953731.927, "dur": 13.253,
    "args": {
      "External id": 2478330,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514655953750.401, "dur": 6655.352,
    "args": {
      "External id": 2478331,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 6034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514655953753.095, "dur": 6652.138,
    "args": {
      "External id": 2478332,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 6035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655953758.626, "dur": 6.433,
    "args": {
      "External id": 2478333,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514655953766.739, "dur": 6635.578,
    "args": {
      "External id": 2478334,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 6037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514655960646.111, "dur": 29.884,
    "args": {
      "External id": 2478335,"Sequence number": 24740421, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 6038
    }
  },
  {
    "ph": "s", "id": 227, "pid": 1336759, "tid": 1336759, "ts": 1514655960646.111,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514655960662.253, "dur": 9.333,
    "args": {
      "External id": 2478336,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 6039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655960667.219, "dur": 4.015,
    "args": {
      "External id": 2478337,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 6040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1336759, "tid": 1336759,
    "ts": 1514655960738.106, "dur": 84.001,
    "args": {
      "External id": 2478338,"Record function id": 0, "Ev Idx": 6041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336759, "tid": 1336759,
    "ts": 1514655960823.768, "dur": 1079.883,
    "args": {
      "External id": 2478339,"Record function id": 0, "Ev Idx": 6042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514655960861.541, "dur": 1029.306,
    "args": {
      "External id": 2478340,"Sequence number": 24740422, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [32, 1], [32, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], "Input Dims": [[16, 4096, 2048], [2048], [2048, 2048], [2048, 2048], [2048, 2048], [8192, 32], [8192, 32], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], "Ev Idx": 6043
    }
  },
  {
    "ph": "s", "id": 226, "pid": 1336759, "tid": 1336759, "ts": 1514655960861.541,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514655960929.374, "dur": 43.978,
    "args": {
      "External id": 2478341,"kernel_hash": "cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514655961036.302, "dur": 96.129,
    "args": {
      "External id": 2478342,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514655961146.153, "dur": 37.881,
    "args": {
      "External id": 2478343,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514655961191.868, "dur": 30.399,
    "args": {
      "External id": 2478344,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514655961249.049, "dur": 25.700,
    "args": {
      "External id": 2478345,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514655961292.817, "dur": 15.701,
    "args": {
      "External id": 2478346,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514655961326.412, "dur": 123.170,
    "args": {
      "External id": 2478347,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 6050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514655961376.511, "dur": 9.769,
    "args": {
      "External id": 2478348,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 6051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655961380.745, "dur": 4.823,
    "args": {
      "External id": 2478349,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514655961388.638, "dur": 3.738,
    "args": {
      "External id": 2478350,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514655961395.713, "dur": 1.228,
    "args": {
      "External id": 2478351,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514655961399.224, "dur": 3.900,
    "args": {
      "External id": 2478352,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514655961459.259, "dur": 42.822,
    "args": {
      "External id": 2478353,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 1336759, "tid": 1336759,
    "ts": 1514655961559.136, "dur": 30.426,
    "args": {
      "External id": 2478354,"kernel_hash": "c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/3l/c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514655961599.147, "dur": 44.242,
    "args": {
      "External id": 2478355,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 6058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514655961653.851, "dur": 34.459,
    "args": {
      "External id": 2478356,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 6059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514655961710.514, "dur": 23.501,
    "args": {
      "External id": 2478357,"kernel_hash": "ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/kq/ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [16, 4096, 5632], []], "Ev Idx": 6060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514655961739.613, "dur": 34.987,
    "args": {
      "External id": 2478358,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 6061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514655961800.815, "dur": 18.042,
    "args": {
      "External id": 2478359,"kernel_hash": "cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u6/cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [65536, 2048], []], "Ev Idx": 6062
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.0)", "pid": 1336759, "tid": 1336759,
    "ts": 1514655961965.194, "dur": 136.723,
    "args": {
      "External id": 2478360,"Record function id": 0, "Ev Idx": 6063
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962177.043, "dur": 48.712,
    "args": {
      "External id": 2478361,"Record function id": 0, "Ev Idx": 6064
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.1)", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962233.818, "dur": 19068.272,
    "args": {
      "External id": 2478362,"Record function id": 0, "Ev Idx": 6065
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.1)", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962242.723, "dur": 933.223,
    "args": {
      "External id": 2478363,"Record function id": 0, "Ev Idx": 6066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962316.524, "dur": 9.476,
    "args": {
      "External id": 2478364,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962338.305, "dur": 42.075,
    "args": {
      "External id": 2478365,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 6068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962345.880, "dur": 4.545,
    "args": {
      "External id": 2478366,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962352.308, "dur": 0.549,
    "args": {
      "External id": 2478367,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962355.189, "dur": 0.567,
    "args": {
      "External id": 2478368,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962359.098, "dur": 0.400,
    "args": {
      "External id": 2478369,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962360.331, "dur": 1.566,
    "args": {
      "External id": 2478370,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962364.861, "dur": 0.303,
    "args": {
      "External id": 2478371,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962367.610, "dur": 0.521,
    "args": {
      "External id": 2478372,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962369.129, "dur": 0.268,
    "args": {
      "External id": 2478373,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962372.396, "dur": 2.337,
    "args": {
      "External id": 2478374,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962390.784, "dur": 43.224,
    "args": {
      "External id": 2478375,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 6078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962465.252, "dur": 150.426,
    "args": {
      "External id": 2478376,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 6079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962477.746, "dur": 3.581,
    "args": {
      "External id": 2478377,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962486.120, "dur": 9.804,
    "args": {
      "External id": 2478378,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 6081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962490.598, "dur": 4.888,
    "args": {
      "External id": 2478379,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 6082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962493.616, "dur": 0.596,
    "args": {
      "External id": 2478380,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 6083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962502.378, "dur": 62.162,
    "args": {
      "External id": 2478381,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 6084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962505.652, "dur": 0.446,
    "args": {
      "External id": 2478382,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962507.488, "dur": 1.633,
    "args": {
      "External id": 2478383,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6086
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962509.910, "dur": 2.474,
    "args": {
      "External id": 2478384,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962514.926, "dur": 0.193,
    "args": {
      "External id": 2478385,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962515.812, "dur": 0.266,
    "args": {
      "External id": 2478386,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962519.153, "dur": 0.202,
    "args": {
      "External id": 2478387,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962522.121, "dur": 0.207,
    "args": {
      "External id": 2478388,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962523.371, "dur": 0.376,
    "args": {
      "External id": 2478389,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962558.586, "dur": 0.760,
    "args": {
      "External id": 2478390,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962577.617, "dur": 28.612,
    "args": {
      "External id": 2478391,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 6094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962667.589, "dur": 413.040,
    "args": {
      "External id": 2478392,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 6095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962700.773, "dur": 374.370,
    "args": {
      "External id": 2478393,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 6096, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1336759,
    "ts": 1514655962711.219, "dur": 357.729,
    "args": {
      "External id": 2478394,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 6097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514655963105.171, "dur": 2.260,
    "args": {
      "External id": 2478395,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 6098, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.1)", "pid": 1336759, "tid": 1336759,
    "ts": 1514655963195.972, "dur": 17929.628,
    "args": {
      "External id": 2478396,"Record function id": 0, "Ev Idx": 6099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655963285.890, "dur": 6.425,
    "args": {
      "External id": 2478397,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 6100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655963295.785, "dur": 1.160,
    "args": {
      "External id": 2478398,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 6101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655963298.678, "dur": 1.159,
    "args": {
      "External id": 2478399,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655963301.392, "dur": 0.978,
    "args": {
      "External id": 2478400,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655963303.638, "dur": 0.957,
    "args": {
      "External id": 2478401,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655963307.510, "dur": 0.948,
    "args": {
      "External id": 2478402,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655963311.705, "dur": 0.841,
    "args": {
      "External id": 2478403,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 6106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655963313.795, "dur": 3.461,
    "args": {
      "External id": 2478404,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655963318.695, "dur": 0.839,
    "args": {
      "External id": 2478405,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655963322.835, "dur": 0.895,
    "args": {
      "External id": 2478406,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514655963340.723, "dur": 17747.513,
    "args": {
      "External id": 2478407,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 6110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514655963357.140, "dur": 17724.785,
    "args": {
      "External id": 2478408,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 6111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514655963378.545, "dur": 11.675,
    "args": {
      "External id": 2478409,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514655963393.323, "dur": 17658.009,
    "args": {
      "External id": 2478410,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 6113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514655963396.035, "dur": 17654.742,
    "args": {
      "External id": 2478411,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 6114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655963401.509, "dur": 4.486,
    "args": {
      "External id": 2478412,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514655963407.585, "dur": 17640.177,
    "args": {
      "External id": 2478413,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 6116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514655981250.137, "dur": 28.485,
    "args": {
      "External id": 2478414,"Sequence number": 24740423, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 6117
    }
  },
  {
    "ph": "s", "id": 225, "pid": 1336759, "tid": 1336759, "ts": 1514655981250.137,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514655981265.073, "dur": 9.204,
    "args": {
      "External id": 2478415,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 6118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655981269.907, "dur": 4.162,
    "args": {
      "External id": 2478416,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 6119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1336759, "tid": 1336759,
    "ts": 1514655981339.953, "dur": 75.885,
    "args": {
      "External id": 2478417,"Record function id": 0, "Ev Idx": 6120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336759, "tid": 1336759,
    "ts": 1514655981417.352, "dur": 1026.400,
    "args": {
      "External id": 2478418,"Record function id": 0, "Ev Idx": 6121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514655981452.873, "dur": 978.844,
    "args": {
      "External id": 2478419,"Sequence number": 24740424, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [32, 1], [32, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], "Input Dims": [[16, 4096, 2048], [2048], [2048, 2048], [2048, 2048], [2048, 2048], [8192, 32], [8192, 32], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], "Ev Idx": 6122
    }
  },
  {
    "ph": "s", "id": 224, "pid": 1336759, "tid": 1336759, "ts": 1514655981452.873,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514655981541.404, "dur": 42.991,
    "args": {
      "External id": 2478420,"kernel_hash": "cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514655981600.777, "dur": 100.672,
    "args": {
      "External id": 2478421,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514655981709.547, "dur": 38.049,
    "args": {
      "External id": 2478422,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514655981753.709, "dur": 30.228,
    "args": {
      "External id": 2478423,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514655981811.457, "dur": 22.923,
    "args": {
      "External id": 2478424,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514655981849.888, "dur": 16.513,
    "args": {
      "External id": 2478425,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514655981880.868, "dur": 162.686,
    "args": {
      "External id": 2478426,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 6129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514655981928.678, "dur": 11.408,
    "args": {
      "External id": 2478427,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 6130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655981933.191, "dur": 6.221,
    "args": {
      "External id": 2478428,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514655981943.937, "dur": 4.724,
    "args": {
      "External id": 2478429,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514655981949.808, "dur": 2.907,
    "args": {
      "External id": 2478430,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514655981954.994, "dur": 2.856,
    "args": {
      "External id": 2478431,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514655982055.836, "dur": 49.400,
    "args": {
      "External id": 2478432,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 1336759, "tid": 1336759,
    "ts": 1514655982134.453, "dur": 25.387,
    "args": {
      "External id": 2478433,"kernel_hash": "c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/3l/c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514655982168.653, "dur": 40.133,
    "args": {
      "External id": 2478434,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 6137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514655982214.685, "dur": 33.859,
    "args": {
      "External id": 2478435,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 6138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514655982272.478, "dur": 23.332,
    "args": {
      "External id": 2478436,"kernel_hash": "ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/kq/ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [16, 4096, 5632], []], "Ev Idx": 6139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514655982301.760, "dur": 33.715,
    "args": {
      "External id": 2478437,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 6140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514655982353.596, "dur": 18.447,
    "args": {
      "External id": 2478438,"kernel_hash": "cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u6/cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [65536, 2048], []], "Ev Idx": 6141
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.1)", "pid": 1336759, "tid": 1336759,
    "ts": 1514655982505.624, "dur": 93.284,
    "args": {
      "External id": 2478439,"Record function id": 0, "Ev Idx": 6142
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336759, "tid": 1336759,
    "ts": 1514655982675.276, "dur": 46.324,
    "args": {
      "External id": 2478440,"Record function id": 0, "Ev Idx": 6143
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.2)", "pid": 1336759, "tid": 1336759,
    "ts": 1514655982730.055, "dur": 18783.555,
    "args": {
      "External id": 2478441,"Record function id": 0, "Ev Idx": 6144
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.2)", "pid": 1336759, "tid": 1336759,
    "ts": 1514655982738.684, "dur": 901.217,
    "args": {
      "External id": 2478442,"Record function id": 0, "Ev Idx": 6145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514655982813.847, "dur": 8.492,
    "args": {
      "External id": 2478443,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514655982836.394, "dur": 37.016,
    "args": {
      "External id": 2478444,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 6147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655982841.156, "dur": 2.124,
    "args": {
      "External id": 2478445,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655982846.626, "dur": 0.628,
    "args": {
      "External id": 2478446,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655982849.675, "dur": 0.792,
    "args": {
      "External id": 2478447,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655982851.086, "dur": 0.535,
    "args": {
      "External id": 2478448,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655982854.239, "dur": 0.201,
    "args": {
      "External id": 2478449,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655982856.806, "dur": 0.595,
    "args": {
      "External id": 2478450,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655982858.092, "dur": 3.103,
    "args": {
      "External id": 2478451,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655982864.389, "dur": 0.456,
    "args": {
      "External id": 2478452,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655982865.931, "dur": 0.451,
    "args": {
      "External id": 2478453,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514655982883.175, "dur": 40.853,
    "args": {
      "External id": 2478454,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 6157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1336759,
    "ts": 1514655982956.942, "dur": 158.164,
    "args": {
      "External id": 2478455,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 6158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514655982968.726, "dur": 3.872,
    "args": {
      "External id": 2478456,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1336759,
    "ts": 1514655982977.609, "dur": 49.899,
    "args": {
      "External id": 2478457,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 6160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983020.479, "dur": 6.542,
    "args": {
      "External id": 2478458,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 6161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983024.278, "dur": 1.014,
    "args": {
      "External id": 2478459,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 6162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983035.325, "dur": 31.947,
    "args": {
      "External id": 2478460,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 6163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983038.933, "dur": 2.097,
    "args": {
      "External id": 2478461,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983041.890, "dur": 0.505,
    "args": {
      "External id": 2478462,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983043.421, "dur": 0.379,
    "args": {
      "External id": 2478463,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983048.043, "dur": 1.949,
    "args": {
      "External id": 2478464,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983050.852, "dur": 0.213,
    "args": {
      "External id": 2478465,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983053.284, "dur": 0.246,
    "args": {
      "External id": 2478466,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983056.195, "dur": 0.374,
    "args": {
      "External id": 2478467,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983057.309, "dur": 0.541,
    "args": {
      "External id": 2478468,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983060.876, "dur": 1.934,
    "args": {
      "External id": 2478469,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983077.878, "dur": 27.824,
    "args": {
      "External id": 2478470,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 6173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983170.527, "dur": 353.237,
    "args": {
      "External id": 2478471,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 6174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983202.166, "dur": 316.813,
    "args": {
      "External id": 2478472,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 6175, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983213.644, "dur": 300.262,
    "args": {
      "External id": 2478473,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 6176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983568.538, "dur": 3.219,
    "args": {
      "External id": 2478474,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 6177, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.2)", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983660.654, "dur": 17678.047,
    "args": {
      "External id": 2478475,"Record function id": 0, "Ev Idx": 6178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983748.962, "dur": 5.996,
    "args": {
      "External id": 2478476,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 6179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983758.428, "dur": 1.171,
    "args": {
      "External id": 2478477,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 6180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983761.221, "dur": 2.228,
    "args": {
      "External id": 2478478,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983765.108, "dur": 0.746,
    "args": {
      "External id": 2478479,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983767.148, "dur": 0.866,
    "args": {
      "External id": 2478480,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983769.426, "dur": 0.747,
    "args": {
      "External id": 2478481,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983773.197, "dur": 0.695,
    "args": {
      "External id": 2478482,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 6185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983775.359, "dur": 2.029,
    "args": {
      "External id": 2478483,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983778.784, "dur": 0.936,
    "args": {
      "External id": 2478484,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983781.194, "dur": 1.004,
    "args": {
      "External id": 2478485,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983801.625, "dur": 17500.032,
    "args": {
      "External id": 2478486,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 6189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983816.565, "dur": 17478.714,
    "args": {
      "External id": 2478487,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 6190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983838.712, "dur": 13.578,
    "args": {
      "External id": 2478488,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983855.046, "dur": 17408.931,
    "args": {
      "External id": 2478489,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 6192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983857.584, "dur": 17405.779,
    "args": {
      "External id": 2478490,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 6193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983862.753, "dur": 4.788,
    "args": {
      "External id": 2478491,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514655983869.151, "dur": 17391.329,
    "args": {
      "External id": 2478492,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 6195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656001461.332, "dur": 28.675,
    "args": {
      "External id": 2478493,"Sequence number": 24740425, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 6196
    }
  },
  {
    "ph": "s", "id": 223, "pid": 1336759, "tid": 1336759, "ts": 1514656001461.332,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514656001477.925, "dur": 7.651,
    "args": {
      "External id": 2478494,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 6197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656001481.152, "dur": 4.179,
    "args": {
      "External id": 2478495,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 6198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1336759, "tid": 1336759,
    "ts": 1514656001571.074, "dur": 74.070,
    "args": {
      "External id": 2478496,"Record function id": 0, "Ev Idx": 6199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656001647.446, "dur": 1071.785,
    "args": {
      "External id": 2478497,"Record function id": 0, "Ev Idx": 6200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656001686.015, "dur": 1020.887,
    "args": {
      "External id": 2478498,"Sequence number": 24740426, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [32, 1], [32, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], "Input Dims": [[16, 4096, 2048], [2048], [2048, 2048], [2048, 2048], [2048, 2048], [8192, 32], [8192, 32], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], "Ev Idx": 6201
    }
  },
  {
    "ph": "s", "id": 222, "pid": 1336759, "tid": 1336759, "ts": 1514656001686.015,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656001751.497, "dur": 43.935,
    "args": {
      "External id": 2478499,"kernel_hash": "cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656001807.275, "dur": 100.036,
    "args": {
      "External id": 2478500,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656001916.558, "dur": 37.305,
    "args": {
      "External id": 2478501,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656001963.306, "dur": 83.450,
    "args": {
      "External id": 2478502,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656002074.872, "dur": 24.484,
    "args": {
      "External id": 2478503,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656002115.043, "dur": 14.676,
    "args": {
      "External id": 2478504,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656002148.259, "dur": 121.851,
    "args": {
      "External id": 2478505,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 6208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656002195.403, "dur": 11.058,
    "args": {
      "External id": 2478506,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 6209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656002200.983, "dur": 4.654,
    "args": {
      "External id": 2478507,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656002208.848, "dur": 5.648,
    "args": {
      "External id": 2478508,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656002215.537, "dur": 1.159,
    "args": {
      "External id": 2478509,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656002219.108, "dur": 3.666,
    "args": {
      "External id": 2478510,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656002280.302, "dur": 47.829,
    "args": {
      "External id": 2478511,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 1336759, "tid": 1336759,
    "ts": 1514656002356.061, "dur": 25.628,
    "args": {
      "External id": 2478512,"kernel_hash": "c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/3l/c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656002389.941, "dur": 39.838,
    "args": {
      "External id": 2478513,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 6216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656002437.396, "dur": 34.265,
    "args": {
      "External id": 2478514,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 6217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656002491.442, "dur": 22.487,
    "args": {
      "External id": 2478515,"kernel_hash": "ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/kq/ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [16, 4096, 5632], []], "Ev Idx": 6218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656002519.664, "dur": 72.325,
    "args": {
      "External id": 2478516,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 6219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656002618.395, "dur": 18.993,
    "args": {
      "External id": 2478517,"kernel_hash": "cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u6/cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [65536, 2048], []], "Ev Idx": 6220
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.2)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656002780.502, "dur": 69.197,
    "args": {
      "External id": 2478518,"Record function id": 0, "Ev Idx": 6221
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656002921.333, "dur": 45.949,
    "args": {
      "External id": 2478519,"Record function id": 0, "Ev Idx": 6222
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.3)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656002976.677, "dur": 19185.101,
    "args": {
      "External id": 2478520,"Record function id": 0, "Ev Idx": 6223
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.3)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003039.349, "dur": 937.642,
    "args": {
      "External id": 2478521,"Record function id": 0, "Ev Idx": 6224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003119.051, "dur": 9.245,
    "args": {
      "External id": 2478522,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003142.003, "dur": 38.515,
    "args": {
      "External id": 2478523,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 6226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003146.998, "dur": 2.246,
    "args": {
      "External id": 2478524,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003155.317, "dur": 0.477,
    "args": {
      "External id": 2478525,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003156.466, "dur": 0.684,
    "args": {
      "External id": 2478526,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003161.391, "dur": 0.450,
    "args": {
      "External id": 2478527,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003164.721, "dur": 0.560,
    "args": {
      "External id": 2478528,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003165.961, "dur": 0.174,
    "args": {
      "External id": 2478529,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003167.748, "dur": 3.059,
    "args": {
      "External id": 2478530,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003171.828, "dur": 0.382,
    "args": {
      "External id": 2478531,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003172.895, "dur": 0.370,
    "args": {
      "External id": 2478532,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003191.909, "dur": 41.687,
    "args": {
      "External id": 2478533,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 6236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003266.020, "dur": 121.834,
    "args": {
      "External id": 2478534,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 6237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003277.840, "dur": 3.937,
    "args": {
      "External id": 2478535,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003286.460, "dur": 9.358,
    "args": {
      "External id": 2478536,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 6239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003290.898, "dur": 4.502,
    "args": {
      "External id": 2478537,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 6240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003293.694, "dur": 0.563,
    "args": {
      "External id": 2478538,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 6241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003303.936, "dur": 28.335,
    "args": {
      "External id": 2478539,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 6242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003305.283, "dur": 2.019,
    "args": {
      "External id": 2478540,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003308.276, "dur": 0.600,
    "args": {
      "External id": 2478541,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003311.326, "dur": 0.643,
    "args": {
      "External id": 2478542,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003314.840, "dur": 1.418,
    "args": {
      "External id": 2478543,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003318.223, "dur": 0.381,
    "args": {
      "External id": 2478544,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003319.549, "dur": 0.145,
    "args": {
      "External id": 2478545,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003322.402, "dur": 0.254,
    "args": {
      "External id": 2478546,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003325.159, "dur": 0.270,
    "args": {
      "External id": 2478547,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003326.202, "dur": 1.844,
    "args": {
      "External id": 2478548,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003354.012, "dur": 25.754,
    "args": {
      "External id": 2478549,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 6252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003439.774, "dur": 435.654,
    "args": {
      "External id": 2478550,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 6253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003470.015, "dur": 400.067,
    "args": {
      "External id": 2478551,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 6254, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003484.884, "dur": 379.329,
    "args": {
      "External id": 2478552,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 6255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656003900.381, "dur": 2.522,
    "args": {
      "External id": 2478553,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 6256, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.3)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656004045.140, "dur": 17886.357,
    "args": {
      "External id": 2478554,"Record function id": 0, "Ev Idx": 6257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656004141.099, "dur": 6.633,
    "args": {
      "External id": 2478555,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 6258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656004151.418, "dur": 1.376,
    "args": {
      "External id": 2478556,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 6259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656004154.696, "dur": 2.054,
    "args": {
      "External id": 2478557,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656004158.546, "dur": 1.215,
    "args": {
      "External id": 2478558,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656004161.187, "dur": 1.096,
    "args": {
      "External id": 2478559,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656004163.521, "dur": 1.083,
    "args": {
      "External id": 2478560,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656004167.990, "dur": 0.776,
    "args": {
      "External id": 2478561,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 6264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656004170.205, "dur": 2.503,
    "args": {
      "External id": 2478562,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656004174.542, "dur": 0.782,
    "args": {
      "External id": 2478563,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656004176.648, "dur": 0.794,
    "args": {
      "External id": 2478564,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656004197.366, "dur": 17692.983,
    "args": {
      "External id": 2478565,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 6268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656004212.815, "dur": 17670.743,
    "args": {
      "External id": 2478566,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 6269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656004234.885, "dur": 13.090,
    "args": {
      "External id": 2478567,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656004251.057, "dur": 17600.647,
    "args": {
      "External id": 2478568,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 6271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656004254.085, "dur": 17596.881,
    "args": {
      "External id": 2478569,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 6272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656004260.178, "dur": 6.166,
    "args": {
      "External id": 2478570,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656004268.436, "dur": 17579.570,
    "args": {
      "External id": 2478571,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 6274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656022100.093, "dur": 37.005,
    "args": {
      "External id": 2478572,"Sequence number": 24740427, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 6275
    }
  },
  {
    "ph": "s", "id": 221, "pid": 1336759, "tid": 1336759, "ts": 1514656022100.093,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514656022124.369, "dur": 8.124,
    "args": {
      "External id": 2478573,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 6276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656022127.519, "dur": 4.676,
    "args": {
      "External id": 2478574,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 6277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1336759, "tid": 1336759,
    "ts": 1514656022198.658, "dur": 71.519,
    "args": {
      "External id": 2478575,"Record function id": 0, "Ev Idx": 6278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656022273.337, "dur": 1074.906,
    "args": {
      "External id": 2478576,"Record function id": 0, "Ev Idx": 6279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656022310.721, "dur": 1024.667,
    "args": {
      "External id": 2478577,"Sequence number": 24740428, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [32, 1], [32, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], "Input Dims": [[16, 4096, 2048], [2048], [2048, 2048], [2048, 2048], [2048, 2048], [8192, 32], [8192, 32], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], "Ev Idx": 6280
    }
  },
  {
    "ph": "s", "id": 220, "pid": 1336759, "tid": 1336759, "ts": 1514656022310.721,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656022372.439, "dur": 41.013,
    "args": {
      "External id": 2478578,"kernel_hash": "cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656022425.029, "dur": 123.452,
    "args": {
      "External id": 2478579,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656022565.204, "dur": 41.596,
    "args": {
      "External id": 2478580,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656022613.174, "dur": 30.671,
    "args": {
      "External id": 2478581,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656022672.913, "dur": 24.634,
    "args": {
      "External id": 2478582,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656022714.234, "dur": 15.010,
    "args": {
      "External id": 2478583,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656022745.796, "dur": 123.548,
    "args": {
      "External id": 2478584,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 6287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656022794.451, "dur": 10.394,
    "args": {
      "External id": 2478585,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 6288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656022798.615, "dur": 5.188,
    "args": {
      "External id": 2478586,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656022807.177, "dur": 4.498,
    "args": {
      "External id": 2478587,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656022813.140, "dur": 0.991,
    "args": {
      "External id": 2478588,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656022818.299, "dur": 4.119,
    "args": {
      "External id": 2478589,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656022879.463, "dur": 42.879,
    "args": {
      "External id": 2478590,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 1336759, "tid": 1336759,
    "ts": 1514656022951.717, "dur": 26.585,
    "args": {
      "External id": 2478591,"kernel_hash": "c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/3l/c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023028.440, "dur": 45.640,
    "args": {
      "External id": 2478592,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 6295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023107.132, "dur": 35.792,
    "args": {
      "External id": 2478593,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 6296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023170.742, "dur": 25.334,
    "args": {
      "External id": 2478594,"kernel_hash": "ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/kq/ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [16, 4096, 5632], []], "Ev Idx": 6297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023202.144, "dur": 34.851,
    "args": {
      "External id": 2478595,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 6298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023254.623, "dur": 19.488,
    "args": {
      "External id": 2478596,"kernel_hash": "cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u6/cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [65536, 2048], []], "Ev Idx": 6299
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.3)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023410.058, "dur": 70.820,
    "args": {
      "External id": 2478597,"Record function id": 0, "Ev Idx": 6300
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023572.307, "dur": 47.781,
    "args": {
      "External id": 2478598,"Record function id": 0, "Ev Idx": 6301
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.4)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023629.013, "dur": 17666.831,
    "args": {
      "External id": 2478599,"Record function id": 0, "Ev Idx": 6302
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.4)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023638.619, "dur": 872.220,
    "args": {
      "External id": 2478600,"Record function id": 0, "Ev Idx": 6303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023713.641, "dur": 8.705,
    "args": {
      "External id": 2478601,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023734.924, "dur": 38.248,
    "args": {
      "External id": 2478602,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 6305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023741.719, "dur": 2.349,
    "args": {
      "External id": 2478603,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023747.748, "dur": 0.479,
    "args": {
      "External id": 2478604,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023749.747, "dur": 0.330,
    "args": {
      "External id": 2478605,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023752.399, "dur": 0.465,
    "args": {
      "External id": 2478606,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023755.674, "dur": 0.337,
    "args": {
      "External id": 2478607,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023758.118, "dur": 0.189,
    "args": {
      "External id": 2478608,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023759.324, "dur": 2.921,
    "args": {
      "External id": 2478609,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023762.942, "dur": 0.347,
    "args": {
      "External id": 2478610,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023765.456, "dur": 0.285,
    "args": {
      "External id": 2478611,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023782.560, "dur": 44.207,
    "args": {
      "External id": 2478612,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 6315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023857.197, "dur": 105.220,
    "args": {
      "External id": 2478613,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 6316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023868.748, "dur": 3.316,
    "args": {
      "External id": 2478614,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023876.912, "dur": 10.891,
    "args": {
      "External id": 2478615,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 6318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023881.279, "dur": 6.116,
    "args": {
      "External id": 2478616,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 6319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023885.828, "dur": 0.458,
    "args": {
      "External id": 2478617,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 6320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023893.387, "dur": 27.876,
    "args": {
      "External id": 2478618,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 6321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023894.880, "dur": 2.200,
    "args": {
      "External id": 2478619,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023899.929, "dur": 0.409,
    "args": {
      "External id": 2478620,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023901.013, "dur": 0.255,
    "args": {
      "External id": 2478621,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023905.368, "dur": 1.340,
    "args": {
      "External id": 2478622,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023907.444, "dur": 0.267,
    "args": {
      "External id": 2478623,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023908.481, "dur": 0.158,
    "args": {
      "External id": 2478624,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023912.138, "dur": 0.321,
    "args": {
      "External id": 2478625,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023912.914, "dur": 0.277,
    "args": {
      "External id": 2478626,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023914.080, "dur": 1.856,
    "args": {
      "External id": 2478627,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656023931.323, "dur": 23.376,
    "args": {
      "External id": 2478628,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 6331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656024054.135, "dur": 363.735,
    "args": {
      "External id": 2478629,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 6332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656024084.618, "dur": 328.277,
    "args": {
      "External id": 2478630,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 6333, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1336759,
    "ts": 1514656024096.084, "dur": 311.189,
    "args": {
      "External id": 2478631,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 6334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656024441.673, "dur": 2.431,
    "args": {
      "External id": 2478632,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 6335, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.4)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656024548.787, "dur": 16570.275,
    "args": {
      "External id": 2478633,"Record function id": 0, "Ev Idx": 6336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656024641.446, "dur": 6.426,
    "args": {
      "External id": 2478634,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 6337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656024651.381, "dur": 1.094,
    "args": {
      "External id": 2478635,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 6338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656024654.330, "dur": 2.187,
    "args": {
      "External id": 2478636,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656024658.481, "dur": 0.661,
    "args": {
      "External id": 2478637,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656024660.599, "dur": 0.729,
    "args": {
      "External id": 2478638,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656024662.451, "dur": 1.162,
    "args": {
      "External id": 2478639,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656024666.564, "dur": 0.926,
    "args": {
      "External id": 2478640,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 6343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656024670.298, "dur": 2.220,
    "args": {
      "External id": 2478641,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656024673.783, "dur": 0.723,
    "args": {
      "External id": 2478642,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656024675.899, "dur": 0.554,
    "args": {
      "External id": 2478643,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656024695.138, "dur": 16384.763,
    "args": {
      "External id": 2478644,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 6347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656024712.155, "dur": 16361.040,
    "args": {
      "External id": 2478645,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 6348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656024733.796, "dur": 12.776,
    "args": {
      "External id": 2478646,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656024749.850, "dur": 16291.590,
    "args": {
      "External id": 2478647,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 6350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656024752.249, "dur": 16288.554,
    "args": {
      "External id": 2478648,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 6351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656024757.911, "dur": 5.072,
    "args": {
      "External id": 2478649,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656024764.770, "dur": 16272.897,
    "args": {
      "External id": 2478650,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 6353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656041243.981, "dur": 27.952,
    "args": {
      "External id": 2478651,"Sequence number": 24740429, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 6354
    }
  },
  {
    "ph": "s", "id": 219, "pid": 1336759, "tid": 1336759, "ts": 1514656041243.981,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514656041259.590, "dur": 8.145,
    "args": {
      "External id": 2478652,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 6355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656041263.417, "dur": 4.087,
    "args": {
      "External id": 2478653,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 6356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1336759, "tid": 1336759,
    "ts": 1514656041331.401, "dur": 74.945,
    "args": {
      "External id": 2478654,"Record function id": 0, "Ev Idx": 6357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656041407.650, "dur": 1053.924,
    "args": {
      "External id": 2478655,"Record function id": 0, "Ev Idx": 6358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656041443.795, "dur": 1004.933,
    "args": {
      "External id": 2478656,"Sequence number": 24740430, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [32, 1], [32, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], "Input Dims": [[16, 4096, 2048], [2048], [2048, 2048], [2048, 2048], [2048, 2048], [8192, 32], [8192, 32], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], "Ev Idx": 6359
    }
  },
  {
    "ph": "s", "id": 218, "pid": 1336759, "tid": 1336759, "ts": 1514656041443.795,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656041506.154, "dur": 62.386,
    "args": {
      "External id": 2478657,"kernel_hash": "cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656041584.131, "dur": 101.380,
    "args": {
      "External id": 2478658,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656041695.580, "dur": 37.794,
    "args": {
      "External id": 2478659,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656041742.306, "dur": 30.548,
    "args": {
      "External id": 2478660,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656041796.968, "dur": 26.725,
    "args": {
      "External id": 2478661,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656041840.287, "dur": 13.814,
    "args": {
      "External id": 2478662,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656041871.101, "dur": 164.710,
    "args": {
      "External id": 2478663,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 6366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656041916.897, "dur": 9.916,
    "args": {
      "External id": 2478664,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 6367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656041920.972, "dur": 5.123,
    "args": {
      "External id": 2478665,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656041929.123, "dur": 6.184,
    "args": {
      "External id": 2478666,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656041938.051, "dur": 0.976,
    "args": {
      "External id": 2478667,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656041941.278, "dur": 3.133,
    "args": {
      "External id": 2478668,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656042048.151, "dur": 53.915,
    "args": {
      "External id": 2478669,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 1336759, "tid": 1336759,
    "ts": 1514656042134.764, "dur": 28.880,
    "args": {
      "External id": 2478670,"kernel_hash": "c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/3l/c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656042171.938, "dur": 40.949,
    "args": {
      "External id": 2478671,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 6374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656042220.709, "dur": 34.398,
    "args": {
      "External id": 2478672,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 6375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656042275.279, "dur": 26.422,
    "args": {
      "External id": 2478673,"kernel_hash": "ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/kq/ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [16, 4096, 5632], []], "Ev Idx": 6376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656042306.878, "dur": 33.662,
    "args": {
      "External id": 2478674,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 6377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656042361.328, "dur": 17.350,
    "args": {
      "External id": 2478675,"kernel_hash": "cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u6/cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [65536, 2048], []], "Ev Idx": 6378
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.4)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656042521.785, "dur": 95.429,
    "args": {
      "External id": 2478676,"Record function id": 0, "Ev Idx": 6379
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656042694.028, "dur": 46.080,
    "args": {
      "External id": 2478677,"Record function id": 0, "Ev Idx": 6380
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.5)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656042748.953, "dur": 18205.986,
    "args": {
      "External id": 2478678,"Record function id": 0, "Ev Idx": 6381
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.5)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656042758.618, "dur": 923.058,
    "args": {
      "External id": 2478679,"Record function id": 0, "Ev Idx": 6382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656042834.767, "dur": 8.410,
    "args": {
      "External id": 2478680,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656042855.935, "dur": 36.657,
    "args": {
      "External id": 2478681,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 6384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656042862.070, "dur": 2.176,
    "args": {
      "External id": 2478682,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656042868.172, "dur": 0.265,
    "args": {
      "External id": 2478683,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656042870.720, "dur": 0.463,
    "args": {
      "External id": 2478684,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656042872.116, "dur": 0.282,
    "args": {
      "External id": 2478685,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656042875.313, "dur": 0.340,
    "args": {
      "External id": 2478686,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656042877.623, "dur": 0.244,
    "args": {
      "External id": 2478687,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656042878.524, "dur": 3.214,
    "args": {
      "External id": 2478688,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656042882.737, "dur": 0.358,
    "args": {
      "External id": 2478689,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656042884.416, "dur": 0.882,
    "args": {
      "External id": 2478690,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656042902.425, "dur": 42.147,
    "args": {
      "External id": 2478691,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 6394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1336759,
    "ts": 1514656042975.518, "dur": 165.310,
    "args": {
      "External id": 2478692,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 6395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043031.381, "dur": 5.323,
    "args": {
      "External id": 2478693,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043042.348, "dur": 9.984,
    "args": {
      "External id": 2478694,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 6397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043046.662, "dur": 5.284,
    "args": {
      "External id": 2478695,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 6398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043049.567, "dur": 0.781,
    "args": {
      "External id": 2478696,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 6399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043060.210, "dur": 31.385,
    "args": {
      "External id": 2478697,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 6400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043062.734, "dur": 2.265,
    "args": {
      "External id": 2478698,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043066.668, "dur": 0.391,
    "args": {
      "External id": 2478699,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043068.353, "dur": 0.542,
    "args": {
      "External id": 2478700,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043071.655, "dur": 1.396,
    "args": {
      "External id": 2478701,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043074.789, "dur": 0.455,
    "args": {
      "External id": 2478702,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043077.061, "dur": 0.383,
    "args": {
      "External id": 2478703,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043080.186, "dur": 0.160,
    "args": {
      "External id": 2478704,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043082.018, "dur": 0.312,
    "args": {
      "External id": 2478705,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043083.656, "dur": 2.441,
    "args": {
      "External id": 2478706,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043104.617, "dur": 26.810,
    "args": {
      "External id": 2478707,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 6410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043196.312, "dur": 391.723,
    "args": {
      "External id": 2478708,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 6411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043227.352, "dur": 355.049,
    "args": {
      "External id": 2478709,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 6412, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043238.676, "dur": 337.923,
    "args": {
      "External id": 2478710,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 6413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043610.853, "dur": 2.592,
    "args": {
      "External id": 2478711,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 6414, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.5)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043701.422, "dur": 17069.250,
    "args": {
      "External id": 2478712,"Record function id": 0, "Ev Idx": 6415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043790.458, "dur": 6.249,
    "args": {
      "External id": 2478713,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 6416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043799.931, "dur": 1.048,
    "args": {
      "External id": 2478714,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 6417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043803.330, "dur": 2.059,
    "args": {
      "External id": 2478715,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043807.389, "dur": 0.973,
    "args": {
      "External id": 2478716,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043809.878, "dur": 0.993,
    "args": {
      "External id": 2478717,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043812.383, "dur": 0.742,
    "args": {
      "External id": 2478718,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043816.161, "dur": 1.104,
    "args": {
      "External id": 2478719,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 6422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043818.763, "dur": 3.440,
    "args": {
      "External id": 2478720,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043823.756, "dur": 0.921,
    "args": {
      "External id": 2478721,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043826.587, "dur": 0.798,
    "args": {
      "External id": 2478722,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043846.015, "dur": 16880.912,
    "args": {
      "External id": 2478723,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 6426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043866.074, "dur": 16854.477,
    "args": {
      "External id": 2478724,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 6427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043886.787, "dur": 13.610,
    "args": {
      "External id": 2478725,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043903.421, "dur": 16786.732,
    "args": {
      "External id": 2478726,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 6429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043906.045, "dur": 16783.480,
    "args": {
      "External id": 2478727,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 6430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043911.720, "dur": 4.732,
    "args": {
      "External id": 2478728,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656043918.218, "dur": 16768.589,
    "args": {
      "External id": 2478729,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 6432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656060903.052, "dur": 28.025,
    "args": {
      "External id": 2478730,"Sequence number": 24740431, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 6433
    }
  },
  {
    "ph": "s", "id": 217, "pid": 1336759, "tid": 1336759, "ts": 1514656060903.052,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514656060919.118, "dur": 7.580,
    "args": {
      "External id": 2478731,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 6434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656060922.253, "dur": 4.265,
    "args": {
      "External id": 2478732,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 6435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1336759, "tid": 1336759,
    "ts": 1514656061023.637, "dur": 75.294,
    "args": {
      "External id": 2478733,"Record function id": 0, "Ev Idx": 6436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656061101.428, "dur": 1049.063,
    "args": {
      "External id": 2478734,"Record function id": 0, "Ev Idx": 6437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656061146.097, "dur": 991.245,
    "args": {
      "External id": 2478735,"Sequence number": 24740432, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [32, 1], [32, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], "Input Dims": [[16, 4096, 2048], [2048], [2048, 2048], [2048, 2048], [2048, 2048], [8192, 32], [8192, 32], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], "Ev Idx": 6438
    }
  },
  {
    "ph": "s", "id": 216, "pid": 1336759, "tid": 1336759, "ts": 1514656061146.097,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656061210.241, "dur": 42.989,
    "args": {
      "External id": 2478736,"kernel_hash": "cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656061265.268, "dur": 100.300,
    "args": {
      "External id": 2478737,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656061376.862, "dur": 37.592,
    "args": {
      "External id": 2478738,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656061422.535, "dur": 31.953,
    "args": {
      "External id": 2478739,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656061477.937, "dur": 24.658,
    "args": {
      "External id": 2478740,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656061519.986, "dur": 32.026,
    "args": {
      "External id": 2478741,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656061578.277, "dur": 126.187,
    "args": {
      "External id": 2478742,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 6445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656061627.471, "dur": 11.219,
    "args": {
      "External id": 2478743,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 6446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656061632.185, "dur": 5.696,
    "args": {
      "External id": 2478744,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656061641.599, "dur": 4.963,
    "args": {
      "External id": 2478745,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656061647.609, "dur": 1.309,
    "args": {
      "External id": 2478746,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656061651.287, "dur": 4.781,
    "args": {
      "External id": 2478747,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656061715.028, "dur": 47.470,
    "args": {
      "External id": 2478748,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 1336759, "tid": 1336759,
    "ts": 1514656061792.022, "dur": 27.088,
    "args": {
      "External id": 2478749,"kernel_hash": "c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/3l/c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656061827.160, "dur": 39.892,
    "args": {
      "External id": 2478750,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 6453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656061876.518, "dur": 34.600,
    "args": {
      "External id": 2478751,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 6454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656061931.222, "dur": 22.846,
    "args": {
      "External id": 2478752,"kernel_hash": "ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/kq/ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [16, 4096, 5632], []], "Ev Idx": 6455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656061959.173, "dur": 68.999,
    "args": {
      "External id": 2478753,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 6456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062054.518, "dur": 18.963,
    "args": {
      "External id": 2478754,"kernel_hash": "cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u6/cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [65536, 2048], []], "Ev Idx": 6457
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.5)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062211.539, "dur": 70.985,
    "args": {
      "External id": 2478755,"Record function id": 0, "Ev Idx": 6458
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062352.528, "dur": 44.224,
    "args": {
      "External id": 2478756,"Record function id": 0, "Ev Idx": 6459
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.6)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062405.396, "dur": 17898.331,
    "args": {
      "External id": 2478757,"Record function id": 0, "Ev Idx": 6460
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.6)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062413.276, "dur": 895.378,
    "args": {
      "External id": 2478758,"Record function id": 0, "Ev Idx": 6461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062485.593, "dur": 7.323,
    "args": {
      "External id": 2478759,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062505.735, "dur": 56.860,
    "args": {
      "External id": 2478760,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 6463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062510.955, "dur": 2.099,
    "args": {
      "External id": 2478761,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062517.493, "dur": 0.434,
    "args": {
      "External id": 2478762,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062519.741, "dur": 0.450,
    "args": {
      "External id": 2478763,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062521.684, "dur": 0.509,
    "args": {
      "External id": 2478764,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062541.971, "dur": 0.640,
    "args": {
      "External id": 2478765,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062545.372, "dur": 0.315,
    "args": {
      "External id": 2478766,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062547.611, "dur": 3.670,
    "args": {
      "External id": 2478767,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062552.851, "dur": 0.535,
    "args": {
      "External id": 2478768,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062554.577, "dur": 0.373,
    "args": {
      "External id": 2478769,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062573.109, "dur": 40.559,
    "args": {
      "External id": 2478770,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 6473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062647.637, "dur": 117.334,
    "args": {
      "External id": 2478771,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 6474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062660.129, "dur": 4.585,
    "args": {
      "External id": 2478772,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062669.872, "dur": 10.113,
    "args": {
      "External id": 2478773,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 6476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062673.910, "dur": 5.682,
    "args": {
      "External id": 2478774,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 6477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062677.803, "dur": 0.557,
    "args": {
      "External id": 2478775,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 6478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062687.302, "dur": 30.825,
    "args": {
      "External id": 2478776,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 6479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062689.817, "dur": 2.190,
    "args": {
      "External id": 2478777,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062693.708, "dur": 0.403,
    "args": {
      "External id": 2478778,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062695.517, "dur": 0.403,
    "args": {
      "External id": 2478779,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062699.128, "dur": 1.338,
    "args": {
      "External id": 2478780,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062701.804, "dur": 0.232,
    "args": {
      "External id": 2478781,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062703.749, "dur": 0.348,
    "args": {
      "External id": 2478782,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062706.978, "dur": 0.408,
    "args": {
      "External id": 2478783,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062708.755, "dur": 0.303,
    "args": {
      "External id": 2478784,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062710.554, "dur": 1.987,
    "args": {
      "External id": 2478785,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062727.932, "dur": 28.673,
    "args": {
      "External id": 2478786,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 6489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062815.633, "dur": 395.206,
    "args": {
      "External id": 2478787,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 6490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062847.246, "dur": 358.227,
    "args": {
      "External id": 2478788,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 6491, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1336759,
    "ts": 1514656062858.170, "dur": 339.783,
    "args": {
      "External id": 2478789,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 6492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656063234.904, "dur": 2.662,
    "args": {
      "External id": 2478790,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 6493, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.6)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656063328.900, "dur": 16792.113,
    "args": {
      "External id": 2478791,"Record function id": 0, "Ev Idx": 6494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656063421.767, "dur": 6.244,
    "args": {
      "External id": 2478792,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 6495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656063432.225, "dur": 1.060,
    "args": {
      "External id": 2478793,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 6496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656063435.308, "dur": 2.100,
    "args": {
      "External id": 2478794,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656063439.109, "dur": 0.978,
    "args": {
      "External id": 2478795,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656063441.568, "dur": 1.005,
    "args": {
      "External id": 2478796,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656063444.298, "dur": 0.771,
    "args": {
      "External id": 2478797,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656063447.896, "dur": 0.828,
    "args": {
      "External id": 2478798,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 6501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656063450.111, "dur": 2.511,
    "args": {
      "External id": 2478799,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656063453.999, "dur": 0.662,
    "args": {
      "External id": 2478800,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656063456.026, "dur": 0.737,
    "args": {
      "External id": 2478801,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656063476.155, "dur": 16599.439,
    "args": {
      "External id": 2478802,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 6505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656063491.633, "dur": 16577.298,
    "args": {
      "External id": 2478803,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 6506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656063514.099, "dur": 30.604,
    "args": {
      "External id": 2478804,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656063549.113, "dur": 16488.234,
    "args": {
      "External id": 2478805,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 6508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656063551.482, "dur": 16485.171,
    "args": {
      "External id": 2478806,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 6509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656063557.346, "dur": 5.309,
    "args": {
      "External id": 2478807,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656063564.415, "dur": 16469.394,
    "args": {
      "External id": 2478808,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 6511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656080253.126, "dur": 27.262,
    "args": {
      "External id": 2478809,"Sequence number": 24740433, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 6512
    }
  },
  {
    "ph": "s", "id": 215, "pid": 1336759, "tid": 1336759, "ts": 1514656080253.126,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514656080268.045, "dur": 7.795,
    "args": {
      "External id": 2478810,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 6513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656080271.408, "dur": 4.225,
    "args": {
      "External id": 2478811,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 6514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1336759, "tid": 1336759,
    "ts": 1514656080340.637, "dur": 71.108,
    "args": {
      "External id": 2478812,"Record function id": 0, "Ev Idx": 6515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656080413.550, "dur": 1036.375,
    "args": {
      "External id": 2478813,"Record function id": 0, "Ev Idx": 6516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656080451.379, "dur": 985.815,
    "args": {
      "External id": 2478814,"Sequence number": 24740434, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [32, 1], [32, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], "Input Dims": [[16, 4096, 2048], [2048], [2048, 2048], [2048, 2048], [2048, 2048], [8192, 32], [8192, 32], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], "Ev Idx": 6517
    }
  },
  {
    "ph": "s", "id": 214, "pid": 1336759, "tid": 1336759, "ts": 1514656080451.379,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656080512.938, "dur": 58.996,
    "args": {
      "External id": 2478815,"kernel_hash": "cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656080586.644, "dur": 101.454,
    "args": {
      "External id": 2478816,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656080697.668, "dur": 39.348,
    "args": {
      "External id": 2478817,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656080745.585, "dur": 31.999,
    "args": {
      "External id": 2478818,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656080801.297, "dur": 24.525,
    "args": {
      "External id": 2478819,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656080842.918, "dur": 13.288,
    "args": {
      "External id": 2478820,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656080873.206, "dur": 159.495,
    "args": {
      "External id": 2478821,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 6524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656080919.756, "dur": 10.956,
    "args": {
      "External id": 2478822,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 6525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656080924.359, "dur": 5.652,
    "args": {
      "External id": 2478823,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656080933.358, "dur": 4.974,
    "args": {
      "External id": 2478824,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656080940.102, "dur": 1.205,
    "args": {
      "External id": 2478825,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656080943.883, "dur": 3.771,
    "args": {
      "External id": 2478826,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656081045.404, "dur": 50.595,
    "args": {
      "External id": 2478827,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 1336759, "tid": 1336759,
    "ts": 1514656081126.634, "dur": 27.726,
    "args": {
      "External id": 2478828,"kernel_hash": "c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/3l/c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656081162.765, "dur": 40.657,
    "args": {
      "External id": 2478829,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 6532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656081211.566, "dur": 35.105,
    "args": {
      "External id": 2478830,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 6533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656081269.149, "dur": 24.496,
    "args": {
      "External id": 2478831,"kernel_hash": "ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/kq/ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [16, 4096, 5632], []], "Ev Idx": 6534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656081299.428, "dur": 35.035,
    "args": {
      "External id": 2478832,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 6535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656081353.616, "dur": 17.103,
    "args": {
      "External id": 2478833,"kernel_hash": "cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u6/cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [65536, 2048], []], "Ev Idx": 6536
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.6)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656081509.863, "dur": 87.637,
    "args": {
      "External id": 2478834,"Record function id": 0, "Ev Idx": 6537
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656081673.519, "dur": 45.386,
    "args": {
      "External id": 2478835,"Record function id": 0, "Ev Idx": 6538
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.7)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656081728.823, "dur": 18064.316,
    "args": {
      "External id": 2478836,"Record function id": 0, "Ev Idx": 6539
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.7)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656081737.166, "dur": 917.367,
    "args": {
      "External id": 2478837,"Record function id": 0, "Ev Idx": 6540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656081813.154, "dur": 8.259,
    "args": {
      "External id": 2478838,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656081834.804, "dur": 38.158,
    "args": {
      "External id": 2478839,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 6542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656081840.324, "dur": 2.373,
    "args": {
      "External id": 2478840,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656081846.676, "dur": 0.400,
    "args": {
      "External id": 2478841,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656081848.422, "dur": 0.323,
    "args": {
      "External id": 2478842,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656081851.120, "dur": 0.437,
    "args": {
      "External id": 2478843,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656081854.298, "dur": 0.583,
    "args": {
      "External id": 2478844,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656081857.030, "dur": 0.371,
    "args": {
      "External id": 2478845,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656081858.596, "dur": 3.141,
    "args": {
      "External id": 2478846,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656081863.757, "dur": 0.312,
    "args": {
      "External id": 2478847,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656081865.331, "dur": 0.167,
    "args": {
      "External id": 2478848,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656081882.811, "dur": 42.704,
    "args": {
      "External id": 2478849,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 6552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1336759,
    "ts": 1514656081957.257, "dur": 159.169,
    "args": {
      "External id": 2478850,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 6553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656081969.232, "dur": 3.487,
    "args": {
      "External id": 2478851,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1336759,
    "ts": 1514656081977.248, "dur": 50.165,
    "args": {
      "External id": 2478852,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 6555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082019.995, "dur": 6.965,
    "args": {
      "External id": 2478853,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 6556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082024.396, "dur": 0.865,
    "args": {
      "External id": 2478854,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 6557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082035.643, "dur": 31.020,
    "args": {
      "External id": 2478855,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 6558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082037.845, "dur": 2.828,
    "args": {
      "External id": 2478856,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082042.864, "dur": 0.508,
    "args": {
      "External id": 2478857,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082044.745, "dur": 0.537,
    "args": {
      "External id": 2478858,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082048.360, "dur": 1.437,
    "args": {
      "External id": 2478859,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082051.055, "dur": 0.502,
    "args": {
      "External id": 2478860,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082053.391, "dur": 0.157,
    "args": {
      "External id": 2478861,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082055.715, "dur": 0.247,
    "args": {
      "External id": 2478862,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082057.271, "dur": 0.437,
    "args": {
      "External id": 2478863,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082058.939, "dur": 2.384,
    "args": {
      "External id": 2478864,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082080.122, "dur": 28.336,
    "args": {
      "External id": 2478865,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 6568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082169.349, "dur": 387.981,
    "args": {
      "External id": 2478866,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 6569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082199.747, "dur": 352.107,
    "args": {
      "External id": 2478867,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 6570, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082212.085, "dur": 332.973,
    "args": {
      "External id": 2478868,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 6571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082582.129, "dur": 2.887,
    "args": {
      "External id": 2478869,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 6572, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.7)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082675.438, "dur": 16937.986,
    "args": {
      "External id": 2478870,"Record function id": 0, "Ev Idx": 6573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082773.018, "dur": 6.612,
    "args": {
      "External id": 2478871,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 6574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082784.457, "dur": 1.013,
    "args": {
      "External id": 2478872,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 6575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082787.305, "dur": 2.322,
    "args": {
      "External id": 2478873,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082792.813, "dur": 0.997,
    "args": {
      "External id": 2478874,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082795.287, "dur": 0.787,
    "args": {
      "External id": 2478875,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082797.533, "dur": 1.002,
    "args": {
      "External id": 2478876,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082800.350, "dur": 0.757,
    "args": {
      "External id": 2478877,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 6580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082804.276, "dur": 2.194,
    "args": {
      "External id": 2478878,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082808.405, "dur": 0.756,
    "args": {
      "External id": 2478879,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082810.785, "dur": 0.600,
    "args": {
      "External id": 2478880,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082829.368, "dur": 16741.841,
    "args": {
      "External id": 2478881,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 6584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082844.644, "dur": 16719.434,
    "args": {
      "External id": 2478882,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 6585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082865.268, "dur": 13.057,
    "args": {
      "External id": 2478883,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082882.671, "dur": 16639.552,
    "args": {
      "External id": 2478884,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 6587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082885.283, "dur": 16636.349,
    "args": {
      "External id": 2478885,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 6588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082890.607, "dur": 4.596,
    "args": {
      "External id": 2478886,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656082896.679, "dur": 16621.900,
    "args": {
      "External id": 2478887,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 6590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656099741.786, "dur": 28.677,
    "args": {
      "External id": 2478888,"Sequence number": 24740435, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 6591
    }
  },
  {
    "ph": "s", "id": 213, "pid": 1336759, "tid": 1336759, "ts": 1514656099741.786,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514656099758.241, "dur": 7.484,
    "args": {
      "External id": 2478889,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 6592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656099761.267, "dur": 4.266,
    "args": {
      "External id": 2478890,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 6593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1336759, "tid": 1336759,
    "ts": 1514656099831.321, "dur": 75.046,
    "args": {
      "External id": 2478891,"Record function id": 0, "Ev Idx": 6594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656099908.418, "dur": 1036.980,
    "args": {
      "External id": 2478892,"Record function id": 0, "Ev Idx": 6595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656099947.525, "dur": 985.471,
    "args": {
      "External id": 2478893,"Sequence number": 24740436, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [32, 1], [32, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], "Input Dims": [[16, 4096, 2048], [2048], [2048, 2048], [2048, 2048], [2048, 2048], [8192, 32], [8192, 32], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], "Ev Idx": 6596
    }
  },
  {
    "ph": "s", "id": 212, "pid": 1336759, "tid": 1336759, "ts": 1514656099947.525,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656100045.222, "dur": 41.929,
    "args": {
      "External id": 2478894,"kernel_hash": "cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656100100.040, "dur": 101.271,
    "args": {
      "External id": 2478895,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656100211.258, "dur": 37.890,
    "args": {
      "External id": 2478896,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656100258.315, "dur": 32.492,
    "args": {
      "External id": 2478897,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656100313.978, "dur": 22.999,
    "args": {
      "External id": 2478898,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656100353.953, "dur": 14.117,
    "args": {
      "External id": 2478899,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656100385.489, "dur": 128.820,
    "args": {
      "External id": 2478900,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 6603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656100433.912, "dur": 10.620,
    "args": {
      "External id": 2478901,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 6604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656100438.640, "dur": 5.239,
    "args": {
      "External id": 2478902,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656100446.986, "dur": 5.962,
    "args": {
      "External id": 2478903,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656100460.206, "dur": 1.024,
    "args": {
      "External id": 2478904,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656100463.710, "dur": 3.707,
    "args": {
      "External id": 2478905,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656100524.503, "dur": 65.442,
    "args": {
      "External id": 2478906,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 1336759, "tid": 1336759,
    "ts": 1514656100625.179, "dur": 26.681,
    "args": {
      "External id": 2478907,"kernel_hash": "c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/3l/c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656100659.928, "dur": 39.987,
    "args": {
      "External id": 2478908,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 6611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656100707.856, "dur": 34.668,
    "args": {
      "External id": 2478909,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 6612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656100764.298, "dur": 25.578,
    "args": {
      "External id": 2478910,"kernel_hash": "ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/kq/ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [16, 4096, 5632], []], "Ev Idx": 6613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656100795.536, "dur": 34.147,
    "args": {
      "External id": 2478911,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 6614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656100850.653, "dur": 16.794,
    "args": {
      "External id": 2478912,"kernel_hash": "cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u6/cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [65536, 2048], []], "Ev Idx": 6615
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.7)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101045.118, "dur": 71.892,
    "args": {
      "External id": 2478913,"Record function id": 0, "Ev Idx": 6616
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101191.991, "dur": 45.191,
    "args": {
      "External id": 2478914,"Record function id": 0, "Ev Idx": 6617
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.8)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101245.514, "dur": 18039.549,
    "args": {
      "External id": 2478915,"Record function id": 0, "Ev Idx": 6618
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.8)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101253.373, "dur": 897.155,
    "args": {
      "External id": 2478916,"Record function id": 0, "Ev Idx": 6619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101331.412, "dur": 7.844,
    "args": {
      "External id": 2478917,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101352.372, "dur": 35.462,
    "args": {
      "External id": 2478918,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 6621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101357.599, "dur": 2.116,
    "args": {
      "External id": 2478919,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101363.947, "dur": 0.479,
    "args": {
      "External id": 2478920,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101366.297, "dur": 0.503,
    "args": {
      "External id": 2478921,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101368.065, "dur": 0.545,
    "args": {
      "External id": 2478922,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101371.706, "dur": 0.163,
    "args": {
      "External id": 2478923,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101373.046, "dur": 0.559,
    "args": {
      "External id": 2478924,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101374.851, "dur": 3.076,
    "args": {
      "External id": 2478925,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101379.457, "dur": 0.209,
    "args": {
      "External id": 2478926,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101381.426, "dur": 0.348,
    "args": {
      "External id": 2478927,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101397.833, "dur": 41.485,
    "args": {
      "External id": 2478928,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 6631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101470.652, "dur": 131.111,
    "args": {
      "External id": 2478929,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 6632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101482.459, "dur": 3.878,
    "args": {
      "External id": 2478930,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101491.289, "dur": 10.012,
    "args": {
      "External id": 2478931,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 6634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101495.894, "dur": 5.009,
    "args": {
      "External id": 2478932,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 6635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101499.108, "dur": 0.531,
    "args": {
      "External id": 2478933,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 6636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101507.392, "dur": 47.595,
    "args": {
      "External id": 2478934,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 6637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101509.454, "dur": 2.114,
    "args": {
      "External id": 2478935,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101512.879, "dur": 0.325,
    "args": {
      "External id": 2478936,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101514.572, "dur": 0.542,
    "args": {
      "External id": 2478937,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101517.933, "dur": 1.381,
    "args": {
      "External id": 2478938,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101520.763, "dur": 0.348,
    "args": {
      "External id": 2478939,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101522.054, "dur": 0.538,
    "args": {
      "External id": 2478940,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101542.827, "dur": 0.747,
    "args": {
      "External id": 2478941,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101546.254, "dur": 0.225,
    "args": {
      "External id": 2478942,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101547.690, "dur": 2.064,
    "args": {
      "External id": 2478943,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101566.651, "dur": 26.678,
    "args": {
      "External id": 2478944,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 6647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101655.038, "dur": 398.966,
    "args": {
      "External id": 2478945,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 6648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101685.532, "dur": 362.650,
    "args": {
      "External id": 2478946,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 6649, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1336759,
    "ts": 1514656101696.905, "dur": 345.263,
    "args": {
      "External id": 2478947,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 6650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656102078.978, "dur": 2.569,
    "args": {
      "External id": 2478948,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 6651, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.8)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656102171.102, "dur": 16934.571,
    "args": {
      "External id": 2478949,"Record function id": 0, "Ev Idx": 6652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656102266.130, "dur": 6.240,
    "args": {
      "External id": 2478950,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 6653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656102276.323, "dur": 1.138,
    "args": {
      "External id": 2478951,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 6654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656102279.150, "dur": 2.204,
    "args": {
      "External id": 2478952,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656102283.592, "dur": 0.939,
    "args": {
      "External id": 2478953,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656102285.864, "dur": 1.075,
    "args": {
      "External id": 2478954,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656102288.391, "dur": 1.119,
    "args": {
      "External id": 2478955,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656102292.528, "dur": 1.177,
    "args": {
      "External id": 2478956,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 6659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656102295.150, "dur": 2.466,
    "args": {
      "External id": 2478957,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656102299.080, "dur": 0.811,
    "args": {
      "External id": 2478958,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656102301.521, "dur": 0.694,
    "args": {
      "External id": 2478959,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656102330.725, "dur": 16733.843,
    "args": {
      "External id": 2478960,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 6663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656102347.038, "dur": 16710.966,
    "args": {
      "External id": 2478961,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 6664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656102366.110, "dur": 13.912,
    "args": {
      "External id": 2478962,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656102383.362, "dur": 16644.875,
    "args": {
      "External id": 2478963,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 6666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656102385.971, "dur": 16641.636,
    "args": {
      "External id": 2478964,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 6667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656102391.644, "dur": 4.627,
    "args": {
      "External id": 2478965,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656102397.764, "dur": 16626.776,
    "args": {
      "External id": 2478966,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 6669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656119234.124, "dur": 28.021,
    "args": {
      "External id": 2478967,"Sequence number": 24740437, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 6670
    }
  },
  {
    "ph": "s", "id": 211, "pid": 1336759, "tid": 1336759, "ts": 1514656119234.124,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514656119249.632, "dur": 7.881,
    "args": {
      "External id": 2478968,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 6671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656119252.485, "dur": 4.819,
    "args": {
      "External id": 2478969,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 6672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1336759, "tid": 1336759,
    "ts": 1514656119321.234, "dur": 72.502,
    "args": {
      "External id": 2478970,"Record function id": 0, "Ev Idx": 6673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656119395.350, "dur": 1025.115,
    "args": {
      "External id": 2478971,"Record function id": 0, "Ev Idx": 6674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656119432.651, "dur": 975.654,
    "args": {
      "External id": 2478972,"Sequence number": 24740438, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [32, 1], [32, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], "Input Dims": [[16, 4096, 2048], [2048], [2048, 2048], [2048, 2048], [2048, 2048], [8192, 32], [8192, 32], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], "Ev Idx": 6675
    }
  },
  {
    "ph": "s", "id": 210, "pid": 1336759, "tid": 1336759, "ts": 1514656119432.651,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656119493.253, "dur": 55.191,
    "args": {
      "External id": 2478973,"kernel_hash": "cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656119563.163, "dur": 102.162,
    "args": {
      "External id": 2478974,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656119674.664, "dur": 37.640,
    "args": {
      "External id": 2478975,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656119721.156, "dur": 31.841,
    "args": {
      "External id": 2478976,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656119776.265, "dur": 24.181,
    "args": {
      "External id": 2478977,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656119817.234, "dur": 14.104,
    "args": {
      "External id": 2478978,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656119848.864, "dur": 119.607,
    "args": {
      "External id": 2478979,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 6682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656119895.185, "dur": 10.414,
    "args": {
      "External id": 2478980,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 6683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656119899.633, "dur": 5.185,
    "args": {
      "External id": 2478981,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656119908.374, "dur": 5.782,
    "args": {
      "External id": 2478982,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656119915.557, "dur": 1.101,
    "args": {
      "External id": 2478983,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656119918.732, "dur": 3.669,
    "args": {
      "External id": 2478984,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656119978.632, "dur": 86.320,
    "args": {
      "External id": 2478985,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120098.793, "dur": 26.473,
    "args": {
      "External id": 2478986,"kernel_hash": "c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/3l/c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120133.633, "dur": 41.353,
    "args": {
      "External id": 2478987,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 6690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120183.284, "dur": 34.464,
    "args": {
      "External id": 2478988,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 6691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120240.200, "dur": 24.403,
    "args": {
      "External id": 2478989,"kernel_hash": "ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/kq/ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [16, 4096, 5632], []], "Ev Idx": 6692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120269.824, "dur": 34.040,
    "args": {
      "External id": 2478990,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 6693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120323.231, "dur": 17.850,
    "args": {
      "External id": 2478991,"kernel_hash": "cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u6/cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [65536, 2048], []], "Ev Idx": 6694
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.8)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120482.120, "dur": 91.461,
    "args": {
      "External id": 2478992,"Record function id": 0, "Ev Idx": 6695
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120650.375, "dur": 46.706,
    "args": {
      "External id": 2478993,"Record function id": 0, "Ev Idx": 6696
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.9)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120705.748, "dur": 17978.250,
    "args": {
      "External id": 2478994,"Record function id": 0, "Ev Idx": 6697
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.9)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120715.092, "dur": 962.652,
    "args": {
      "External id": 2478995,"Record function id": 0, "Ev Idx": 6698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120792.622, "dur": 8.247,
    "args": {
      "External id": 2478996,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120813.419, "dur": 32.823,
    "args": {
      "External id": 2478997,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 6700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120818.242, "dur": 2.405,
    "args": {
      "External id": 2478998,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120824.726, "dur": 0.309,
    "args": {
      "External id": 2478999,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120825.912, "dur": 0.516,
    "args": {
      "External id": 2479000,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120827.381, "dur": 0.456,
    "args": {
      "External id": 2479001,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120830.375, "dur": 0.442,
    "args": {
      "External id": 2479002,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120831.916, "dur": 0.310,
    "args": {
      "External id": 2479003,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120832.945, "dur": 3.726,
    "args": {
      "External id": 2479004,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120838.067, "dur": 0.346,
    "args": {
      "External id": 2479005,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120839.333, "dur": 0.345,
    "args": {
      "External id": 2479006,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120856.652, "dur": 41.602,
    "args": {
      "External id": 2479007,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 6710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120930.295, "dur": 163.161,
    "args": {
      "External id": 2479008,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 6711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120942.735, "dur": 3.764,
    "args": {
      "External id": 2479009,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120951.226, "dur": 9.815,
    "args": {
      "External id": 2479010,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 6713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120955.510, "dur": 5.137,
    "args": {
      "External id": 2479011,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 6714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120958.837, "dur": 0.529,
    "args": {
      "External id": 2479012,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 6715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120967.983, "dur": 68.708,
    "args": {
      "External id": 2479013,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 6716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120969.679, "dur": 2.016,
    "args": {
      "External id": 2479014,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120975.454, "dur": 0.472,
    "args": {
      "External id": 2479015,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120977.453, "dur": 0.510,
    "args": {
      "External id": 2479016,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656120980.538, "dur": 38.828,
    "args": {
      "External id": 2479017,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656121022.199, "dur": 0.301,
    "args": {
      "External id": 2479018,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656121023.594, "dur": 2.239,
    "args": {
      "External id": 2479019,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656121027.063, "dur": 0.402,
    "args": {
      "External id": 2479020,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656121028.918, "dur": 0.313,
    "args": {
      "External id": 2479021,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656121031.790, "dur": 0.172,
    "args": {
      "External id": 2479022,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656121055.879, "dur": 29.057,
    "args": {
      "External id": 2479023,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 6726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656121150.125, "dur": 428.010,
    "args": {
      "External id": 2479024,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 6727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656121179.057, "dur": 393.007,
    "args": {
      "External id": 2479025,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 6728, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1336759,
    "ts": 1514656121191.059, "dur": 374.583,
    "args": {
      "External id": 2479026,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 6729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656121602.539, "dur": 2.809,
    "args": {
      "External id": 2479027,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 6730, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.9)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656121699.281, "dur": 16783.830,
    "args": {
      "External id": 2479028,"Record function id": 0, "Ev Idx": 6731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656121790.853, "dur": 6.393,
    "args": {
      "External id": 2479029,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 6732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656121800.824, "dur": 1.019,
    "args": {
      "External id": 2479030,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 6733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656121803.735, "dur": 2.280,
    "args": {
      "External id": 2479031,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656121807.855, "dur": 0.886,
    "args": {
      "External id": 2479032,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656121810.309, "dur": 0.840,
    "args": {
      "External id": 2479033,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656121812.453, "dur": 0.702,
    "args": {
      "External id": 2479034,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656121816.485, "dur": 0.875,
    "args": {
      "External id": 2479035,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 6738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656121818.678, "dur": 2.520,
    "args": {
      "External id": 2479036,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656121822.576, "dur": 0.693,
    "args": {
      "External id": 2479037,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656121824.550, "dur": 0.808,
    "args": {
      "External id": 2479038,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656121845.200, "dur": 16592.794,
    "args": {
      "External id": 2479039,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 6742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656121860.206, "dur": 16571.010,
    "args": {
      "External id": 2479040,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 6743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656121880.551, "dur": 13.844,
    "args": {
      "External id": 2479041,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656121897.375, "dur": 16503.665,
    "args": {
      "External id": 2479042,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 6745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656121899.940, "dur": 16500.326,
    "args": {
      "External id": 2479043,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 6746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656121905.271, "dur": 5.763,
    "args": {
      "External id": 2479044,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656121913.054, "dur": 16484.377,
    "args": {
      "External id": 2479045,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 6748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656138629.878, "dur": 28.869,
    "args": {
      "External id": 2479046,"Sequence number": 24740439, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 6749
    }
  },
  {
    "ph": "s", "id": 209, "pid": 1336759, "tid": 1336759, "ts": 1514656138629.878,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514656138645.772, "dur": 8.339,
    "args": {
      "External id": 2479047,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 6750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656138649.062, "dur": 4.696,
    "args": {
      "External id": 2479048,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 6751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1336759, "tid": 1336759,
    "ts": 1514656138721.529, "dur": 75.096,
    "args": {
      "External id": 2479049,"Record function id": 0, "Ev Idx": 6752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656138798.062, "dur": 1034.940,
    "args": {
      "External id": 2479050,"Record function id": 0, "Ev Idx": 6753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656138837.400, "dur": 982.625,
    "args": {
      "External id": 2479051,"Sequence number": 24740440, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [32, 1], [32, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], "Input Dims": [[16, 4096, 2048], [2048], [2048, 2048], [2048, 2048], [2048, 2048], [8192, 32], [8192, 32], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], "Ev Idx": 6754
    }
  },
  {
    "ph": "s", "id": 208, "pid": 1336759, "tid": 1336759, "ts": 1514656138837.400,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656138899.934, "dur": 39.225,
    "args": {
      "External id": 2479052,"kernel_hash": "cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656138951.379, "dur": 137.799,
    "args": {
      "External id": 2479053,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656139104.118, "dur": 41.738,
    "args": {
      "External id": 2479054,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656139153.520, "dur": 30.581,
    "args": {
      "External id": 2479055,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656139211.367, "dur": 25.541,
    "args": {
      "External id": 2479056,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656139253.510, "dur": 15.007,
    "args": {
      "External id": 2479057,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656139286.543, "dur": 122.367,
    "args": {
      "External id": 2479058,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 6761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656139333.639, "dur": 10.907,
    "args": {
      "External id": 2479059,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 6762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656139338.489, "dur": 5.309,
    "args": {
      "External id": 2479060,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656139347.499, "dur": 4.529,
    "args": {
      "External id": 2479061,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656139353.431, "dur": 0.798,
    "args": {
      "External id": 2479062,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656139357.117, "dur": 4.471,
    "args": {
      "External id": 2479063,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656139418.814, "dur": 42.363,
    "args": {
      "External id": 2479064,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 1336759, "tid": 1336759,
    "ts": 1514656139487.779, "dur": 24.980,
    "args": {
      "External id": 2479065,"kernel_hash": "c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/3l/c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656139525.260, "dur": 61.100,
    "args": {
      "External id": 2479066,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 6769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656139597.161, "dur": 35.696,
    "args": {
      "External id": 2479067,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 6770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656139656.766, "dur": 26.044,
    "args": {
      "External id": 2479068,"kernel_hash": "ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/kq/ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [16, 4096, 5632], []], "Ev Idx": 6771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656139688.928, "dur": 33.089,
    "args": {
      "External id": 2479069,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 6772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656139740.976, "dur": 16.802,
    "args": {
      "External id": 2479070,"kernel_hash": "cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u6/cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [65536, 2048], []], "Ev Idx": 6773
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.9)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656139895.239, "dur": 70.265,
    "args": {
      "External id": 2479071,"Record function id": 0, "Ev Idx": 6774
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140080.533, "dur": 48.244,
    "args": {
      "External id": 2479072,"Record function id": 0, "Ev Idx": 6775
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.10)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140138.881, "dur": 18076.515,
    "args": {
      "External id": 2479073,"Record function id": 0, "Ev Idx": 6776
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.10)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140147.354, "dur": 934.845,
    "args": {
      "External id": 2479074,"Record function id": 0, "Ev Idx": 6777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140225.809, "dur": 8.515,
    "args": {
      "External id": 2479075,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140247.114, "dur": 36.365,
    "args": {
      "External id": 2479076,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 6779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140252.805, "dur": 1.932,
    "args": {
      "External id": 2479077,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140258.991, "dur": 0.443,
    "args": {
      "External id": 2479078,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140260.855, "dur": 0.434,
    "args": {
      "External id": 2479079,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140262.411, "dur": 0.430,
    "args": {
      "External id": 2479080,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140265.624, "dur": 0.600,
    "args": {
      "External id": 2479081,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140267.472, "dur": 0.668,
    "args": {
      "External id": 2479082,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140269.593, "dur": 3.846,
    "args": {
      "External id": 2479083,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140274.713, "dur": 0.321,
    "args": {
      "External id": 2479084,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140276.723, "dur": 0.197,
    "args": {
      "External id": 2479085,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140293.601, "dur": 40.676,
    "args": {
      "External id": 2479086,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 6789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140365.425, "dur": 107.580,
    "args": {
      "External id": 2479087,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 6790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140377.506, "dur": 4.112,
    "args": {
      "External id": 2479088,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140386.410, "dur": 10.443,
    "args": {
      "External id": 2479089,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 6792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140390.837, "dur": 5.590,
    "args": {
      "External id": 2479090,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 6793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140394.544, "dur": 0.763,
    "args": {
      "External id": 2479091,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 6794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140403.506, "dur": 29.324,
    "args": {
      "External id": 2479092,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 6795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140405.720, "dur": 2.045,
    "args": {
      "External id": 2479093,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140409.518, "dur": 0.303,
    "args": {
      "External id": 2479094,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140411.337, "dur": 0.270,
    "args": {
      "External id": 2479095,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140414.770, "dur": 1.580,
    "args": {
      "External id": 2479096,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140417.457, "dur": 0.646,
    "args": {
      "External id": 2479097,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140419.315, "dur": 0.276,
    "args": {
      "External id": 2479098,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140422.534, "dur": 0.310,
    "args": {
      "External id": 2479099,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140424.213, "dur": 0.384,
    "args": {
      "External id": 2479100,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140425.946, "dur": 1.704,
    "args": {
      "External id": 2479101,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140442.555, "dur": 22.920,
    "args": {
      "External id": 2479102,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 6805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140522.347, "dur": 420.874,
    "args": {
      "External id": 2479103,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 6806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140568.726, "dur": 369.143,
    "args": {
      "External id": 2479104,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 6807, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140612.226, "dur": 319.669,
    "args": {
      "External id": 2479105,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 6808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656140968.662, "dur": 2.264,
    "args": {
      "External id": 2479106,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 6809, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.10)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656141104.566, "dur": 16921.556,
    "args": {
      "External id": 2479107,"Record function id": 0, "Ev Idx": 6810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656141199.648, "dur": 6.560,
    "args": {
      "External id": 2479108,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 6811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656141209.782, "dur": 0.907,
    "args": {
      "External id": 2479109,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 6812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656141212.638, "dur": 2.336,
    "args": {
      "External id": 2479110,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656141216.808, "dur": 1.025,
    "args": {
      "External id": 2479111,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656141219.317, "dur": 0.897,
    "args": {
      "External id": 2479112,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656141222.252, "dur": 0.567,
    "args": {
      "External id": 2479113,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656141225.899, "dur": 0.851,
    "args": {
      "External id": 2479114,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 6817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656141228.092, "dur": 2.052,
    "args": {
      "External id": 2479115,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656141231.651, "dur": 0.818,
    "args": {
      "External id": 2479116,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656141234.068, "dur": 0.612,
    "args": {
      "External id": 2479117,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656141254.165, "dur": 16710.193,
    "args": {
      "External id": 2479118,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 6821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656141270.083, "dur": 16687.961,
    "args": {
      "External id": 2479119,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 6822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656141289.257, "dur": 14.385,
    "args": {
      "External id": 2479120,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656141306.598, "dur": 16620.470,
    "args": {
      "External id": 2479121,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 6824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656141308.949, "dur": 16617.573,
    "args": {
      "External id": 2479122,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 6825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656141314.877, "dur": 5.796,
    "args": {
      "External id": 2479123,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656141322.524, "dur": 16601.345,
    "args": {
      "External id": 2479124,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 6827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656158159.017, "dur": 32.261,
    "args": {
      "External id": 2479125,"Sequence number": 24740441, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 6828
    }
  },
  {
    "ph": "s", "id": 207, "pid": 1336759, "tid": 1336759, "ts": 1514656158159.017,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514656158179.533, "dur": 7.157,
    "args": {
      "External id": 2479126,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 6829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656158182.211, "dur": 4.232,
    "args": {
      "External id": 2479127,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 6830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1336759, "tid": 1336759,
    "ts": 1514656158251.828, "dur": 71.710,
    "args": {
      "External id": 2479128,"Record function id": 0, "Ev Idx": 6831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656158325.120, "dur": 1025.282,
    "args": {
      "External id": 2479129,"Record function id": 0, "Ev Idx": 6832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656158364.233, "dur": 973.272,
    "args": {
      "External id": 2479130,"Sequence number": 24740442, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [32, 1], [32, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], "Input Dims": [[16, 4096, 2048], [2048], [2048, 2048], [2048, 2048], [2048, 2048], [8192, 32], [8192, 32], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], "Ev Idx": 6833
    }
  },
  {
    "ph": "s", "id": 206, "pid": 1336759, "tid": 1336759, "ts": 1514656158364.233,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656158424.094, "dur": 40.002,
    "args": {
      "External id": 2479131,"kernel_hash": "cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656158476.768, "dur": 118.170,
    "args": {
      "External id": 2479132,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656158608.101, "dur": 40.299,
    "args": {
      "External id": 2479133,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656158655.449, "dur": 31.780,
    "args": {
      "External id": 2479134,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656158712.918, "dur": 23.715,
    "args": {
      "External id": 2479135,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656158752.518, "dur": 13.295,
    "args": {
      "External id": 2479136,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656158783.287, "dur": 124.154,
    "args": {
      "External id": 2479137,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 6840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656158831.780, "dur": 10.405,
    "args": {
      "External id": 2479138,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 6841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656158836.701, "dur": 4.695,
    "args": {
      "External id": 2479139,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656158845.099, "dur": 5.017,
    "args": {
      "External id": 2479140,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656158851.260, "dur": 1.037,
    "args": {
      "External id": 2479141,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656158854.903, "dur": 4.866,
    "args": {
      "External id": 2479142,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656158917.698, "dur": 42.832,
    "args": {
      "External id": 2479143,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159025.697, "dur": 29.676,
    "args": {
      "External id": 2479144,"kernel_hash": "c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/3l/c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159067.531, "dur": 44.288,
    "args": {
      "External id": 2479145,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 6848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159119.743, "dur": 34.592,
    "args": {
      "External id": 2479146,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 6849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159175.287, "dur": 24.528,
    "args": {
      "External id": 2479147,"kernel_hash": "ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/kq/ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [16, 4096, 5632], []], "Ev Idx": 6850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159205.312, "dur": 33.803,
    "args": {
      "External id": 2479148,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 6851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159258.487, "dur": 17.060,
    "args": {
      "External id": 2479149,"kernel_hash": "cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u6/cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [65536, 2048], []], "Ev Idx": 6852
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.10)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159412.682, "dur": 69.346,
    "args": {
      "External id": 2479150,"Record function id": 0, "Ev Idx": 6853
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159572.955, "dur": 49.094,
    "args": {
      "External id": 2479151,"Record function id": 0, "Ev Idx": 6854
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.11)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159630.872, "dur": 18030.184,
    "args": {
      "External id": 2479152,"Record function id": 0, "Ev Idx": 6855
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.11)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159639.022, "dur": 880.722,
    "args": {
      "External id": 2479153,"Record function id": 0, "Ev Idx": 6856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159714.425, "dur": 8.062,
    "args": {
      "External id": 2479154,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159736.229, "dur": 33.223,
    "args": {
      "External id": 2479155,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 6858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159741.552, "dur": 2.121,
    "args": {
      "External id": 2479156,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159747.978, "dur": 0.366,
    "args": {
      "External id": 2479157,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159750.028, "dur": 0.295,
    "args": {
      "External id": 2479158,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159751.607, "dur": 0.424,
    "args": {
      "External id": 2479159,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159754.919, "dur": 0.372,
    "args": {
      "External id": 2479160,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159756.683, "dur": 0.322,
    "args": {
      "External id": 2479161,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159758.038, "dur": 2.835,
    "args": {
      "External id": 2479162,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159762.172, "dur": 0.306,
    "args": {
      "External id": 2479163,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159763.665, "dur": 0.173,
    "args": {
      "External id": 2479164,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159781.917, "dur": 42.732,
    "args": {
      "External id": 2479165,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 6868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159855.986, "dur": 111.221,
    "args": {
      "External id": 2479166,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 6869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159868.236, "dur": 5.143,
    "args": {
      "External id": 2479167,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159878.515, "dur": 10.182,
    "args": {
      "External id": 2479168,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 6871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159882.890, "dur": 5.385,
    "args": {
      "External id": 2479169,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 6872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159886.657, "dur": 0.524,
    "args": {
      "External id": 2479170,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 6873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159895.541, "dur": 26.556,
    "args": {
      "External id": 2479171,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 6874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159897.617, "dur": 0.256,
    "args": {
      "External id": 2479172,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159899.465, "dur": 2.023,
    "args": {
      "External id": 2479173,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159902.761, "dur": 0.482,
    "args": {
      "External id": 2479174,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159904.438, "dur": 1.808,
    "args": {
      "External id": 2479175,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159909.522, "dur": 0.459,
    "args": {
      "External id": 2479176,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159911.175, "dur": 0.160,
    "args": {
      "External id": 2479177,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159912.476, "dur": 0.402,
    "args": {
      "External id": 2479178,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159916.017, "dur": 0.478,
    "args": {
      "External id": 2479179,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159917.748, "dur": 0.368,
    "args": {
      "External id": 2479180,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656159934.445, "dur": 24.528,
    "args": {
      "External id": 2479181,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 6884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656160062.696, "dur": 363.129,
    "args": {
      "External id": 2479182,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 6885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656160092.480, "dur": 328.823,
    "args": {
      "External id": 2479183,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 6886, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1336759,
    "ts": 1514656160103.384, "dur": 312.369,
    "args": {
      "External id": 2479184,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 6887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656160451.215, "dur": 2.541,
    "args": {
      "External id": 2479185,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 6888, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.11)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656160557.644, "dur": 16903.327,
    "args": {
      "External id": 2479186,"Record function id": 0, "Ev Idx": 6889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656160651.079, "dur": 6.882,
    "args": {
      "External id": 2479187,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 6890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656160661.363, "dur": 0.975,
    "args": {
      "External id": 2479188,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 6891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656160663.766, "dur": 2.708,
    "args": {
      "External id": 2479189,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656160668.072, "dur": 1.703,
    "args": {
      "External id": 2479190,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656160671.307, "dur": 0.910,
    "args": {
      "External id": 2479191,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656160673.478, "dur": 0.874,
    "args": {
      "External id": 2479192,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656160677.412, "dur": 0.754,
    "args": {
      "External id": 2479193,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 6896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656160679.830, "dur": 2.033,
    "args": {
      "External id": 2479194,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656160683.263, "dur": 0.743,
    "args": {
      "External id": 2479195,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656160685.518, "dur": 0.730,
    "args": {
      "External id": 2479196,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656160706.286, "dur": 16717.005,
    "args": {
      "External id": 2479197,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 6900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656160722.287, "dur": 16694.359,
    "args": {
      "External id": 2479198,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 6901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656160745.510, "dur": 12.545,
    "args": {
      "External id": 2479199,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656160761.122, "dur": 16624.466,
    "args": {
      "External id": 2479200,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 6903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656160763.609, "dur": 16621.353,
    "args": {
      "External id": 2479201,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 6904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656160769.143, "dur": 5.917,
    "args": {
      "External id": 2479202,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656160776.882, "dur": 16605.286,
    "args": {
      "External id": 2479203,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 6906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656177600.954, "dur": 35.483,
    "args": {
      "External id": 2479204,"Sequence number": 24740443, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 6907
    }
  },
  {
    "ph": "s", "id": 205, "pid": 1336759, "tid": 1336759, "ts": 1514656177600.954,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514656177623.789, "dur": 8.171,
    "args": {
      "External id": 2479205,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 6908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656177627.169, "dur": 4.458,
    "args": {
      "External id": 2479206,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 6909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1336759, "tid": 1336759,
    "ts": 1514656177697.829, "dur": 76.212,
    "args": {
      "External id": 2479207,"Record function id": 0, "Ev Idx": 6910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656177775.395, "dur": 1021.475,
    "args": {
      "External id": 2479208,"Record function id": 0, "Ev Idx": 6911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656177812.798, "dur": 971.913,
    "args": {
      "External id": 2479209,"Sequence number": 24740444, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [32, 1], [32, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], "Input Dims": [[16, 4096, 2048], [2048], [2048, 2048], [2048, 2048], [2048, 2048], [8192, 32], [8192, 32], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], "Ev Idx": 6912
    }
  },
  {
    "ph": "s", "id": 204, "pid": 1336759, "tid": 1336759, "ts": 1514656177812.798,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656177874.285, "dur": 40.199,
    "args": {
      "External id": 2479210,"kernel_hash": "cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656177926.906, "dur": 133.243,
    "args": {
      "External id": 2479211,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656178073.147, "dur": 41.899,
    "args": {
      "External id": 2479212,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656178121.921, "dur": 30.290,
    "args": {
      "External id": 2479213,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656178177.505, "dur": 23.616,
    "args": {
      "External id": 2479214,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656178220.166, "dur": 14.721,
    "args": {
      "External id": 2479215,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656178252.346, "dur": 124.835,
    "args": {
      "External id": 2479216,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 6919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656178301.467, "dur": 10.036,
    "args": {
      "External id": 2479217,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 6920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656178305.936, "dur": 4.686,
    "args": {
      "External id": 2479218,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656178314.472, "dur": 5.046,
    "args": {
      "External id": 2479219,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656178320.977, "dur": 0.973,
    "args": {
      "External id": 2479220,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656178324.339, "dur": 4.922,
    "args": {
      "External id": 2479221,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656178387.530, "dur": 42.817,
    "args": {
      "External id": 2479222,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 1336759, "tid": 1336759,
    "ts": 1514656178456.366, "dur": 25.991,
    "args": {
      "External id": 2479223,"kernel_hash": "c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/3l/c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656178490.631, "dur": 54.852,
    "args": {
      "External id": 2479224,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 6927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656178557.376, "dur": 37.837,
    "args": {
      "External id": 2479225,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 6928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656178617.268, "dur": 25.128,
    "args": {
      "External id": 2479226,"kernel_hash": "ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/kq/ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [16, 4096, 5632], []], "Ev Idx": 6929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656178648.737, "dur": 33.486,
    "args": {
      "External id": 2479227,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 6930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656178701.356, "dur": 17.662,
    "args": {
      "External id": 2479228,"kernel_hash": "cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u6/cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [65536, 2048], []], "Ev Idx": 6931
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.11)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656178856.748, "dur": 71.124,
    "args": {
      "External id": 2479229,"Record function id": 0, "Ev Idx": 6932
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179040.556, "dur": 47.469,
    "args": {
      "External id": 2479230,"Record function id": 0, "Ev Idx": 6933
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.12)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179097.122, "dur": 18105.540,
    "args": {
      "External id": 2479231,"Record function id": 0, "Ev Idx": 6934
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.12)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179104.890, "dur": 853.718,
    "args": {
      "External id": 2479232,"Record function id": 0, "Ev Idx": 6935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179182.517, "dur": 8.852,
    "args": {
      "External id": 2479233,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179207.112, "dur": 36.098,
    "args": {
      "External id": 2479234,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 6937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179212.884, "dur": 2.338,
    "args": {
      "External id": 2479235,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179218.795, "dur": 0.350,
    "args": {
      "External id": 2479236,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179220.469, "dur": 0.551,
    "args": {
      "External id": 2479237,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179222.522, "dur": 2.509,
    "args": {
      "External id": 2479238,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179226.598, "dur": 0.494,
    "args": {
      "External id": 2479239,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179228.607, "dur": 0.483,
    "args": {
      "External id": 2479240,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179231.953, "dur": 1.976,
    "args": {
      "External id": 2479241,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179235.214, "dur": 0.214,
    "args": {
      "External id": 2479242,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179236.852, "dur": 0.435,
    "args": {
      "External id": 2479243,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179253.974, "dur": 40.766,
    "args": {
      "External id": 2479244,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 6947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179327.392, "dur": 107.620,
    "args": {
      "External id": 2479245,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 6948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179339.435, "dur": 3.852,
    "args": {
      "External id": 2479246,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179348.328, "dur": 11.516,
    "args": {
      "External id": 2479247,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 6950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179352.558, "dur": 6.891,
    "args": {
      "External id": 2479248,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 6951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179356.064, "dur": 2.248,
    "args": {
      "External id": 2479249,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 6952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179366.857, "dur": 28.369,
    "args": {
      "External id": 2479250,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 6953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179369.203, "dur": 0.438,
    "args": {
      "External id": 2479251,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179371.463, "dur": 0.449,
    "args": {
      "External id": 2479252,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179373.353, "dur": 0.245,
    "args": {
      "External id": 2479253,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179376.332, "dur": 2.020,
    "args": {
      "External id": 2479254,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179380.029, "dur": 0.353,
    "args": {
      "External id": 2479255,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179381.653, "dur": 2.026,
    "args": {
      "External id": 2479256,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179384.796, "dur": 0.254,
    "args": {
      "External id": 2479257,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179386.674, "dur": 0.283,
    "args": {
      "External id": 2479258,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179389.949, "dur": 0.399,
    "args": {
      "External id": 2479259,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 6962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179404.724, "dur": 22.946,
    "args": {
      "External id": 2479260,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 6963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179484.025, "dur": 379.542,
    "args": {
      "External id": 2479261,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 6964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179512.316, "dur": 346.181,
    "args": {
      "External id": 2479262,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 6965, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179523.782, "dur": 329.176,
    "args": {
      "External id": 2479263,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 6966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179887.362, "dur": 2.671,
    "args": {
      "External id": 2479264,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 6967, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.12)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656179978.986, "dur": 17040.369,
    "args": {
      "External id": 2479265,"Record function id": 0, "Ev Idx": 6968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656180110.472, "dur": 6.802,
    "args": {
      "External id": 2479266,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 6969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656180120.815, "dur": 1.159,
    "args": {
      "External id": 2479267,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 6970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656180123.760, "dur": 2.296,
    "args": {
      "External id": 2479268,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656180127.719, "dur": 1.177,
    "args": {
      "External id": 2479269,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656180130.200, "dur": 1.062,
    "args": {
      "External id": 2479270,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656180132.627, "dur": 0.775,
    "args": {
      "External id": 2479271,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 6974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656180136.620, "dur": 0.766,
    "args": {
      "External id": 2479272,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 6975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656180138.838, "dur": 1.739,
    "args": {
      "External id": 2479273,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656180142.140, "dur": 0.625,
    "args": {
      "External id": 2479274,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656180144.398, "dur": 0.619,
    "args": {
      "External id": 2479275,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 6978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656180165.999, "dur": 16793.144,
    "args": {
      "External id": 2479276,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 6979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656180181.463, "dur": 16771.025,
    "args": {
      "External id": 2479277,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 6980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656180207.171, "dur": 14.040,
    "args": {
      "External id": 2479278,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656180224.288, "dur": 16698.534,
    "args": {
      "External id": 2479279,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 6982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656180227.231, "dur": 16695.143,
    "args": {
      "External id": 2479280,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 6983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656180232.921, "dur": 5.696,
    "args": {
      "External id": 2479281,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 6984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656180240.314, "dur": 16679.231,
    "args": {
      "External id": 2479282,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 6985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656197147.005, "dur": 33.074,
    "args": {
      "External id": 2479283,"Sequence number": 24740445, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 6986
    }
  },
  {
    "ph": "s", "id": 203, "pid": 1336759, "tid": 1336759, "ts": 1514656197147.005,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514656197167.897, "dur": 7.437,
    "args": {
      "External id": 2479284,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 6987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656197170.908, "dur": 4.195,
    "args": {
      "External id": 2479285,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 6988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1336759, "tid": 1336759,
    "ts": 1514656197239.236, "dur": 72.893,
    "args": {
      "External id": 2479286,"Record function id": 0, "Ev Idx": 6989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656197313.624, "dur": 1046.141,
    "args": {
      "External id": 2479287,"Record function id": 0, "Ev Idx": 6990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656197353.382, "dur": 992.771,
    "args": {
      "External id": 2479288,"Sequence number": 24740446, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [32, 1], [32, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], "Input Dims": [[16, 4096, 2048], [2048], [2048, 2048], [2048, 2048], [2048, 2048], [8192, 32], [8192, 32], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], "Ev Idx": 6991
    }
  },
  {
    "ph": "s", "id": 202, "pid": 1336759, "tid": 1336759, "ts": 1514656197353.382,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656197415.208, "dur": 38.736,
    "args": {
      "External id": 2479289,"kernel_hash": "cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 6992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656197465.669, "dur": 121.883,
    "args": {
      "External id": 2479290,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656197600.006, "dur": 40.944,
    "args": {
      "External id": 2479291,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656197649.243, "dur": 31.596,
    "args": {
      "External id": 2479292,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 6995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656197706.111, "dur": 26.119,
    "args": {
      "External id": 2479293,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656197749.570, "dur": 14.474,
    "args": {
      "External id": 2479294,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 6997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656197781.670, "dur": 122.575,
    "args": {
      "External id": 2479295,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 6998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656197829.182, "dur": 10.966,
    "args": {
      "External id": 2479296,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 6999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656197834.063, "dur": 5.303,
    "args": {
      "External id": 2479297,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656197842.977, "dur": 5.043,
    "args": {
      "External id": 2479298,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656197849.316, "dur": 0.845,
    "args": {
      "External id": 2479299,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656197853.001, "dur": 3.933,
    "args": {
      "External id": 2479300,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656197914.400, "dur": 45.806,
    "args": {
      "External id": 2479301,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198027.140, "dur": 28.952,
    "args": {
      "External id": 2479302,"kernel_hash": "c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/3l/c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 7005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198067.274, "dur": 45.287,
    "args": {
      "External id": 2479303,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 7006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198120.652, "dur": 34.976,
    "args": {
      "External id": 2479304,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 7007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198181.921, "dur": 25.056,
    "args": {
      "External id": 2479305,"kernel_hash": "ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/kq/ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [16, 4096, 5632], []], "Ev Idx": 7008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198213.903, "dur": 33.791,
    "args": {
      "External id": 2479306,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 7009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198265.647, "dur": 17.206,
    "args": {
      "External id": 2479307,"kernel_hash": "cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u6/cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [65536, 2048], []], "Ev Idx": 7010
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.12)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198418.686, "dur": 71.583,
    "args": {
      "External id": 2479308,"Record function id": 0, "Ev Idx": 7011
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198581.685, "dur": 48.783,
    "args": {
      "External id": 2479309,"Record function id": 0, "Ev Idx": 7012
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.13)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198639.878, "dur": 18127.191,
    "args": {
      "External id": 2479310,"Record function id": 0, "Ev Idx": 7013
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.13)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198649.869, "dur": 853.258,
    "args": {
      "External id": 2479311,"Record function id": 0, "Ev Idx": 7014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198727.534, "dur": 8.699,
    "args": {
      "External id": 2479312,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198749.114, "dur": 34.525,
    "args": {
      "External id": 2479313,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 7016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198754.366, "dur": 2.233,
    "args": {
      "External id": 2479314,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198760.852, "dur": 0.422,
    "args": {
      "External id": 2479315,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198762.646, "dur": 0.233,
    "args": {
      "External id": 2479316,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198764.203, "dur": 0.453,
    "args": {
      "External id": 2479317,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198767.704, "dur": 0.495,
    "args": {
      "External id": 2479318,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198769.275, "dur": 0.518,
    "args": {
      "External id": 2479319,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198770.973, "dur": 3.339,
    "args": {
      "External id": 2479320,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198775.732, "dur": 0.175,
    "args": {
      "External id": 2479321,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198777.606, "dur": 0.152,
    "args": {
      "External id": 2479322,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198794.357, "dur": 39.721,
    "args": {
      "External id": 2479323,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 7026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198865.927, "dur": 107.448,
    "args": {
      "External id": 2479324,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 7027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198878.142, "dur": 3.643,
    "args": {
      "External id": 2479325,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198886.432, "dur": 10.037,
    "args": {
      "External id": 2479326,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 7029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198890.723, "dur": 5.363,
    "args": {
      "External id": 2479327,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 7030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198894.369, "dur": 0.504,
    "args": {
      "External id": 2479328,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 7031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198903.095, "dur": 28.969,
    "args": {
      "External id": 2479329,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 7032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198905.836, "dur": 2.143,
    "args": {
      "External id": 2479330,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198910.343, "dur": 0.509,
    "args": {
      "External id": 2479331,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198912.334, "dur": 0.295,
    "args": {
      "External id": 2479332,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198915.791, "dur": 1.278,
    "args": {
      "External id": 2479333,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198918.391, "dur": 0.376,
    "args": {
      "External id": 2479334,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198919.898, "dur": 0.306,
    "args": {
      "External id": 2479335,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198923.301, "dur": 0.243,
    "args": {
      "External id": 2479336,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198925.096, "dur": 0.304,
    "args": {
      "External id": 2479337,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198926.777, "dur": 1.510,
    "args": {
      "External id": 2479338,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656198943.470, "dur": 22.270,
    "args": {
      "External id": 2479339,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 7042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656199065.537, "dur": 348.013,
    "args": {
      "External id": 2479340,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 7043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656199096.686, "dur": 311.998,
    "args": {
      "External id": 2479341,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 7044, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1336759,
    "ts": 1514656199107.119, "dur": 296.293,
    "args": {
      "External id": 2479342,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 7045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656199435.055, "dur": 2.464,
    "args": {
      "External id": 2479343,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 7046, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.13)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656199522.866, "dur": 17054.564,
    "args": {
      "External id": 2479344,"Record function id": 0, "Ev Idx": 7047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656199635.024, "dur": 6.675,
    "args": {
      "External id": 2479345,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 7048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656199645.047, "dur": 1.117,
    "args": {
      "External id": 2479346,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 7049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656199647.887, "dur": 2.083,
    "args": {
      "External id": 2479347,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656199651.701, "dur": 0.879,
    "args": {
      "External id": 2479348,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656199654.215, "dur": 0.853,
    "args": {
      "External id": 2479349,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656199656.486, "dur": 0.752,
    "args": {
      "External id": 2479350,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656199660.782, "dur": 0.839,
    "args": {
      "External id": 2479351,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 7054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656199663.124, "dur": 2.329,
    "args": {
      "External id": 2479352,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656199667.024, "dur": 0.684,
    "args": {
      "External id": 2479353,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656199669.194, "dur": 0.790,
    "args": {
      "External id": 2479354,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656199689.218, "dur": 16847.883,
    "args": {
      "External id": 2479355,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 7058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656199704.659, "dur": 16815.087,
    "args": {
      "External id": 2479356,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 7059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656199728.399, "dur": 13.421,
    "args": {
      "External id": 2479357,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656199744.562, "dur": 16745.091,
    "args": {
      "External id": 2479358,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 7061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656199747.057, "dur": 16742.101,
    "args": {
      "External id": 2479359,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 7062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656199753.214, "dur": 5.052,
    "args": {
      "External id": 2479360,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656199759.831, "dur": 16726.516,
    "args": {
      "External id": 2479361,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 7064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656216709.945, "dur": 32.611,
    "args": {
      "External id": 2479362,"Sequence number": 24740447, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 7065
    }
  },
  {
    "ph": "s", "id": 201, "pid": 1336759, "tid": 1336759, "ts": 1514656216709.945,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514656216730.378, "dur": 7.662,
    "args": {
      "External id": 2479363,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 7066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656216733.222, "dur": 4.545,
    "args": {
      "External id": 2479364,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 7067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1336759, "tid": 1336759,
    "ts": 1514656216803.514, "dur": 73.316,
    "args": {
      "External id": 2479365,"Record function id": 0, "Ev Idx": 7068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656216878.614, "dur": 1036.565,
    "args": {
      "External id": 2479366,"Record function id": 0, "Ev Idx": 7069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656216917.568, "dur": 985.159,
    "args": {
      "External id": 2479367,"Sequence number": 24740448, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [32, 1], [32, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], "Input Dims": [[16, 4096, 2048], [2048], [2048, 2048], [2048, 2048], [2048, 2048], [8192, 32], [8192, 32], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], "Ev Idx": 7070
    }
  },
  {
    "ph": "s", "id": 200, "pid": 1336759, "tid": 1336759, "ts": 1514656216917.568,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656216980.867, "dur": 85.861,
    "args": {
      "External id": 2479368,"kernel_hash": "cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 7071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656217081.001, "dur": 102.278,
    "args": {
      "External id": 2479369,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656217192.990, "dur": 38.426,
    "args": {
      "External id": 2479370,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656217239.585, "dur": 30.449,
    "args": {
      "External id": 2479371,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656217295.885, "dur": 23.465,
    "args": {
      "External id": 2479372,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 7075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656217336.535, "dur": 13.941,
    "args": {
      "External id": 2479373,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 7076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656217368.421, "dur": 123.998,
    "args": {
      "External id": 2479374,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 7077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656217417.260, "dur": 11.051,
    "args": {
      "External id": 2479375,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 7078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656217422.169, "dur": 5.300,
    "args": {
      "External id": 2479376,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656217431.117, "dur": 5.509,
    "args": {
      "External id": 2479377,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656217437.837, "dur": 1.260,
    "args": {
      "External id": 2479378,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656217441.663, "dur": 3.603,
    "args": {
      "External id": 2479379,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656217502.304, "dur": 60.788,
    "args": {
      "External id": 2479380,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 1336759, "tid": 1336759,
    "ts": 1514656217595.614, "dur": 25.555,
    "args": {
      "External id": 2479381,"kernel_hash": "c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/3l/c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 7084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656217630.186, "dur": 43.507,
    "args": {
      "External id": 2479382,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 7085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656217681.820, "dur": 35.229,
    "args": {
      "External id": 2479383,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 7086
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656217738.968, "dur": 23.737,
    "args": {
      "External id": 2479384,"kernel_hash": "ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/kq/ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [16, 4096, 5632], []], "Ev Idx": 7087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656217768.048, "dur": 34.116,
    "args": {
      "External id": 2479385,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 7088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656217821.099, "dur": 16.874,
    "args": {
      "External id": 2479386,"kernel_hash": "cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u6/cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [65536, 2048], []], "Ev Idx": 7089
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.13)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656217977.277, "dur": 109.705,
    "args": {
      "External id": 2479387,"Record function id": 0, "Ev Idx": 7090
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218163.241, "dur": 46.207,
    "args": {
      "External id": 2479388,"Record function id": 0, "Ev Idx": 7091
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.14)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218218.961, "dur": 18090.938,
    "args": {
      "External id": 2479389,"Record function id": 0, "Ev Idx": 7092
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.14)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218227.856, "dur": 942.157,
    "args": {
      "External id": 2479390,"Record function id": 0, "Ev Idx": 7093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218305.701, "dur": 8.129,
    "args": {
      "External id": 2479391,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218326.679, "dur": 42.414,
    "args": {
      "External id": 2479392,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 7095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218335.324, "dur": 2.248,
    "args": {
      "External id": 2479393,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218342.760, "dur": 0.569,
    "args": {
      "External id": 2479394,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218344.727, "dur": 0.416,
    "args": {
      "External id": 2479395,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218346.305, "dur": 1.961,
    "args": {
      "External id": 2479396,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218349.756, "dur": 0.269,
    "args": {
      "External id": 2479397,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218353.182, "dur": 0.285,
    "args": {
      "External id": 2479398,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218356.558, "dur": 1.515,
    "args": {
      "External id": 2479399,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218359.317, "dur": 0.434,
    "args": {
      "External id": 2479400,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218361.025, "dur": 1.902,
    "args": {
      "External id": 2479401,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218379.601, "dur": 38.471,
    "args": {
      "External id": 2479402,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 7105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218449.714, "dur": 120.343,
    "args": {
      "External id": 2479403,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 7106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218460.346, "dur": 3.462,
    "args": {
      "External id": 2479404,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218468.831, "dur": 9.875,
    "args": {
      "External id": 2479405,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 7108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218473.328, "dur": 4.993,
    "args": {
      "External id": 2479406,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 7109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218476.680, "dur": 0.437,
    "args": {
      "External id": 2479407,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 7110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218485.638, "dur": 26.536,
    "args": {
      "External id": 2479408,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 7111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218487.918, "dur": 0.315,
    "args": {
      "External id": 2479409,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218489.957, "dur": 0.200,
    "args": {
      "External id": 2479410,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218491.362, "dur": 2.044,
    "args": {
      "External id": 2479411,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218494.614, "dur": 1.436,
    "args": {
      "External id": 2479412,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218497.200, "dur": 0.196,
    "args": {
      "External id": 2479413,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218500.322, "dur": 0.316,
    "args": {
      "External id": 2479414,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218502.331, "dur": 0.247,
    "args": {
      "External id": 2479415,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218504.041, "dur": 0.392,
    "args": {
      "External id": 2479416,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218507.318, "dur": 0.276,
    "args": {
      "External id": 2479417,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218521.842, "dur": 39.107,
    "args": {
      "External id": 2479418,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 7121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218624.253, "dur": 440.801,
    "args": {
      "External id": 2479419,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 7122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218655.814, "dur": 403.870,
    "args": {
      "External id": 2479420,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 7123, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1336759,
    "ts": 1514656218666.374, "dur": 385.628,
    "args": {
      "External id": 2479421,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 7124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656219092.414, "dur": 2.517,
    "args": {
      "External id": 2479422,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 7125, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.14)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656219190.931, "dur": 16932.650,
    "args": {
      "External id": 2479423,"Record function id": 0, "Ev Idx": 7126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656219284.928, "dur": 6.509,
    "args": {
      "External id": 2479424,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 7127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656219294.867, "dur": 0.872,
    "args": {
      "External id": 2479425,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 7128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656219297.403, "dur": 2.198,
    "args": {
      "External id": 2479426,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656219301.105, "dur": 0.875,
    "args": {
      "External id": 2479427,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656219303.329, "dur": 0.763,
    "args": {
      "External id": 2479428,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656219305.298, "dur": 0.934,
    "args": {
      "External id": 2479429,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656219309.600, "dur": 0.981,
    "args": {
      "External id": 2479430,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 7133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656219312.052, "dur": 1.972,
    "args": {
      "External id": 2479431,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656219315.404, "dur": 0.535,
    "args": {
      "External id": 2479432,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656219317.417, "dur": 0.543,
    "args": {
      "External id": 2479433,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656219337.421, "dur": 16749.037,
    "args": {
      "External id": 2479434,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 7137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656219353.263, "dur": 16726.733,
    "args": {
      "External id": 2479435,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 7138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656219377.271, "dur": 12.846,
    "args": {
      "External id": 2479436,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656219393.325, "dur": 16655.540,
    "args": {
      "External id": 2479437,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 7140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656219395.719, "dur": 16652.543,
    "args": {
      "External id": 2479438,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 7141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656219401.411, "dur": 5.656,
    "args": {
      "External id": 2479439,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656219408.966, "dur": 16636.194,
    "args": {
      "External id": 2479440,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 7143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656236253.184, "dur": 32.522,
    "args": {
      "External id": 2479441,"Sequence number": 24740449, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 7144
    }
  },
  {
    "ph": "s", "id": 199, "pid": 1336759, "tid": 1336759, "ts": 1514656236253.184,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514656236273.370, "dur": 7.575,
    "args": {
      "External id": 2479442,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 7145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656236276.454, "dur": 4.194,
    "args": {
      "External id": 2479443,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 7146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1336759, "tid": 1336759,
    "ts": 1514656236346.739, "dur": 74.583,
    "args": {
      "External id": 2479444,"Record function id": 0, "Ev Idx": 7147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656236422.803, "dur": 1047.650,
    "args": {
      "External id": 2479445,"Record function id": 0, "Ev Idx": 7148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656236460.843, "dur": 996.619,
    "args": {
      "External id": 2479446,"Sequence number": 24740450, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [32, 1], [32, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], "Input Dims": [[16, 4096, 2048], [2048], [2048, 2048], [2048, 2048], [2048, 2048], [8192, 32], [8192, 32], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], "Ev Idx": 7149
    }
  },
  {
    "ph": "s", "id": 198, "pid": 1336759, "tid": 1336759, "ts": 1514656236460.843,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656236521.794, "dur": 55.330,
    "args": {
      "External id": 2479447,"kernel_hash": "cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 7150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656236592.697, "dur": 102.067,
    "args": {
      "External id": 2479448,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656236704.139, "dur": 39.279,
    "args": {
      "External id": 2479449,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656236751.503, "dur": 32.136,
    "args": {
      "External id": 2479450,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656236808.695, "dur": 24.133,
    "args": {
      "External id": 2479451,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 7154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656236849.308, "dur": 13.038,
    "args": {
      "External id": 2479452,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 7155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656236879.017, "dur": 173.339,
    "args": {
      "External id": 2479453,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 7156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656236933.220, "dur": 16.984,
    "args": {
      "External id": 2479454,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 7157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656236943.735, "dur": 5.492,
    "args": {
      "External id": 2479455,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656236952.793, "dur": 5.606,
    "args": {
      "External id": 2479456,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656236959.767, "dur": 1.179,
    "args": {
      "External id": 2479457,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656236963.669, "dur": 3.477,
    "args": {
      "External id": 2479458,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656237064.912, "dur": 51.048,
    "args": {
      "External id": 2479459,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 1336759, "tid": 1336759,
    "ts": 1514656237148.397, "dur": 26.458,
    "args": {
      "External id": 2479460,"kernel_hash": "c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/3l/c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 7163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656237183.815, "dur": 41.459,
    "args": {
      "External id": 2479461,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 7164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656237233.833, "dur": 34.473,
    "args": {
      "External id": 2479462,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 7165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656237289.850, "dur": 25.089,
    "args": {
      "External id": 2479463,"kernel_hash": "ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/kq/ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [16, 4096, 5632], []], "Ev Idx": 7166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656237320.708, "dur": 33.538,
    "args": {
      "External id": 2479464,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 7167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656237375.340, "dur": 17.191,
    "args": {
      "External id": 2479465,"kernel_hash": "cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u6/cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [65536, 2048], []], "Ev Idx": 7168
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.14)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656237547.299, "dur": 73.018,
    "args": {
      "External id": 2479466,"Record function id": 0, "Ev Idx": 7169
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656237693.290, "dur": 44.909,
    "args": {
      "External id": 2479467,"Record function id": 0, "Ev Idx": 7170
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.15)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656237747.255, "dur": 18089.377,
    "args": {
      "External id": 2479468,"Record function id": 0, "Ev Idx": 7171
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.15)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656237755.474, "dur": 890.099,
    "args": {
      "External id": 2479469,"Record function id": 0, "Ev Idx": 7172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656237830.428, "dur": 8.986,
    "args": {
      "External id": 2479470,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656237852.489, "dur": 34.002,
    "args": {
      "External id": 2479471,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 7174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656237857.750, "dur": 2.236,
    "args": {
      "External id": 2479472,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656237864.426, "dur": 0.343,
    "args": {
      "External id": 2479473,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656237866.154, "dur": 0.421,
    "args": {
      "External id": 2479474,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656237867.844, "dur": 0.210,
    "args": {
      "External id": 2479475,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656237870.849, "dur": 0.305,
    "args": {
      "External id": 2479476,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656237872.342, "dur": 0.361,
    "args": {
      "External id": 2479477,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656237874.061, "dur": 3.403,
    "args": {
      "External id": 2479478,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656237878.630, "dur": 0.208,
    "args": {
      "External id": 2479479,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656237880.098, "dur": 0.276,
    "args": {
      "External id": 2479480,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656237896.680, "dur": 40.665,
    "args": {
      "External id": 2479481,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 7184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1336759,
    "ts": 1514656237969.777, "dur": 164.901,
    "args": {
      "External id": 2479482,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 7185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656237980.053, "dur": 42.260,
    "args": {
      "External id": 2479483,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238029.720, "dur": 11.387,
    "args": {
      "External id": 2479484,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 7187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238034.021, "dur": 6.707,
    "args": {
      "External id": 2479485,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 7188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238038.268, "dur": 0.772,
    "args": {
      "External id": 2479486,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 7189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238049.234, "dur": 28.569,
    "args": {
      "External id": 2479487,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 7190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238051.751, "dur": 2.170,
    "args": {
      "External id": 2479488,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238055.159, "dur": 0.341,
    "args": {
      "External id": 2479489,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238056.813, "dur": 0.447,
    "args": {
      "External id": 2479490,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238060.499, "dur": 1.777,
    "args": {
      "External id": 2479491,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238063.536, "dur": 0.269,
    "args": {
      "External id": 2479492,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238065.118, "dur": 0.317,
    "args": {
      "External id": 2479493,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238068.163, "dur": 0.402,
    "args": {
      "External id": 2479494,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238069.792, "dur": 0.323,
    "args": {
      "External id": 2479495,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238071.450, "dur": 1.993,
    "args": {
      "External id": 2479496,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238098.378, "dur": 26.918,
    "args": {
      "External id": 2479497,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 7200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238188.435, "dur": 362.082,
    "args": {
      "External id": 2479498,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 7201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238218.771, "dur": 326.344,
    "args": {
      "External id": 2479499,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 7202, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238230.417, "dur": 291.968,
    "args": {
      "External id": 2479500,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 7203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238575.007, "dur": 2.689,
    "args": {
      "External id": 2479501,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 7204, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.15)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238665.807, "dur": 16987.219,
    "args": {
      "External id": 2479502,"Record function id": 0, "Ev Idx": 7205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238758.201, "dur": 6.273,
    "args": {
      "External id": 2479503,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 7206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238768.035, "dur": 1.056,
    "args": {
      "External id": 2479504,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 7207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238770.724, "dur": 2.348,
    "args": {
      "External id": 2479505,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238774.787, "dur": 0.838,
    "args": {
      "External id": 2479506,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238777.008, "dur": 0.856,
    "args": {
      "External id": 2479507,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238779.212, "dur": 0.897,
    "args": {
      "External id": 2479508,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238783.138, "dur": 1.118,
    "args": {
      "External id": 2479509,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 7212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238785.658, "dur": 2.090,
    "args": {
      "External id": 2479510,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238789.163, "dur": 0.677,
    "args": {
      "External id": 2479511,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238791.417, "dur": 0.855,
    "args": {
      "External id": 2479512,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238810.914, "dur": 16804.566,
    "args": {
      "External id": 2479513,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 7216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238826.304, "dur": 16782.676,
    "args": {
      "External id": 2479514,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 7217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238849.069, "dur": 14.962,
    "args": {
      "External id": 2479515,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238867.314, "dur": 16711.411,
    "args": {
      "External id": 2479516,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 7219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238869.896, "dur": 16708.235,
    "args": {
      "External id": 2479517,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 7220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238875.472, "dur": 5.143,
    "args": {
      "External id": 2479518,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656238882.356, "dur": 16692.963,
    "args": {
      "External id": 2479519,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 7222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656255781.268, "dur": 32.510,
    "args": {
      "External id": 2479520,"Sequence number": 24740451, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 7223
    }
  },
  {
    "ph": "s", "id": 197, "pid": 1336759, "tid": 1336759, "ts": 1514656255781.268,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514656255802.069, "dur": 7.367,
    "args": {
      "External id": 2479521,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 7224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656255804.976, "dur": 4.283,
    "args": {
      "External id": 2479522,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 7225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1336759, "tid": 1336759,
    "ts": 1514656255871.977, "dur": 77.305,
    "args": {
      "External id": 2479523,"Record function id": 0, "Ev Idx": 7226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656255950.640, "dur": 1085.310,
    "args": {
      "External id": 2479524,"Record function id": 0, "Ev Idx": 7227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656256033.571, "dur": 987.485,
    "args": {
      "External id": 2479525,"Sequence number": 24740452, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [32, 1], [32, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], "Input Dims": [[16, 4096, 2048], [2048], [2048, 2048], [2048, 2048], [2048, 2048], [8192, 32], [8192, 32], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], "Ev Idx": 7228
    }
  },
  {
    "ph": "s", "id": 196, "pid": 1336759, "tid": 1336759, "ts": 1514656256033.571,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656256100.463, "dur": 41.632,
    "args": {
      "External id": 2479526,"kernel_hash": "cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 7229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656256154.781, "dur": 102.792,
    "args": {
      "External id": 2479527,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656256267.485, "dur": 38.411,
    "args": {
      "External id": 2479528,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656256314.188, "dur": 31.210,
    "args": {
      "External id": 2479529,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656256369.600, "dur": 23.573,
    "args": {
      "External id": 2479530,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 7233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656256410.333, "dur": 14.765,
    "args": {
      "External id": 2479531,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 7234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656256442.064, "dur": 142.990,
    "args": {
      "External id": 2479532,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 7235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656256487.529, "dur": 11.799,
    "args": {
      "External id": 2479533,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 7236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656256492.571, "dur": 5.781,
    "args": {
      "External id": 2479534,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656256502.334, "dur": 5.316,
    "args": {
      "External id": 2479535,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656256508.981, "dur": 1.500,
    "args": {
      "External id": 2479536,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656256512.887, "dur": 3.950,
    "args": {
      "External id": 2479537,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656256596.989, "dur": 50.485,
    "args": {
      "External id": 2479538,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 1336759, "tid": 1336759,
    "ts": 1514656256677.137, "dur": 27.567,
    "args": {
      "External id": 2479539,"kernel_hash": "c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/3l/c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 7242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656256713.424, "dur": 41.053,
    "args": {
      "External id": 2479540,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 7243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656256762.817, "dur": 34.537,
    "args": {
      "External id": 2479541,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 7244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656256819.484, "dur": 24.327,
    "args": {
      "External id": 2479542,"kernel_hash": "ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/kq/ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [16, 4096, 5632], []], "Ev Idx": 7245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656256849.033, "dur": 33.725,
    "args": {
      "External id": 2479543,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 7246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656256901.330, "dur": 17.080,
    "args": {
      "External id": 2479544,"kernel_hash": "cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u6/cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [65536, 2048], []], "Ev Idx": 7247
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.15)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257098.757, "dur": 72.497,
    "args": {
      "External id": 2479545,"Record function id": 0, "Ev Idx": 7248
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257244.400, "dur": 46.644,
    "args": {
      "External id": 2479546,"Record function id": 0, "Ev Idx": 7249
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.16)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257300.075, "dur": 18161.473,
    "args": {
      "External id": 2479547,"Record function id": 0, "Ev Idx": 7250
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.16)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257308.100, "dur": 903.504,
    "args": {
      "External id": 2479548,"Record function id": 0, "Ev Idx": 7251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257385.794, "dur": 8.679,
    "args": {
      "External id": 2479549,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257407.877, "dur": 34.975,
    "args": {
      "External id": 2479550,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 7253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257413.302, "dur": 2.180,
    "args": {
      "External id": 2479551,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257420.314, "dur": 0.416,
    "args": {
      "External id": 2479552,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257421.901, "dur": 0.558,
    "args": {
      "External id": 2479553,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257423.877, "dur": 0.299,
    "args": {
      "External id": 2479554,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257427.041, "dur": 0.469,
    "args": {
      "External id": 2479555,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257428.763, "dur": 0.370,
    "args": {
      "External id": 2479556,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257430.413, "dur": 3.287,
    "args": {
      "External id": 2479557,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257435.052, "dur": 0.274,
    "args": {
      "External id": 2479558,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257436.782, "dur": 0.208,
    "args": {
      "External id": 2479559,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257456.235, "dur": 39.311,
    "args": {
      "External id": 2479560,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 7263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257543.917, "dur": 112.314,
    "args": {
      "External id": 2479561,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 7264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257556.405, "dur": 6.627,
    "args": {
      "External id": 2479562,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257568.293, "dur": 10.951,
    "args": {
      "External id": 2479563,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 7266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257572.741, "dur": 5.961,
    "args": {
      "External id": 2479564,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 7267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257576.439, "dur": 0.702,
    "args": {
      "External id": 2479565,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 7268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257586.472, "dur": 27.391,
    "args": {
      "External id": 2479566,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 7269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257589.173, "dur": 0.426,
    "args": {
      "External id": 2479567,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257590.983, "dur": 2.124,
    "args": {
      "External id": 2479568,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257594.279, "dur": 0.483,
    "args": {
      "External id": 2479569,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257596.130, "dur": 1.419,
    "args": {
      "External id": 2479570,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257600.599, "dur": 0.317,
    "args": {
      "External id": 2479571,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257602.142, "dur": 0.414,
    "args": {
      "External id": 2479572,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257603.788, "dur": 0.314,
    "args": {
      "External id": 2479573,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257606.956, "dur": 0.457,
    "args": {
      "External id": 2479574,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257608.840, "dur": 0.481,
    "args": {
      "External id": 2479575,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257625.132, "dur": 23.399,
    "args": {
      "External id": 2479576,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 7279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257709.649, "dur": 406.367,
    "args": {
      "External id": 2479577,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 7280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257741.102, "dur": 369.444,
    "args": {
      "External id": 2479578,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 7281, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1336759,
    "ts": 1514656257751.385, "dur": 352.691,
    "args": {
      "External id": 2479579,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 7282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656258140.950, "dur": 2.592,
    "args": {
      "External id": 2479580,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 7283, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.16)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656258233.011, "dur": 17044.656,
    "args": {
      "External id": 2479581,"Record function id": 0, "Ev Idx": 7284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656258328.847, "dur": 6.474,
    "args": {
      "External id": 2479582,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 7285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656258338.496, "dur": 1.077,
    "args": {
      "External id": 2479583,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 7286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656258341.289, "dur": 2.225,
    "args": {
      "External id": 2479584,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656258345.171, "dur": 1.154,
    "args": {
      "External id": 2479585,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656258347.613, "dur": 0.842,
    "args": {
      "External id": 2479586,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656258349.651, "dur": 0.747,
    "args": {
      "External id": 2479587,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656258354.073, "dur": 0.794,
    "args": {
      "External id": 2479588,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 7291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656258356.343, "dur": 1.928,
    "args": {
      "External id": 2479589,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656258359.707, "dur": 0.832,
    "args": {
      "External id": 2479590,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656258362.042, "dur": 0.747,
    "args": {
      "External id": 2479591,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656258382.026, "dur": 16860.263,
    "args": {
      "External id": 2479592,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 7295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656258397.281, "dur": 16838.596,
    "args": {
      "External id": 2479593,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 7296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656258421.213, "dur": 13.768,
    "args": {
      "External id": 2479594,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656258437.967, "dur": 16768.292,
    "args": {
      "External id": 2479595,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 7298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656258440.629, "dur": 16765.035,
    "args": {
      "External id": 2479596,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 7299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656258445.905, "dur": 5.165,
    "args": {
      "External id": 2479597,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656258452.878, "dur": 16749.865,
    "args": {
      "External id": 2479598,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 7301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656275404.489, "dur": 33.663,
    "args": {
      "External id": 2479599,"Sequence number": 24740453, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 7302
    }
  },
  {
    "ph": "s", "id": 195, "pid": 1336759, "tid": 1336759, "ts": 1514656275404.489,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514656275425.978, "dur": 7.461,
    "args": {
      "External id": 2479600,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 7303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656275429.031, "dur": 4.175,
    "args": {
      "External id": 2479601,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 7304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1336759, "tid": 1336759,
    "ts": 1514656275497.682, "dur": 87.743,
    "args": {
      "External id": 2479602,"Record function id": 0, "Ev Idx": 7305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656275588.615, "dur": 1080.553,
    "args": {
      "External id": 2479603,"Record function id": 0, "Ev Idx": 7306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656275628.295, "dur": 1028.315,
    "args": {
      "External id": 2479604,"Sequence number": 24740454, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [32, 1], [32, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], "Input Dims": [[16, 4096, 2048], [2048], [2048, 2048], [2048, 2048], [2048, 2048], [8192, 32], [8192, 32], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], "Ev Idx": 7307
    }
  },
  {
    "ph": "s", "id": 194, "pid": 1336759, "tid": 1336759, "ts": 1514656275628.295,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656275690.609, "dur": 42.120,
    "args": {
      "External id": 2479605,"kernel_hash": "cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 7308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656275744.075, "dur": 100.922,
    "args": {
      "External id": 2479606,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656275855.122, "dur": 36.975,
    "args": {
      "External id": 2479607,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656275900.488, "dur": 31.600,
    "args": {
      "External id": 2479608,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656275964.565, "dur": 60.796,
    "args": {
      "External id": 2479609,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 7312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656276046.496, "dur": 18.691,
    "args": {
      "External id": 2479610,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 7313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656276082.765, "dur": 127.481,
    "args": {
      "External id": 2479611,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 7314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656276133.102, "dur": 11.515,
    "args": {
      "External id": 2479612,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 7315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656276138.101, "dur": 5.827,
    "args": {
      "External id": 2479613,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656276147.232, "dur": 6.003,
    "args": {
      "External id": 2479614,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656276154.595, "dur": 1.301,
    "args": {
      "External id": 2479615,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656276158.486, "dur": 4.766,
    "args": {
      "External id": 2479616,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656276246.454, "dur": 51.053,
    "args": {
      "External id": 2479617,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 1336759, "tid": 1336759,
    "ts": 1514656276330.067, "dur": 26.823,
    "args": {
      "External id": 2479618,"kernel_hash": "c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/3l/c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 7321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656276365.971, "dur": 41.027,
    "args": {
      "External id": 2479619,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 7322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656276415.187, "dur": 34.798,
    "args": {
      "External id": 2479620,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 7323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656276471.318, "dur": 24.184,
    "args": {
      "External id": 2479621,"kernel_hash": "ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/kq/ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [16, 4096, 5632], []], "Ev Idx": 7324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656276500.654, "dur": 47.748,
    "args": {
      "External id": 2479622,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 7325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656276572.482, "dur": 19.029,
    "args": {
      "External id": 2479623,"kernel_hash": "cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u6/cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [65536, 2048], []], "Ev Idx": 7326
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.16)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656276729.656, "dur": 72.778,
    "args": {
      "External id": 2479624,"Record function id": 0, "Ev Idx": 7327
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656276875.800, "dur": 47.341,
    "args": {
      "External id": 2479625,"Record function id": 0, "Ev Idx": 7328
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.17)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656276932.294, "dur": 18365.086,
    "args": {
      "External id": 2479626,"Record function id": 0, "Ev Idx": 7329
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.17)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656276940.437, "dur": 899.644,
    "args": {
      "External id": 2479627,"Record function id": 0, "Ev Idx": 7330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277056.718, "dur": 8.945,
    "args": {
      "External id": 2479628,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277079.665, "dur": 36.190,
    "args": {
      "External id": 2479629,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 7332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277085.145, "dur": 2.319,
    "args": {
      "External id": 2479630,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277091.853, "dur": 0.298,
    "args": {
      "External id": 2479631,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277093.463, "dur": 0.453,
    "args": {
      "External id": 2479632,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277095.278, "dur": 0.389,
    "args": {
      "External id": 2479633,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277099.078, "dur": 0.186,
    "args": {
      "External id": 2479634,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277100.951, "dur": 0.380,
    "args": {
      "External id": 2479635,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277102.733, "dur": 3.085,
    "args": {
      "External id": 2479636,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277107.584, "dur": 0.284,
    "args": {
      "External id": 2479637,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277109.357, "dur": 0.502,
    "args": {
      "External id": 2479638,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277126.340, "dur": 39.266,
    "args": {
      "External id": 2479639,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 7342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277198.971, "dur": 105.397,
    "args": {
      "External id": 2479640,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 7343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277209.681, "dur": 4.087,
    "args": {
      "External id": 2479641,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277218.400, "dur": 9.827,
    "args": {
      "External id": 2479642,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 7345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277222.763, "dur": 5.076,
    "args": {
      "External id": 2479643,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 7346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277226.255, "dur": 0.492,
    "args": {
      "External id": 2479644,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 7347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277234.633, "dur": 29.135,
    "args": {
      "External id": 2479645,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 7348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277236.911, "dur": 2.453,
    "args": {
      "External id": 2479646,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277241.087, "dur": 0.344,
    "args": {
      "External id": 2479647,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277242.602, "dur": 0.339,
    "args": {
      "External id": 2479648,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277246.266, "dur": 1.328,
    "args": {
      "External id": 2479649,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277248.947, "dur": 0.277,
    "args": {
      "External id": 2479650,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277250.882, "dur": 0.393,
    "args": {
      "External id": 2479651,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277254.215, "dur": 0.414,
    "args": {
      "External id": 2479652,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277255.909, "dur": 0.275,
    "args": {
      "External id": 2479653,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277257.517, "dur": 1.933,
    "args": {
      "External id": 2479654,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277275.583, "dur": 21.412,
    "args": {
      "External id": 2479655,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 7358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277355.085, "dur": 388.691,
    "args": {
      "External id": 2479656,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 7359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277386.542, "dur": 352.050,
    "args": {
      "External id": 2479657,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 7360, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277397.040, "dur": 336.183,
    "args": {
      "External id": 2479658,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 7361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277768.514, "dur": 2.362,
    "args": {
      "External id": 2479659,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 7362, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.17)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277861.360, "dur": 17249.374,
    "args": {
      "External id": 2479660,"Record function id": 0, "Ev Idx": 7363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277958.046, "dur": 6.483,
    "args": {
      "External id": 2479661,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 7364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277967.738, "dur": 0.958,
    "args": {
      "External id": 2479662,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 7365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277970.428, "dur": 1.609,
    "args": {
      "External id": 2479663,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277973.584, "dur": 0.686,
    "args": {
      "External id": 2479664,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277975.713, "dur": 0.673,
    "args": {
      "External id": 2479665,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656277977.640, "dur": 0.753,
    "args": {
      "External id": 2479666,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656278018.370, "dur": 1.515,
    "args": {
      "External id": 2479667,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 7370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656278023.233, "dur": 2.164,
    "args": {
      "External id": 2479668,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656278026.921, "dur": 0.744,
    "args": {
      "External id": 2479669,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656278029.323, "dur": 0.591,
    "args": {
      "External id": 2479670,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656278051.745, "dur": 17020.639,
    "args": {
      "External id": 2479671,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 7374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656278067.846, "dur": 16997.912,
    "args": {
      "External id": 2479672,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 7375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656278094.869, "dur": 13.709,
    "args": {
      "External id": 2479673,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656278111.805, "dur": 16922.986,
    "args": {
      "External id": 2479674,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 7377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656278114.016, "dur": 16920.236,
    "args": {
      "External id": 2479675,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 7378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656278119.544, "dur": 5.595,
    "args": {
      "External id": 2479676,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656278126.715, "dur": 16904.657,
    "args": {
      "External id": 2479677,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 7380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656295241.357, "dur": 31.890,
    "args": {
      "External id": 2479678,"Sequence number": 24740455, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 7381
    }
  },
  {
    "ph": "s", "id": 193, "pid": 1336759, "tid": 1336759, "ts": 1514656295241.357,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514656295261.660, "dur": 7.191,
    "args": {
      "External id": 2479679,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 7382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656295264.643, "dur": 4.025,
    "args": {
      "External id": 2479680,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 7383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1336759, "tid": 1336759,
    "ts": 1514656295333.838, "dur": 73.648,
    "args": {
      "External id": 2479681,"Record function id": 0, "Ev Idx": 7384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656295408.983, "dur": 1038.443,
    "args": {
      "External id": 2479682,"Record function id": 0, "Ev Idx": 7385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656295447.871, "dur": 986.917,
    "args": {
      "External id": 2479683,"Sequence number": 24740456, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [32, 1], [32, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], "Input Dims": [[16, 4096, 2048], [2048], [2048, 2048], [2048, 2048], [2048, 2048], [8192, 32], [8192, 32], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], "Ev Idx": 7386
    }
  },
  {
    "ph": "s", "id": 192, "pid": 1336759, "tid": 1336759, "ts": 1514656295447.871,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656295512.867, "dur": 55.639,
    "args": {
      "External id": 2479684,"kernel_hash": "cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 7387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656295583.965, "dur": 103.374,
    "args": {
      "External id": 2479685,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656295697.117, "dur": 36.646,
    "args": {
      "External id": 2479686,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656295741.702, "dur": 30.959,
    "args": {
      "External id": 2479687,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656295797.114, "dur": 22.915,
    "args": {
      "External id": 2479688,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 7391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656295836.683, "dur": 15.544,
    "args": {
      "External id": 2479689,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 7392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656295869.691, "dur": 162.745,
    "args": {
      "External id": 2479690,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 7393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656295917.789, "dur": 11.386,
    "args": {
      "External id": 2479691,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 7394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656295922.822, "dur": 5.448,
    "args": {
      "External id": 2479692,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656295931.771, "dur": 5.240,
    "args": {
      "External id": 2479693,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656295938.330, "dur": 1.484,
    "args": {
      "External id": 2479694,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656295942.283, "dur": 3.637,
    "args": {
      "External id": 2479695,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656296044.804, "dur": 50.749,
    "args": {
      "External id": 2479696,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 1336759, "tid": 1336759,
    "ts": 1514656296126.942, "dur": 26.973,
    "args": {
      "External id": 2479697,"kernel_hash": "c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/3l/c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 7400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656296162.369, "dur": 41.065,
    "args": {
      "External id": 2479698,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 7401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656296211.673, "dur": 34.562,
    "args": {
      "External id": 2479699,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 7402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656296267.932, "dur": 24.820,
    "args": {
      "External id": 2479700,"kernel_hash": "ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/kq/ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [16, 4096, 5632], []], "Ev Idx": 7403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656296298.095, "dur": 32.895,
    "args": {
      "External id": 2479701,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 7404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656296349.331, "dur": 16.712,
    "args": {
      "External id": 2479702,"kernel_hash": "cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u6/cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [65536, 2048], []], "Ev Idx": 7405
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.17)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656296508.137, "dur": 86.113,
    "args": {
      "External id": 2479703,"Record function id": 0, "Ev Idx": 7406
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656296670.024, "dur": 46.825,
    "args": {
      "External id": 2479704,"Record function id": 0, "Ev Idx": 7407
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.18)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656296725.896, "dur": 18424.806,
    "args": {
      "External id": 2479705,"Record function id": 0, "Ev Idx": 7408
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.18)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656296733.647, "dur": 902.855,
    "args": {
      "External id": 2479706,"Record function id": 0, "Ev Idx": 7409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656296809.338, "dur": 8.246,
    "args": {
      "External id": 2479707,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656296830.227, "dur": 35.614,
    "args": {
      "External id": 2479708,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 7411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656296835.652, "dur": 2.271,
    "args": {
      "External id": 2479709,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656296841.886, "dur": 0.395,
    "args": {
      "External id": 2479710,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656296844.068, "dur": 0.586,
    "args": {
      "External id": 2479711,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656296845.907, "dur": 0.471,
    "args": {
      "External id": 2479712,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656296849.028, "dur": 0.455,
    "args": {
      "External id": 2479713,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656296851.154, "dur": 0.362,
    "args": {
      "External id": 2479714,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656296853.102, "dur": 3.159,
    "args": {
      "External id": 2479715,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656296857.503, "dur": 0.517,
    "args": {
      "External id": 2479716,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656296859.240, "dur": 0.400,
    "args": {
      "External id": 2479717,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656296876.804, "dur": 38.654,
    "args": {
      "External id": 2479718,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 7421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1336759,
    "ts": 1514656296947.252, "dur": 152.973,
    "args": {
      "External id": 2479719,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 7422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656296960.551, "dur": 5.310,
    "args": {
      "External id": 2479720,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1336759,
    "ts": 1514656296970.389, "dur": 9.773,
    "args": {
      "External id": 2479721,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 7424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656296974.560, "dur": 5.215,
    "args": {
      "External id": 2479722,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 7425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656296978.043, "dur": 0.449,
    "args": {
      "External id": 2479723,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 7426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297025.786, "dur": 28.940,
    "args": {
      "External id": 2479724,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 7427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297028.621, "dur": 0.790,
    "args": {
      "External id": 2479725,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297031.013, "dur": 2.517,
    "args": {
      "External id": 2479726,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297035.086, "dur": 0.233,
    "args": {
      "External id": 2479727,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297036.968, "dur": 1.683,
    "args": {
      "External id": 2479728,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297041.573, "dur": 0.241,
    "args": {
      "External id": 2479729,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297042.864, "dur": 0.262,
    "args": {
      "External id": 2479730,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297044.800, "dur": 0.526,
    "args": {
      "External id": 2479731,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297047.988, "dur": 0.423,
    "args": {
      "External id": 2479732,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297049.808, "dur": 0.292,
    "args": {
      "External id": 2479733,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297066.377, "dur": 25.388,
    "args": {
      "External id": 2479734,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 7437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297156.332, "dur": 366.473,
    "args": {
      "External id": 2479735,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 7438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297187.171, "dur": 330.774,
    "args": {
      "External id": 2479736,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 7439, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297197.400, "dur": 314.976,
    "args": {
      "External id": 2479737,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 7440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297562.676, "dur": 2.975,
    "args": {
      "External id": 2479738,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 7441, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.18)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297658.004, "dur": 17274.405,
    "args": {
      "External id": 2479739,"Record function id": 0, "Ev Idx": 7442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297754.529, "dur": 6.285,
    "args": {
      "External id": 2479740,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 7443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297764.322, "dur": 1.138,
    "args": {
      "External id": 2479741,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 7444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297767.101, "dur": 2.383,
    "args": {
      "External id": 2479742,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297771.273, "dur": 0.682,
    "args": {
      "External id": 2479743,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297773.190, "dur": 0.962,
    "args": {
      "External id": 2479744,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297775.346, "dur": 0.704,
    "args": {
      "External id": 2479745,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297778.816, "dur": 0.707,
    "args": {
      "External id": 2479746,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 7449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297780.835, "dur": 2.066,
    "args": {
      "External id": 2479747,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297784.188, "dur": 0.796,
    "args": {
      "External id": 2479748,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297786.412, "dur": 0.829,
    "args": {
      "External id": 2479749,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297806.468, "dur": 17088.330,
    "args": {
      "External id": 2479750,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 7453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297821.935, "dur": 17066.482,
    "args": {
      "External id": 2479751,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 7454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297843.816, "dur": 13.560,
    "args": {
      "External id": 2479752,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297860.334, "dur": 16997.550,
    "args": {
      "External id": 2479753,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 7456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297862.757, "dur": 16994.469,
    "args": {
      "External id": 2479754,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 7457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297868.330, "dur": 6.181,
    "args": {
      "External id": 2479755,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656297876.046, "dur": 16978.200,
    "args": {
      "External id": 2479756,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 7459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656315093.033, "dur": 32.753,
    "args": {
      "External id": 2479757,"Sequence number": 24740457, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 7460
    }
  },
  {
    "ph": "s", "id": 191, "pid": 1336759, "tid": 1336759, "ts": 1514656315093.033,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514656315113.361, "dur": 7.566,
    "args": {
      "External id": 2479758,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 7461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656315116.469, "dur": 4.097,
    "args": {
      "External id": 2479759,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 7462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1336759, "tid": 1336759,
    "ts": 1514656315187.601, "dur": 74.027,
    "args": {
      "External id": 2479760,"Record function id": 0, "Ev Idx": 7463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656315263.092, "dur": 1052.550,
    "args": {
      "External id": 2479761,"Record function id": 0, "Ev Idx": 7464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656315303.042, "dur": 999.778,
    "args": {
      "External id": 2479762,"Sequence number": 24740458, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [32, 1], [32, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], "Input Dims": [[16, 4096, 2048], [2048], [2048, 2048], [2048, 2048], [2048, 2048], [8192, 32], [8192, 32], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], "Ev Idx": 7465
    }
  },
  {
    "ph": "s", "id": 190, "pid": 1336759, "tid": 1336759, "ts": 1514656315303.042,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656315375.371, "dur": 40.116,
    "args": {
      "External id": 2479763,"kernel_hash": "cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 7466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656315428.954, "dur": 118.900,
    "args": {
      "External id": 2479764,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656315561.854, "dur": 41.856,
    "args": {
      "External id": 2479765,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656315610.288, "dur": 30.468,
    "args": {
      "External id": 2479766,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656315668.820, "dur": 24.945,
    "args": {
      "External id": 2479767,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 7470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656315709.049, "dur": 16.452,
    "args": {
      "External id": 2479768,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 7471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656315740.790, "dur": 123.090,
    "args": {
      "External id": 2479769,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 7472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656315787.889, "dur": 10.645,
    "args": {
      "External id": 2479770,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 7473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656315792.687, "dur": 5.081,
    "args": {
      "External id": 2479771,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656315801.247, "dur": 6.430,
    "args": {
      "External id": 2479772,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656315808.951, "dur": 2.708,
    "args": {
      "External id": 2479773,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656315814.211, "dur": 3.437,
    "args": {
      "External id": 2479774,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656315873.863, "dur": 43.230,
    "args": {
      "External id": 2479775,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 1336759, "tid": 1336759,
    "ts": 1514656315946.648, "dur": 25.044,
    "args": {
      "External id": 2479776,"kernel_hash": "c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/3l/c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 7479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656315980.311, "dur": 84.583,
    "args": {
      "External id": 2479777,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 7480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316075.320, "dur": 35.489,
    "args": {
      "External id": 2479778,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 7481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316136.479, "dur": 25.589,
    "args": {
      "External id": 2479779,"kernel_hash": "ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/kq/ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [16, 4096, 5632], []], "Ev Idx": 7482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316167.493, "dur": 33.548,
    "args": {
      "External id": 2479780,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 7483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316218.856, "dur": 19.595,
    "args": {
      "External id": 2479781,"kernel_hash": "cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u6/cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [65536, 2048], []], "Ev Idx": 7484
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.18)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316376.945, "dur": 70.034,
    "args": {
      "External id": 2479782,"Record function id": 0, "Ev Idx": 7485
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316519.257, "dur": 66.050,
    "args": {
      "External id": 2479783,"Record function id": 0, "Ev Idx": 7486
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.19)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316595.601, "dur": 18322.922,
    "args": {
      "External id": 2479784,"Record function id": 0, "Ev Idx": 7487
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.19)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316604.904, "dur": 942.285,
    "args": {
      "External id": 2479785,"Record function id": 0, "Ev Idx": 7488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316683.631, "dur": 9.206,
    "args": {
      "External id": 2479786,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316706.370, "dur": 33.172,
    "args": {
      "External id": 2479787,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 7490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316711.742, "dur": 2.236,
    "args": {
      "External id": 2479788,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316718.166, "dur": 0.374,
    "args": {
      "External id": 2479789,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316719.827, "dur": 0.441,
    "args": {
      "External id": 2479790,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316721.675, "dur": 0.326,
    "args": {
      "External id": 2479791,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316724.624, "dur": 0.610,
    "args": {
      "External id": 2479792,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316726.476, "dur": 0.291,
    "args": {
      "External id": 2479793,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316727.968, "dur": 2.630,
    "args": {
      "External id": 2479794,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316731.927, "dur": 0.184,
    "args": {
      "External id": 2479795,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316733.321, "dur": 0.280,
    "args": {
      "External id": 2479796,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316749.910, "dur": 40.293,
    "args": {
      "External id": 2479797,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 7500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316822.315, "dur": 104.633,
    "args": {
      "External id": 2479798,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 7501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316832.759, "dur": 3.567,
    "args": {
      "External id": 2479799,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316840.998, "dur": 9.729,
    "args": {
      "External id": 2479800,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 7503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316845.125, "dur": 5.183,
    "args": {
      "External id": 2479801,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 7504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316848.666, "dur": 0.585,
    "args": {
      "External id": 2479802,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 7505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316858.040, "dur": 27.747,
    "args": {
      "External id": 2479803,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 7506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316860.081, "dur": 2.262,
    "args": {
      "External id": 2479804,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316863.905, "dur": 0.459,
    "args": {
      "External id": 2479805,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316865.557, "dur": 0.323,
    "args": {
      "External id": 2479806,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316869.199, "dur": 1.321,
    "args": {
      "External id": 2479807,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316871.653, "dur": 0.537,
    "args": {
      "External id": 2479808,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316873.511, "dur": 0.280,
    "args": {
      "External id": 2479809,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316876.585, "dur": 0.277,
    "args": {
      "External id": 2479810,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316878.561, "dur": 0.306,
    "args": {
      "External id": 2479811,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316880.266, "dur": 1.780,
    "args": {
      "External id": 2479812,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316896.772, "dur": 22.598,
    "args": {
      "External id": 2479813,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 7516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656316977.630, "dur": 453.710,
    "args": {
      "External id": 2479814,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 7517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656317046.408, "dur": 379.599,
    "args": {
      "External id": 2479815,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 7518, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1336759,
    "ts": 1514656317057.856, "dur": 362.601,
    "args": {
      "External id": 2479816,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 7519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656317456.686, "dur": 2.658,
    "args": {
      "External id": 2479817,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 7520, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.19)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656317571.442, "dur": 17162.518,
    "args": {
      "External id": 2479818,"Record function id": 0, "Ev Idx": 7521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656317669.913, "dur": 6.629,
    "args": {
      "External id": 2479819,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 7522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656317691.217, "dur": 1.479,
    "args": {
      "External id": 2479820,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 7523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656317694.371, "dur": 2.579,
    "args": {
      "External id": 2479821,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656317699.788, "dur": 1.109,
    "args": {
      "External id": 2479822,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656317701.946, "dur": 0.579,
    "args": {
      "External id": 2479823,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656317703.730, "dur": 0.577,
    "args": {
      "External id": 2479824,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656317706.255, "dur": 0.878,
    "args": {
      "External id": 2479825,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 7528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656317710.567, "dur": 1.729,
    "args": {
      "External id": 2479826,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656317713.763, "dur": 1.013,
    "args": {
      "External id": 2479827,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656317716.366, "dur": 0.687,
    "args": {
      "External id": 2479828,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656317736.805, "dur": 16959.395,
    "args": {
      "External id": 2479829,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 7532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656317752.218, "dur": 16937.689,
    "args": {
      "External id": 2479830,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 7533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656317770.961, "dur": 12.835,
    "args": {
      "External id": 2479831,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656317788.565, "dur": 16870.157,
    "args": {
      "External id": 2479832,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 7535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656317791.158, "dur": 16866.994,
    "args": {
      "External id": 2479833,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 7536
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656317796.411, "dur": 5.577,
    "args": {
      "External id": 2479834,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656317803.621, "dur": 16851.888,
    "args": {
      "External id": 2479835,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 7538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656334862.138, "dur": 32.787,
    "args": {
      "External id": 2479836,"Sequence number": 24740459, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 7539
    }
  },
  {
    "ph": "s", "id": 189, "pid": 1336759, "tid": 1336759, "ts": 1514656334862.138,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514656334883.251, "dur": 7.188,
    "args": {
      "External id": 2479837,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 7540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656334886.078, "dur": 4.111,
    "args": {
      "External id": 2479838,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 7541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1336759, "tid": 1336759,
    "ts": 1514656334954.682, "dur": 101.991,
    "args": {
      "External id": 2479839,"Record function id": 0, "Ev Idx": 7542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656335059.646, "dur": 1047.438,
    "args": {
      "External id": 2479840,"Record function id": 0, "Ev Idx": 7543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656335100.300, "dur": 994.607,
    "args": {
      "External id": 2479841,"Sequence number": 24740460, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [32, 1], [32, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], "Input Dims": [[16, 4096, 2048], [2048], [2048, 2048], [2048, 2048], [2048, 2048], [8192, 32], [8192, 32], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], "Ev Idx": 7544
    }
  },
  {
    "ph": "s", "id": 188, "pid": 1336759, "tid": 1336759, "ts": 1514656335100.300,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656335166.097, "dur": 42.376,
    "args": {
      "External id": 2479842,"kernel_hash": "cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 7545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656335220.891, "dur": 103.525,
    "args": {
      "External id": 2479843,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656335333.967, "dur": 37.095,
    "args": {
      "External id": 2479844,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656335379.733, "dur": 32.383,
    "args": {
      "External id": 2479845,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656335434.428, "dur": 23.513,
    "args": {
      "External id": 2479846,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 7549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656335476.335, "dur": 14.126,
    "args": {
      "External id": 2479847,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 7550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656335507.355, "dur": 145.442,
    "args": {
      "External id": 2479848,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 7551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656335572.100, "dur": 11.864,
    "args": {
      "External id": 2479849,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 7552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656335577.096, "dur": 5.981,
    "args": {
      "External id": 2479850,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656335586.902, "dur": 5.833,
    "args": {
      "External id": 2479851,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656335594.029, "dur": 1.183,
    "args": {
      "External id": 2479852,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656335597.639, "dur": 4.336,
    "args": {
      "External id": 2479853,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656335664.299, "dur": 48.711,
    "args": {
      "External id": 2479854,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 1336759, "tid": 1336759,
    "ts": 1514656335743.681, "dur": 27.693,
    "args": {
      "External id": 2479855,"kernel_hash": "c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/3l/c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 7558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656335780.276, "dur": 40.671,
    "args": {
      "External id": 2479856,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 7559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656335829.629, "dur": 34.577,
    "args": {
      "External id": 2479857,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 7560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656335885.542, "dur": 25.020,
    "args": {
      "External id": 2479858,"kernel_hash": "ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/kq/ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [16, 4096, 5632], []], "Ev Idx": 7561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656335916.043, "dur": 33.950,
    "args": {
      "External id": 2479859,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 7562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656335970.443, "dur": 51.899,
    "args": {
      "External id": 2479860,"kernel_hash": "cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u6/cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [65536, 2048], []], "Ev Idx": 7563
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.19)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336169.049, "dur": 74.591,
    "args": {
      "External id": 2479861,"Record function id": 0, "Ev Idx": 7564
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336315.853, "dur": 46.322,
    "args": {
      "External id": 2479862,"Record function id": 0, "Ev Idx": 7565
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.20)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336371.308, "dur": 18262.960,
    "args": {
      "External id": 2479863,"Record function id": 0, "Ev Idx": 7566
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.20)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336378.967, "dur": 909.179,
    "args": {
      "External id": 2479864,"Record function id": 0, "Ev Idx": 7567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336456.092, "dur": 7.937,
    "args": {
      "External id": 2479865,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336476.401, "dur": 35.144,
    "args": {
      "External id": 2479866,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 7569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336481.532, "dur": 2.287,
    "args": {
      "External id": 2479867,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336488.401, "dur": 0.299,
    "args": {
      "External id": 2479868,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336490.600, "dur": 0.312,
    "args": {
      "External id": 2479869,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336492.499, "dur": 0.420,
    "args": {
      "External id": 2479870,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336495.693, "dur": 0.306,
    "args": {
      "External id": 2479871,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336497.545, "dur": 0.312,
    "args": {
      "External id": 2479872,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336499.134, "dur": 3.221,
    "args": {
      "External id": 2479873,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336503.623, "dur": 0.286,
    "args": {
      "External id": 2479874,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336505.290, "dur": 0.309,
    "args": {
      "External id": 2479875,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336521.726, "dur": 55.211,
    "args": {
      "External id": 2479876,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 7579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336612.775, "dur": 113.055,
    "args": {
      "External id": 2479877,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 7580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336623.619, "dur": 5.383,
    "args": {
      "External id": 2479878,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336634.142, "dur": 10.319,
    "args": {
      "External id": 2479879,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 7582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336638.370, "dur": 5.688,
    "args": {
      "External id": 2479880,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 7583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336642.064, "dur": 0.497,
    "args": {
      "External id": 2479881,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 7584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336651.721, "dur": 32.264,
    "args": {
      "External id": 2479882,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 7585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336654.031, "dur": 2.111,
    "args": {
      "External id": 2479883,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336657.492, "dur": 0.365,
    "args": {
      "External id": 2479884,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336662.303, "dur": 0.513,
    "args": {
      "External id": 2479885,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336666.197, "dur": 1.598,
    "args": {
      "External id": 2479886,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336668.986, "dur": 0.500,
    "args": {
      "External id": 2479887,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336670.627, "dur": 2.148,
    "args": {
      "External id": 2479888,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336674.152, "dur": 0.336,
    "args": {
      "External id": 2479889,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336675.671, "dur": 0.287,
    "args": {
      "External id": 2479890,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336679.201, "dur": 0.422,
    "args": {
      "External id": 2479891,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336695.157, "dur": 22.504,
    "args": {
      "External id": 2479892,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 7595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336780.347, "dur": 409.229,
    "args": {
      "External id": 2479893,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 7596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336810.709, "dur": 373.331,
    "args": {
      "External id": 2479894,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 7597, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1336759,
    "ts": 1514656336820.904, "dur": 357.896,
    "args": {
      "External id": 2479895,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 7598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656337214.327, "dur": 2.723,
    "args": {
      "External id": 2479896,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 7599, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.20)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656337310.261, "dur": 17120.371,
    "args": {
      "External id": 2479897,"Record function id": 0, "Ev Idx": 7600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656337405.498, "dur": 6.268,
    "args": {
      "External id": 2479898,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 7601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656337415.262, "dur": 1.451,
    "args": {
      "External id": 2479899,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 7602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656337418.233, "dur": 2.494,
    "args": {
      "External id": 2479900,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656337422.426, "dur": 0.891,
    "args": {
      "External id": 2479901,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656337424.457, "dur": 1.003,
    "args": {
      "External id": 2479902,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656337426.593, "dur": 0.895,
    "args": {
      "External id": 2479903,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656337430.860, "dur": 0.836,
    "args": {
      "External id": 2479904,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 7607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656337433.134, "dur": 1.985,
    "args": {
      "External id": 2479905,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656337436.597, "dur": 0.735,
    "args": {
      "External id": 2479906,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656337438.832, "dur": 0.793,
    "args": {
      "External id": 2479907,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656337459.104, "dur": 16933.817,
    "args": {
      "External id": 2479908,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 7611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656337474.934, "dur": 16911.489,
    "args": {
      "External id": 2479909,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 7612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656337496.981, "dur": 12.942,
    "args": {
      "External id": 2479910,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656337513.062, "dur": 16843.221,
    "args": {
      "External id": 2479911,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 7614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656337515.437, "dur": 16840.285,
    "args": {
      "External id": 2479912,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 7615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656337521.154, "dur": 21.166,
    "args": {
      "External id": 2479913,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656337545.252, "dur": 16807.242,
    "args": {
      "External id": 2479914,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 7617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656354573.499, "dur": 35.751,
    "args": {
      "External id": 2479915,"Sequence number": 24740461, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 7618
    }
  },
  {
    "ph": "s", "id": 187, "pid": 1336759, "tid": 1336759, "ts": 1514656354573.499,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514656354596.394, "dur": 8.016,
    "args": {
      "External id": 2479916,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 7619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656354599.425, "dur": 4.639,
    "args": {
      "External id": 2479917,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 7620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1336759, "tid": 1336759,
    "ts": 1514656354670.617, "dur": 70.345,
    "args": {
      "External id": 2479918,"Record function id": 0, "Ev Idx": 7621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656354742.570, "dur": 1022.573,
    "args": {
      "External id": 2479919,"Record function id": 0, "Ev Idx": 7622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656354780.449, "dur": 972.219,
    "args": {
      "External id": 2479920,"Sequence number": 24740462, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [32, 1], [32, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], "Input Dims": [[16, 4096, 2048], [2048], [2048, 2048], [2048, 2048], [2048, 2048], [8192, 32], [8192, 32], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], "Ev Idx": 7623
    }
  },
  {
    "ph": "s", "id": 186, "pid": 1336759, "tid": 1336759, "ts": 1514656354780.449,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656354844.645, "dur": 42.680,
    "args": {
      "External id": 2479921,"kernel_hash": "cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 7624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656354899.276, "dur": 124.058,
    "args": {
      "External id": 2479922,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656355039.235, "dur": 42.516,
    "args": {
      "External id": 2479923,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656355088.655, "dur": 30.484,
    "args": {
      "External id": 2479924,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656355146.684, "dur": 25.412,
    "args": {
      "External id": 2479925,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 7628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656355187.415, "dur": 13.654,
    "args": {
      "External id": 2479926,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 7629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656355218.365, "dur": 123.879,
    "args": {
      "External id": 2479927,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 7630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656355265.958, "dur": 10.652,
    "args": {
      "External id": 2479928,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 7631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656355270.618, "dur": 5.167,
    "args": {
      "External id": 2479929,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656355279.292, "dur": 5.362,
    "args": {
      "External id": 2479930,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656355285.871, "dur": 1.109,
    "args": {
      "External id": 2479931,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656355290.943, "dur": 3.651,
    "args": {
      "External id": 2479932,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656355352.844, "dur": 43.501,
    "args": {
      "External id": 2479933,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 1336759, "tid": 1336759,
    "ts": 1514656355425.039, "dur": 26.098,
    "args": {
      "External id": 2479934,"kernel_hash": "c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/3l/c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 7637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656355460.463, "dur": 41.007,
    "args": {
      "External id": 2479935,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 7638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656355507.764, "dur": 50.998,
    "args": {
      "External id": 2479936,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 7639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656355584.669, "dur": 26.508,
    "args": {
      "External id": 2479937,"kernel_hash": "ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/kq/ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [16, 4096, 5632], []], "Ev Idx": 7640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656355616.858, "dur": 36.426,
    "args": {
      "External id": 2479938,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 7641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656355670.279, "dur": 17.712,
    "args": {
      "External id": 2479939,"kernel_hash": "cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u6/cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [65536, 2048], []], "Ev Idx": 7642
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.20)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656355824.082, "dur": 69.507,
    "args": {
      "External id": 2479940,"Record function id": 0, "Ev Idx": 7643
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656355963.846, "dur": 83.396,
    "args": {
      "External id": 2479941,"Record function id": 0, "Ev Idx": 7644
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.21)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356057.987, "dur": 18223.212,
    "args": {
      "External id": 2479942,"Record function id": 0, "Ev Idx": 7645
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.21)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356067.422, "dur": 867.474,
    "args": {
      "External id": 2479943,"Record function id": 0, "Ev Idx": 7646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356143.975, "dur": 8.826,
    "args": {
      "External id": 2479944,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356166.023, "dur": 34.350,
    "args": {
      "External id": 2479945,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 7648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356171.352, "dur": 2.203,
    "args": {
      "External id": 2479946,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356177.510, "dur": 0.383,
    "args": {
      "External id": 2479947,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356179.087, "dur": 0.401,
    "args": {
      "External id": 2479948,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356181.080, "dur": 0.420,
    "args": {
      "External id": 2479949,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356184.118, "dur": 0.442,
    "args": {
      "External id": 2479950,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356185.879, "dur": 0.343,
    "args": {
      "External id": 2479951,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356187.730, "dur": 3.416,
    "args": {
      "External id": 2479952,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356192.584, "dur": 0.184,
    "args": {
      "External id": 2479953,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356193.919, "dur": 0.257,
    "args": {
      "External id": 2479954,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356210.667, "dur": 39.580,
    "args": {
      "External id": 2479955,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 7658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356282.375, "dur": 106.152,
    "args": {
      "External id": 2479956,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 7659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356293.107, "dur": 4.075,
    "args": {
      "External id": 2479957,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356301.758, "dur": 9.864,
    "args": {
      "External id": 2479958,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 7661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356305.894, "dur": 5.344,
    "args": {
      "External id": 2479959,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 7662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356309.637, "dur": 0.464,
    "args": {
      "External id": 2479960,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 7663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356318.984, "dur": 27.654,
    "args": {
      "External id": 2479961,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 7664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356321.338, "dur": 2.220,
    "args": {
      "External id": 2479962,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356324.940, "dur": 0.254,
    "args": {
      "External id": 2479963,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356326.396, "dur": 0.439,
    "args": {
      "External id": 2479964,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356330.109, "dur": 1.220,
    "args": {
      "External id": 2479965,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356332.576, "dur": 0.296,
    "args": {
      "External id": 2479966,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356334.302, "dur": 0.286,
    "args": {
      "External id": 2479967,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356337.466, "dur": 0.266,
    "args": {
      "External id": 2479968,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356339.158, "dur": 0.161,
    "args": {
      "External id": 2479969,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356340.363, "dur": 1.810,
    "args": {
      "External id": 2479970,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356358.316, "dur": 22.449,
    "args": {
      "External id": 2479971,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 7674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356448.958, "dur": 386.070,
    "args": {
      "External id": 2479972,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 7675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356482.061, "dur": 347.614,
    "args": {
      "External id": 2479973,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 7676, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356492.301, "dur": 331.397,
    "args": {
      "External id": 2479974,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 7677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356861.673, "dur": 2.477,
    "args": {
      "External id": 2479975,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 7678, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.21)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656356957.288, "dur": 17140.500,
    "args": {
      "External id": 2479976,"Record function id": 0, "Ev Idx": 7679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656357091.835, "dur": 6.752,
    "args": {
      "External id": 2479977,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 7680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656357102.198, "dur": 0.966,
    "args": {
      "External id": 2479978,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 7681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656357105.387, "dur": 2.199,
    "args": {
      "External id": 2479979,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656357109.341, "dur": 1.156,
    "args": {
      "External id": 2479980,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656357111.801, "dur": 0.866,
    "args": {
      "External id": 2479981,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656357113.957, "dur": 0.850,
    "args": {
      "External id": 2479982,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656357118.080, "dur": 0.951,
    "args": {
      "External id": 2479983,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 7686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656357120.504, "dur": 1.548,
    "args": {
      "External id": 2479984,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656357123.384, "dur": 0.847,
    "args": {
      "External id": 2479985,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656357125.782, "dur": 0.664,
    "args": {
      "External id": 2479986,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656357147.491, "dur": 16912.151,
    "args": {
      "External id": 2479987,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 7690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656357163.563, "dur": 16889.371,
    "args": {
      "External id": 2479988,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 7691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656357185.178, "dur": 14.239,
    "args": {
      "External id": 2479989,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656357202.342, "dur": 16819.630,
    "args": {
      "External id": 2479990,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 7693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656357204.655, "dur": 16816.729,
    "args": {
      "External id": 2479991,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 7694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656357210.296, "dur": 5.487,
    "args": {
      "External id": 2479992,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656357217.467, "dur": 16801.248,
    "args": {
      "External id": 2479993,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 7696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656374224.629, "dur": 32.857,
    "args": {
      "External id": 2479994,"Sequence number": 24740463, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 7697
    }
  },
  {
    "ph": "s", "id": 185, "pid": 1336759, "tid": 1336759, "ts": 1514656374224.629,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514656374244.608, "dur": 7.764,
    "args": {
      "External id": 2479995,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 7698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656374247.987, "dur": 4.160,
    "args": {
      "External id": 2479996,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 7699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1336759, "tid": 1336759,
    "ts": 1514656374316.644, "dur": 73.252,
    "args": {
      "External id": 2479997,"Record function id": 0, "Ev Idx": 7700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656374391.678, "dur": 1043.360,
    "args": {
      "External id": 2479998,"Record function id": 0, "Ev Idx": 7701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656374432.016, "dur": 990.800,
    "args": {
      "External id": 2479999,"Sequence number": 24740464, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [32, 1], [32, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], "Input Dims": [[16, 4096, 2048], [2048], [2048, 2048], [2048, 2048], [2048, 2048], [8192, 32], [8192, 32], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], "Ev Idx": 7702
    }
  },
  {
    "ph": "s", "id": 184, "pid": 1336759, "tid": 1336759, "ts": 1514656374432.016,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656374496.622, "dur": 55.010,
    "args": {
      "External id": 2480000,"kernel_hash": "cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 7703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656374567.319, "dur": 102.289,
    "args": {
      "External id": 2480001,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656374679.598, "dur": 37.693,
    "args": {
      "External id": 2480002,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656374726.251, "dur": 30.212,
    "args": {
      "External id": 2480003,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656374781.434, "dur": 24.810,
    "args": {
      "External id": 2480004,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 7707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656374825.246, "dur": 15.599,
    "args": {
      "External id": 2480005,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 7708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656374858.602, "dur": 121.395,
    "args": {
      "External id": 2480006,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 7709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656374905.188, "dur": 10.942,
    "args": {
      "External id": 2480007,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 7710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656374909.699, "dur": 5.440,
    "args": {
      "External id": 2480008,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656374918.783, "dur": 5.246,
    "args": {
      "External id": 2480009,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656374925.325, "dur": 1.063,
    "args": {
      "External id": 2480010,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656374929.054, "dur": 3.883,
    "args": {
      "External id": 2480011,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375030.076, "dur": 50.316,
    "args": {
      "External id": 2480012,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375110.713, "dur": 27.377,
    "args": {
      "External id": 2480013,"kernel_hash": "c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/3l/c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 7716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375147.315, "dur": 40.271,
    "args": {
      "External id": 2480014,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 7717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375195.484, "dur": 34.415,
    "args": {
      "External id": 2480015,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 7718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375253.144, "dur": 25.047,
    "args": {
      "External id": 2480016,"kernel_hash": "ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/kq/ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [16, 4096, 5632], []], "Ev Idx": 7719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375284.052, "dur": 33.598,
    "args": {
      "External id": 2480017,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 7720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375337.082, "dur": 18.020,
    "args": {
      "External id": 2480018,"kernel_hash": "cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u6/cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [65536, 2048], []], "Ev Idx": 7721
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.21)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375494.869, "dur": 87.936,
    "args": {
      "External id": 2480019,"Record function id": 0, "Ev Idx": 7722
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375658.441, "dur": 46.648,
    "args": {
      "External id": 2480020,"Record function id": 0, "Ev Idx": 7723
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.22)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375714.723, "dur": 18083.514,
    "args": {
      "External id": 2480021,"Record function id": 0, "Ev Idx": 7724
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.22)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375722.294, "dur": 879.018,
    "args": {
      "External id": 2480022,"Record function id": 0, "Ev Idx": 7725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375798.433, "dur": 8.578,
    "args": {
      "External id": 2480023,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375820.479, "dur": 35.594,
    "args": {
      "External id": 2480024,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 7727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375825.982, "dur": 2.193,
    "args": {
      "External id": 2480025,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375832.244, "dur": 0.376,
    "args": {
      "External id": 2480026,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375833.881, "dur": 0.472,
    "args": {
      "External id": 2480027,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375836.095, "dur": 0.243,
    "args": {
      "External id": 2480028,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375840.003, "dur": 0.360,
    "args": {
      "External id": 2480029,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375841.476, "dur": 0.337,
    "args": {
      "External id": 2480030,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375843.324, "dur": 3.114,
    "args": {
      "External id": 2480031,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375848.094, "dur": 0.310,
    "args": {
      "External id": 2480032,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375849.758, "dur": 0.404,
    "args": {
      "External id": 2480033,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375867.051, "dur": 39.220,
    "args": {
      "External id": 2480034,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 7737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375937.648, "dur": 154.664,
    "args": {
      "External id": 2480035,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 7738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375947.990, "dur": 4.111,
    "args": {
      "External id": 2480036,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375956.970, "dur": 9.954,
    "args": {
      "External id": 2480037,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 7740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375961.154, "dur": 5.342,
    "args": {
      "External id": 2480038,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 7741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375964.757, "dur": 0.655,
    "args": {
      "External id": 2480039,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 7742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375973.388, "dur": 70.118,
    "args": {
      "External id": 2480040,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 7743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375975.599, "dur": 2.077,
    "args": {
      "External id": 2480041,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656375979.263, "dur": 0.593,
    "args": {
      "External id": 2480042,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656376018.203, "dur": 0.815,
    "args": {
      "External id": 2480043,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656376023.386, "dur": 1.271,
    "args": {
      "External id": 2480044,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656376025.915, "dur": 0.457,
    "args": {
      "External id": 2480045,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656376027.667, "dur": 0.253,
    "args": {
      "External id": 2480046,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656376030.603, "dur": 0.374,
    "args": {
      "External id": 2480047,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656376032.353, "dur": 0.150,
    "args": {
      "External id": 2480048,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656376037.251, "dur": 1.795,
    "args": {
      "External id": 2480049,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656376058.287, "dur": 25.624,
    "args": {
      "External id": 2480050,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 7753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656376146.361, "dur": 346.331,
    "args": {
      "External id": 2480051,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 7754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656376177.426, "dur": 310.465,
    "args": {
      "External id": 2480052,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 7755, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1336759,
    "ts": 1514656376187.440, "dur": 293.387,
    "args": {
      "External id": 2480053,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 7756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656376513.894, "dur": 2.199,
    "args": {
      "External id": 2480054,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 7757, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.22)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656376623.227, "dur": 16990.916,
    "args": {
      "External id": 2480055,"Record function id": 0, "Ev Idx": 7758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656376716.016, "dur": 6.203,
    "args": {
      "External id": 2480056,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 7759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656376725.487, "dur": 1.527,
    "args": {
      "External id": 2480057,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 7760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656376728.711, "dur": 2.389,
    "args": {
      "External id": 2480058,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656376732.794, "dur": 0.880,
    "args": {
      "External id": 2480059,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656376734.950, "dur": 0.671,
    "args": {
      "External id": 2480060,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656376736.819, "dur": 0.985,
    "args": {
      "External id": 2480061,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656376740.949, "dur": 1.013,
    "args": {
      "External id": 2480062,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 7765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656376743.394, "dur": 1.664,
    "args": {
      "External id": 2480063,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656376746.366, "dur": 0.643,
    "args": {
      "External id": 2480064,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656376748.494, "dur": 0.744,
    "args": {
      "External id": 2480065,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656376767.739, "dur": 16809.529,
    "args": {
      "External id": 2480066,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 7769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656376787.842, "dur": 16782.762,
    "args": {
      "External id": 2480067,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 7770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656376809.890, "dur": 13.703,
    "args": {
      "External id": 2480068,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656376826.568, "dur": 16712.828,
    "args": {
      "External id": 2480069,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 7772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656376829.102, "dur": 16709.629,
    "args": {
      "External id": 2480070,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 7773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656376834.534, "dur": 5.721,
    "args": {
      "External id": 2480071,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656376841.917, "dur": 16683.719,
    "args": {
      "External id": 2480072,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 7775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656393742.988, "dur": 31.816,
    "args": {
      "External id": 2480073,"Sequence number": 24740465, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 7776
    }
  },
  {
    "ph": "s", "id": 183, "pid": 1336759, "tid": 1336759, "ts": 1514656393742.988,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514656393762.703, "dur": 7.098,
    "args": {
      "External id": 2480074,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 7777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656393765.592, "dur": 3.985,
    "args": {
      "External id": 2480075,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 7778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1336759, "tid": 1336759,
    "ts": 1514656393833.937, "dur": 72.114,
    "args": {
      "External id": 2480076,"Record function id": 0, "Ev Idx": 7779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656393907.444, "dur": 1045.415,
    "args": {
      "External id": 2480077,"Record function id": 0, "Ev Idx": 7780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656393946.092, "dur": 994.401,
    "args": {
      "External id": 2480078,"Sequence number": 24740466, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [32, 1], [32, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], "Input Dims": [[16, 4096, 2048], [2048], [2048, 2048], [2048, 2048], [2048, 2048], [8192, 32], [8192, 32], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], "Ev Idx": 7781
    }
  },
  {
    "ph": "s", "id": 182, "pid": 1336759, "tid": 1336759, "ts": 1514656393946.092,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656394053.209, "dur": 43.457,
    "args": {
      "External id": 2480079,"kernel_hash": "cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 7782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656394110.290, "dur": 102.317,
    "args": {
      "External id": 2480080,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656394222.171, "dur": 37.699,
    "args": {
      "External id": 2480081,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656394267.833, "dur": 31.093,
    "args": {
      "External id": 2480082,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656394323.906, "dur": 25.107,
    "args": {
      "External id": 2480083,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 7786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656394366.462, "dur": 14.519,
    "args": {
      "External id": 2480084,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 7787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656394398.296, "dur": 122.529,
    "args": {
      "External id": 2480085,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 7788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656394445.413, "dur": 10.690,
    "args": {
      "External id": 2480086,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 7789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656394450.074, "dur": 5.364,
    "args": {
      "External id": 2480087,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656394458.864, "dur": 5.218,
    "args": {
      "External id": 2480088,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656394465.469, "dur": 1.319,
    "args": {
      "External id": 2480089,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656394469.223, "dur": 4.063,
    "args": {
      "External id": 2480090,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656394548.213, "dur": 49.777,
    "args": {
      "External id": 2480091,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 1336759, "tid": 1336759,
    "ts": 1514656394628.297, "dur": 27.091,
    "args": {
      "External id": 2480092,"kernel_hash": "c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/3l/c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 7795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656394664.580, "dur": 41.890,
    "args": {
      "External id": 2480093,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 7796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656394714.759, "dur": 34.271,
    "args": {
      "External id": 2480094,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 7797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656394772.114, "dur": 27.291,
    "args": {
      "External id": 2480095,"kernel_hash": "ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/kq/ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [16, 4096, 5632], []], "Ev Idx": 7798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656394805.504, "dur": 34.134,
    "args": {
      "External id": 2480096,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 7799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656394858.543, "dur": 17.166,
    "args": {
      "External id": 2480097,"kernel_hash": "cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u6/cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [65536, 2048], []], "Ev Idx": 7800
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.22)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395053.339, "dur": 71.923,
    "args": {
      "External id": 2480098,"Record function id": 0, "Ev Idx": 7801
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395199.580, "dur": 46.888,
    "args": {
      "External id": 2480099,"Record function id": 0, "Ev Idx": 7802
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.23)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395256.025, "dur": 18335.993,
    "args": {
      "External id": 2480100,"Record function id": 0, "Ev Idx": 7803
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.23)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395265.139, "dur": 928.549,
    "args": {
      "External id": 2480101,"Record function id": 0, "Ev Idx": 7804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395345.642, "dur": 9.088,
    "args": {
      "External id": 2480102,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395367.341, "dur": 32.639,
    "args": {
      "External id": 2480103,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 7806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395372.786, "dur": 2.113,
    "args": {
      "External id": 2480104,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395378.722, "dur": 0.316,
    "args": {
      "External id": 2480105,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395380.441, "dur": 0.222,
    "args": {
      "External id": 2480106,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395382.379, "dur": 0.617,
    "args": {
      "External id": 2480107,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395385.504, "dur": 0.315,
    "args": {
      "External id": 2480108,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395387.337, "dur": 0.198,
    "args": {
      "External id": 2480109,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395388.637, "dur": 3.044,
    "args": {
      "External id": 2480110,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395392.925, "dur": 0.340,
    "args": {
      "External id": 2480111,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395394.543, "dur": 0.309,
    "args": {
      "External id": 2480112,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395411.097, "dur": 38.538,
    "args": {
      "External id": 2480113,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 7816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395482.133, "dur": 160.136,
    "args": {
      "External id": 2480114,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 7817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395492.450, "dur": 4.366,
    "args": {
      "External id": 2480115,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395501.493, "dur": 10.263,
    "args": {
      "External id": 2480116,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 7819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395505.653, "dur": 5.696,
    "args": {
      "External id": 2480117,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 7820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395509.705, "dur": 0.505,
    "args": {
      "External id": 2480118,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 7821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395518.412, "dur": 47.643,
    "args": {
      "External id": 2480119,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 7822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395520.903, "dur": 2.182,
    "args": {
      "External id": 2480120,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395524.597, "dur": 0.368,
    "args": {
      "External id": 2480121,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395541.770, "dur": 0.804,
    "args": {
      "External id": 2480122,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395547.162, "dur": 1.448,
    "args": {
      "External id": 2480123,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395549.789, "dur": 0.700,
    "args": {
      "External id": 2480124,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395551.803, "dur": 0.404,
    "args": {
      "External id": 2480125,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395555.065, "dur": 0.617,
    "args": {
      "External id": 2480126,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395557.010, "dur": 0.344,
    "args": {
      "External id": 2480127,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395559.216, "dur": 2.138,
    "args": {
      "External id": 2480128,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395607.220, "dur": 26.839,
    "args": {
      "External id": 2480129,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 7832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395696.772, "dur": 401.112,
    "args": {
      "External id": 2480130,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 7833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395726.723, "dur": 365.963,
    "args": {
      "External id": 2480131,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 7834, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1336759,
    "ts": 1514656395738.934, "dur": 347.714,
    "args": {
      "External id": 2480132,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 7835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656396121.540, "dur": 2.577,
    "args": {
      "External id": 2480133,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 7836, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.23)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656396214.240, "dur": 17178.082,
    "args": {
      "External id": 2480134,"Record function id": 0, "Ev Idx": 7837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656396307.770, "dur": 6.372,
    "args": {
      "External id": 2480135,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 7838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656396317.602, "dur": 0.854,
    "args": {
      "External id": 2480136,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 7839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656396320.147, "dur": 1.904,
    "args": {
      "External id": 2480137,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656396323.585, "dur": 0.588,
    "args": {
      "External id": 2480138,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656396325.316, "dur": 0.562,
    "args": {
      "External id": 2480139,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656396327.009, "dur": 0.822,
    "args": {
      "External id": 2480140,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656396331.231, "dur": 0.738,
    "args": {
      "External id": 2480141,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 7844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656396333.340, "dur": 1.664,
    "args": {
      "External id": 2480142,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656396336.237, "dur": 0.617,
    "args": {
      "External id": 2480143,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656396338.162, "dur": 0.667,
    "args": {
      "External id": 2480144,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656396357.514, "dur": 16997.583,
    "args": {
      "External id": 2480145,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 7848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656396372.773, "dur": 16975.908,
    "args": {
      "External id": 2480146,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 7849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656396393.393, "dur": 12.999,
    "args": {
      "External id": 2480147,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656396409.382, "dur": 16909.800,
    "args": {
      "External id": 2480148,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 7851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656396411.890, "dur": 16906.702,
    "args": {
      "External id": 2480149,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 7852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656396417.539, "dur": 5.184,
    "args": {
      "External id": 2480150,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656396424.288, "dur": 16891.419,
    "args": {
      "External id": 2480151,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 7854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656413519.032, "dur": 46.808,
    "args": {
      "External id": 2480152,"Sequence number": 24740467, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 7855
    }
  },
  {
    "ph": "s", "id": 181, "pid": 1336759, "tid": 1336759, "ts": 1514656413519.032,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514656413552.935, "dur": 7.813,
    "args": {
      "External id": 2480153,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 7856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656413555.962, "dur": 4.464,
    "args": {
      "External id": 2480154,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 7857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1336759, "tid": 1336759,
    "ts": 1514656413628.687, "dur": 72.476,
    "args": {
      "External id": 2480155,"Record function id": 0, "Ev Idx": 7858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656413702.794, "dur": 1047.665,
    "args": {
      "External id": 2480156,"Record function id": 0, "Ev Idx": 7859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656413745.507, "dur": 992.294,
    "args": {
      "External id": 2480157,"Sequence number": 24740468, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [32, 1], [32, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], "Input Dims": [[16, 4096, 2048], [2048], [2048, 2048], [2048, 2048], [2048, 2048], [8192, 32], [8192, 32], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], "Ev Idx": 7860
    }
  },
  {
    "ph": "s", "id": 180, "pid": 1336759, "tid": 1336759, "ts": 1514656413745.507,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656413809.187, "dur": 42.610,
    "args": {
      "External id": 2480158,"kernel_hash": "cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 7861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656413863.771, "dur": 101.332,
    "args": {
      "External id": 2480159,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656413974.761, "dur": 78.143,
    "args": {
      "External id": 2480160,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656414065.440, "dur": 32.473,
    "args": {
      "External id": 2480161,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656414123.475, "dur": 26.686,
    "args": {
      "External id": 2480162,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 7865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656414166.440, "dur": 15.542,
    "args": {
      "External id": 2480163,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 7866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656414199.661, "dur": 122.294,
    "args": {
      "External id": 2480164,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 7867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656414247.279, "dur": 11.113,
    "args": {
      "External id": 2480165,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 7868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656414252.112, "dur": 5.531,
    "args": {
      "External id": 2480166,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656414260.885, "dur": 5.076,
    "args": {
      "External id": 2480167,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656414267.252, "dur": 1.183,
    "args": {
      "External id": 2480168,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656414270.784, "dur": 3.920,
    "args": {
      "External id": 2480169,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656414332.745, "dur": 42.873,
    "args": {
      "External id": 2480170,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 1336759, "tid": 1336759,
    "ts": 1514656414404.002, "dur": 27.114,
    "args": {
      "External id": 2480171,"kernel_hash": "c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/3l/c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 7874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656414439.270, "dur": 40.669,
    "args": {
      "External id": 2480172,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 7875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656414487.834, "dur": 34.452,
    "args": {
      "External id": 2480173,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 7876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656414560.169, "dur": 26.834,
    "args": {
      "External id": 2480174,"kernel_hash": "ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/kq/ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [16, 4096, 5632], []], "Ev Idx": 7877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656414593.559, "dur": 37.695,
    "args": {
      "External id": 2480175,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 7878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656414651.106, "dur": 16.527,
    "args": {
      "External id": 2480176,"kernel_hash": "cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u6/cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [65536, 2048], []], "Ev Idx": 7879
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.23)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656414811.208, "dur": 67.260,
    "args": {
      "External id": 2480177,"Record function id": 0, "Ev Idx": 7880
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656414947.041, "dur": 82.966,
    "args": {
      "External id": 2480178,"Record function id": 0, "Ev Idx": 7881
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.24)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415041.040, "dur": 18346.429,
    "args": {
      "External id": 2480179,"Record function id": 0, "Ev Idx": 7882
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.24)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415049.585, "dur": 836.501,
    "args": {
      "External id": 2480180,"Record function id": 0, "Ev Idx": 7883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415128.855, "dur": 8.620,
    "args": {
      "External id": 2480181,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415150.853, "dur": 31.122,
    "args": {
      "External id": 2480182,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 7885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415156.366, "dur": 1.942,
    "args": {
      "External id": 2480183,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415162.468, "dur": 0.303,
    "args": {
      "External id": 2480184,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415163.808, "dur": 0.545,
    "args": {
      "External id": 2480185,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415165.301, "dur": 0.333,
    "args": {
      "External id": 2480186,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415168.297, "dur": 0.338,
    "args": {
      "External id": 2480187,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415169.620, "dur": 0.520,
    "args": {
      "External id": 2480188,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415171.037, "dur": 2.711,
    "args": {
      "External id": 2480189,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415174.717, "dur": 0.249,
    "args": {
      "External id": 2480190,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415175.981, "dur": 0.193,
    "args": {
      "External id": 2480191,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415192.769, "dur": 38.302,
    "args": {
      "External id": 2480192,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 7895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415263.085, "dur": 99.646,
    "args": {
      "External id": 2480193,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 7896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415273.471, "dur": 4.466,
    "args": {
      "External id": 2480194,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415282.843, "dur": 9.680,
    "args": {
      "External id": 2480195,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 7898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415287.227, "dur": 4.881,
    "args": {
      "External id": 2480196,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 7899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415290.462, "dur": 0.468,
    "args": {
      "External id": 2480197,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 7900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415299.082, "dur": 23.603,
    "args": {
      "External id": 2480198,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 7901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415300.825, "dur": 1.764,
    "args": {
      "External id": 2480199,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415303.630, "dur": 0.347,
    "args": {
      "External id": 2480200,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415305.250, "dur": 0.403,
    "args": {
      "External id": 2480201,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415308.092, "dur": 1.040,
    "args": {
      "External id": 2480202,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415310.282, "dur": 0.181,
    "args": {
      "External id": 2480203,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415311.452, "dur": 0.313,
    "args": {
      "External id": 2480204,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415314.461, "dur": 0.299,
    "args": {
      "External id": 2480205,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415315.696, "dur": 0.332,
    "args": {
      "External id": 2480206,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415316.817, "dur": 1.856,
    "args": {
      "External id": 2480207,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415332.796, "dur": 22.033,
    "args": {
      "External id": 2480208,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 7911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415414.252, "dur": 377.065,
    "args": {
      "External id": 2480209,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 7912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415449.721, "dur": 336.276,
    "args": {
      "External id": 2480210,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 7913, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415459.687, "dur": 320.139,
    "args": {
      "External id": 2480211,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 7914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415814.366, "dur": 2.427,
    "args": {
      "External id": 2480212,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 7915, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.24)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656415907.519, "dur": 17294.583,
    "args": {
      "External id": 2480213,"Record function id": 0, "Ev Idx": 7916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656416041.365, "dur": 7.018,
    "args": {
      "External id": 2480214,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 7917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656416056.996, "dur": 1.087,
    "args": {
      "External id": 2480215,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 7918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656416059.750, "dur": 1.825,
    "args": {
      "External id": 2480216,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656416064.533, "dur": 0.847,
    "args": {
      "External id": 2480217,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656416066.470, "dur": 0.895,
    "args": {
      "External id": 2480218,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656416068.491, "dur": 0.784,
    "args": {
      "External id": 2480219,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656416073.603, "dur": 0.774,
    "args": {
      "External id": 2480220,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 7923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656416077.387, "dur": 1.477,
    "args": {
      "External id": 2480221,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656416080.398, "dur": 0.693,
    "args": {
      "External id": 2480222,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656416082.576, "dur": 0.636,
    "args": {
      "External id": 2480223,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 7926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656416104.429, "dur": 17058.244,
    "args": {
      "External id": 2480224,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 7927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656416119.870, "dur": 17036.431,
    "args": {
      "External id": 2480225,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 7928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656416138.399, "dur": 13.623,
    "args": {
      "External id": 2480226,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656416157.917, "dur": 16968.167,
    "args": {
      "External id": 2480227,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 7930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656416160.626, "dur": 16964.940,
    "args": {
      "External id": 2480228,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 7931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656416166.037, "dur": 5.060,
    "args": {
      "External id": 2480229,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656416172.752, "dur": 16949.979,
    "args": {
      "External id": 2480230,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 7933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656433330.808, "dur": 32.575,
    "args": {
      "External id": 2480231,"Sequence number": 24740469, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 7934
    }
  },
  {
    "ph": "s", "id": 179, "pid": 1336759, "tid": 1336759, "ts": 1514656433330.808,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514656433351.535, "dur": 6.938,
    "args": {
      "External id": 2480232,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 7935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656433354.346, "dur": 3.918,
    "args": {
      "External id": 2480233,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 7936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1336759, "tid": 1336759,
    "ts": 1514656433424.300, "dur": 73.152,
    "args": {
      "External id": 2480234,"Record function id": 0, "Ev Idx": 7937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656433499.169, "dur": 1025.382,
    "args": {
      "External id": 2480235,"Record function id": 0, "Ev Idx": 7938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656433551.879, "dur": 959.918,
    "args": {
      "External id": 2480236,"Sequence number": 24740470, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [32, 1], [32, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], "Input Dims": [[16, 4096, 2048], [2048], [2048, 2048], [2048, 2048], [2048, 2048], [8192, 32], [8192, 32], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], "Ev Idx": 7939
    }
  },
  {
    "ph": "s", "id": 178, "pid": 1336759, "tid": 1336759, "ts": 1514656433551.879,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656433618.757, "dur": 41.459,
    "args": {
      "External id": 2480237,"kernel_hash": "cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 7940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656433672.804, "dur": 98.685,
    "args": {
      "External id": 2480238,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656433781.207, "dur": 36.822,
    "args": {
      "External id": 2480239,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656433826.088, "dur": 31.530,
    "args": {
      "External id": 2480240,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656433879.434, "dur": 24.303,
    "args": {
      "External id": 2480241,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 7944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656433921.489, "dur": 14.595,
    "args": {
      "External id": 2480242,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 7945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656433953.582, "dur": 160.153,
    "args": {
      "External id": 2480243,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 7946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656434036.543, "dur": 12.186,
    "args": {
      "External id": 2480244,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 7947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656434041.538, "dur": 6.230,
    "args": {
      "External id": 2480245,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656434051.241, "dur": 5.673,
    "args": {
      "External id": 2480246,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656434058.098, "dur": 1.046,
    "args": {
      "External id": 2480247,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656434061.555, "dur": 2.969,
    "args": {
      "External id": 2480248,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656434124.300, "dur": 48.836,
    "args": {
      "External id": 2480249,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 7952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 1336759, "tid": 1336759,
    "ts": 1514656434202.819, "dur": 26.555,
    "args": {
      "External id": 2480250,"kernel_hash": "c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/3l/c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 7953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656434238.131, "dur": 40.122,
    "args": {
      "External id": 2480251,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 7954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656434286.120, "dur": 34.199,
    "args": {
      "External id": 2480252,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 7955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656434348.699, "dur": 24.592,
    "args": {
      "External id": 2480253,"kernel_hash": "ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/kq/ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [16, 4096, 5632], []], "Ev Idx": 7956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656434380.632, "dur": 34.565,
    "args": {
      "External id": 2480254,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 7957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656434433.062, "dur": 17.020,
    "args": {
      "External id": 2480255,"kernel_hash": "cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u6/cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [65536, 2048], []], "Ev Idx": 7958
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.24)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656434602.020, "dur": 69.002,
    "args": {
      "External id": 2480256,"Record function id": 0, "Ev Idx": 7959
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656434743.943, "dur": 45.747,
    "args": {
      "External id": 2480257,"Record function id": 0, "Ev Idx": 7960
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.25)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656434799.948, "dur": 18450.685,
    "args": {
      "External id": 2480258,"Record function id": 0, "Ev Idx": 7961
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.25)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656434808.561, "dur": 945.987,
    "args": {
      "External id": 2480259,"Record function id": 0, "Ev Idx": 7962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656434884.167, "dur": 8.878,
    "args": {
      "External id": 2480260,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656434906.340, "dur": 31.275,
    "args": {
      "External id": 2480261,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 7964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656434911.293, "dur": 2.378,
    "args": {
      "External id": 2480262,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656434917.387, "dur": 0.222,
    "args": {
      "External id": 2480263,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656434918.626, "dur": 0.220,
    "args": {
      "External id": 2480264,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656434919.886, "dur": 0.466,
    "args": {
      "External id": 2480265,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656434923.232, "dur": 0.343,
    "args": {
      "External id": 2480266,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656434924.701, "dur": 0.328,
    "args": {
      "External id": 2480267,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656434925.803, "dur": 3.116,
    "args": {
      "External id": 2480268,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656434930.038, "dur": 0.263,
    "args": {
      "External id": 2480269,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656434931.553, "dur": 0.499,
    "args": {
      "External id": 2480270,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656434947.693, "dur": 74.871,
    "args": {
      "External id": 2480271,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 7974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435060.396, "dur": 110.477,
    "args": {
      "External id": 2480272,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 7975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435071.752, "dur": 5.168,
    "args": {
      "External id": 2480273,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 7976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435081.637, "dur": 10.544,
    "args": {
      "External id": 2480274,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 7977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435086.144, "dur": 5.648,
    "args": {
      "External id": 2480275,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 7978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435089.609, "dur": 0.723,
    "args": {
      "External id": 2480276,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 7979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435099.484, "dur": 27.058,
    "args": {
      "External id": 2480277,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 7980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435101.859, "dur": 2.246,
    "args": {
      "External id": 2480278,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435105.521, "dur": 0.151,
    "args": {
      "External id": 2480279,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435107.348, "dur": 0.282,
    "args": {
      "External id": 2480280,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435110.616, "dur": 1.135,
    "args": {
      "External id": 2480281,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435112.847, "dur": 0.256,
    "args": {
      "External id": 2480282,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435114.461, "dur": 0.303,
    "args": {
      "External id": 2480283,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435117.883, "dur": 0.355,
    "args": {
      "External id": 2480284,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435119.434, "dur": 0.447,
    "args": {
      "External id": 2480285,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435120.816, "dur": 1.730,
    "args": {
      "External id": 2480286,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 7989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435138.703, "dur": 24.468,
    "args": {
      "External id": 2480287,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 7990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435223.602, "dur": 434.883,
    "args": {
      "External id": 2480288,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 7991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435253.176, "dur": 400.087,
    "args": {
      "External id": 2480289,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 7992, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435263.169, "dur": 384.346,
    "args": {
      "External id": 2480290,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 7993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435681.944, "dur": 2.562,
    "args": {
      "External id": 2480291,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 7994, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.25)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435774.684, "dur": 17290.284,
    "args": {
      "External id": 2480292,"Record function id": 0, "Ev Idx": 7995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435868.979, "dur": 6.530,
    "args": {
      "External id": 2480293,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 7996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435878.551, "dur": 1.254,
    "args": {
      "External id": 2480294,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 7997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435881.339, "dur": 1.895,
    "args": {
      "External id": 2480295,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435884.817, "dur": 0.914,
    "args": {
      "External id": 2480296,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 7999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435886.930, "dur": 1.065,
    "args": {
      "External id": 2480297,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 8000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435889.246, "dur": 0.684,
    "args": {
      "External id": 2480298,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 8001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435893.437, "dur": 1.225,
    "args": {
      "External id": 2480299,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 8002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435896.071, "dur": 1.493,
    "args": {
      "External id": 2480300,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 8003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435898.843, "dur": 0.821,
    "args": {
      "External id": 2480301,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 8004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435901.174, "dur": 0.393,
    "args": {
      "External id": 2480302,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 8005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435920.789, "dur": 17106.774,
    "args": {
      "External id": 2480303,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 8006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435935.587, "dur": 17085.082,
    "args": {
      "External id": 2480304,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 8007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435959.200, "dur": 13.328,
    "args": {
      "External id": 2480305,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435975.435, "dur": 16990.925,
    "args": {
      "External id": 2480306,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 8009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656435977.620, "dur": 16988.167,
    "args": {
      "External id": 2480307,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 8010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656436024.662, "dur": 6.446,
    "args": {
      "External id": 2480308,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656436033.074, "dur": 16929.905,
    "args": {
      "External id": 2480309,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 8012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656453195.571, "dur": 31.561,
    "args": {
      "External id": 2480310,"Sequence number": 24740471, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 8013
    }
  },
  {
    "ph": "s", "id": 177, "pid": 1336759, "tid": 1336759, "ts": 1514656453195.571,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514656453215.007, "dur": 7.391,
    "args": {
      "External id": 2480311,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 8014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656453217.707, "dur": 4.487,
    "args": {
      "External id": 2480312,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1336759, "tid": 1336759,
    "ts": 1514656453286.613, "dur": 72.580,
    "args": {
      "External id": 2480313,"Record function id": 0, "Ev Idx": 8016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656453360.581, "dur": 1042.929,
    "args": {
      "External id": 2480314,"Record function id": 0, "Ev Idx": 8017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656453401.187, "dur": 989.283,
    "args": {
      "External id": 2480315,"Sequence number": 24740472, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [32, 1], [32, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], "Input Dims": [[16, 4096, 2048], [2048], [2048, 2048], [2048, 2048], [2048, 2048], [8192, 32], [8192, 32], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], "Ev Idx": 8018
    }
  },
  {
    "ph": "s", "id": 176, "pid": 1336759, "tid": 1336759, "ts": 1514656453401.187,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656453463.471, "dur": 39.192,
    "args": {
      "External id": 2480316,"kernel_hash": "cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 8019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656453514.345, "dur": 119.620,
    "args": {
      "External id": 2480317,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 8020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656453646.910, "dur": 38.668,
    "args": {
      "External id": 2480318,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 8021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656453694.195, "dur": 31.002,
    "args": {
      "External id": 2480319,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 8022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656453750.562, "dur": 25.431,
    "args": {
      "External id": 2480320,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 8023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656453792.660, "dur": 14.915,
    "args": {
      "External id": 2480321,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 8024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656453824.584, "dur": 118.585,
    "args": {
      "External id": 2480322,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 8025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656453870.575, "dur": 10.288,
    "args": {
      "External id": 2480323,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 8026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656453875.108, "dur": 5.039,
    "args": {
      "External id": 2480324,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656453883.356, "dur": 5.200,
    "args": {
      "External id": 2480325,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656453889.727, "dur": 1.195,
    "args": {
      "External id": 2480326,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656453893.325, "dur": 3.333,
    "args": {
      "External id": 2480327,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656453953.200, "dur": 84.287,
    "args": {
      "External id": 2480328,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 8031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454071.672, "dur": 28.561,
    "args": {
      "External id": 2480329,"kernel_hash": "c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/3l/c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 8032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454108.819, "dur": 44.943,
    "args": {
      "External id": 2480330,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 8033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454162.316, "dur": 34.231,
    "args": {
      "External id": 2480331,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 8034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454219.074, "dur": 28.492,
    "args": {
      "External id": 2480332,"kernel_hash": "ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/kq/ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [16, 4096, 5632], []], "Ev Idx": 8035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454253.177, "dur": 34.057,
    "args": {
      "External id": 2480333,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 8036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454306.442, "dur": 16.600,
    "args": {
      "External id": 2480334,"kernel_hash": "cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u6/cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [65536, 2048], []], "Ev Idx": 8037
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.25)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454466.246, "dur": 88.396,
    "args": {
      "External id": 2480335,"Record function id": 0, "Ev Idx": 8038
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454629.076, "dur": 47.318,
    "args": {
      "External id": 2480336,"Record function id": 0, "Ev Idx": 8039
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.26)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454685.567, "dur": 18506.908,
    "args": {
      "External id": 2480337,"Record function id": 0, "Ev Idx": 8040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.26)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454693.764, "dur": 873.192,
    "args": {
      "External id": 2480338,"Record function id": 0, "Ev Idx": 8041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454772.377, "dur": 8.981,
    "args": {
      "External id": 2480339,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454794.072, "dur": 31.738,
    "args": {
      "External id": 2480340,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 8043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454799.124, "dur": 2.342,
    "args": {
      "External id": 2480341,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454805.226, "dur": 0.240,
    "args": {
      "External id": 2480342,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454806.681, "dur": 0.261,
    "args": {
      "External id": 2480343,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454807.993, "dur": 0.346,
    "args": {
      "External id": 2480344,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454810.958, "dur": 0.248,
    "args": {
      "External id": 2480345,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454812.227, "dur": 0.334,
    "args": {
      "External id": 2480346,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454813.773, "dur": 3.034,
    "args": {
      "External id": 2480347,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454818.398, "dur": 0.375,
    "args": {
      "External id": 2480348,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454819.962, "dur": 0.195,
    "args": {
      "External id": 2480349,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454836.372, "dur": 36.827,
    "args": {
      "External id": 2480350,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 8053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454904.886, "dur": 143.394,
    "args": {
      "External id": 2480351,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 8054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454915.361, "dur": 3.944,
    "args": {
      "External id": 2480352,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454924.104, "dur": 10.045,
    "args": {
      "External id": 2480353,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 8056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454928.465, "dur": 5.290,
    "args": {
      "External id": 2480354,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 8057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454932.011, "dur": 0.540,
    "args": {
      "External id": 2480355,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 8058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454940.780, "dur": 27.518,
    "args": {
      "External id": 2480356,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 8059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454942.683, "dur": 2.162,
    "args": {
      "External id": 2480357,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454946.607, "dur": 0.332,
    "args": {
      "External id": 2480358,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454950.416, "dur": 0.152,
    "args": {
      "External id": 2480359,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454953.657, "dur": 0.500,
    "args": {
      "External id": 2480360,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454955.113, "dur": 0.382,
    "args": {
      "External id": 2480361,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454956.288, "dur": 2.166,
    "args": {
      "External id": 2480362,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454959.291, "dur": 0.715,
    "args": {
      "External id": 2480363,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454960.917, "dur": 0.586,
    "args": {
      "External id": 2480364,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454964.346, "dur": 0.514,
    "args": {
      "External id": 2480365,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656454977.333, "dur": 61.056,
    "args": {
      "External id": 2480366,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 8069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656455103.121, "dur": 351.978,
    "args": {
      "External id": 2480367,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 8070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656455132.993, "dur": 317.159,
    "args": {
      "External id": 2480368,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 8071, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1336759,
    "ts": 1514656455143.618, "dur": 301.374,
    "args": {
      "External id": 2480369,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 8072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656455475.811, "dur": 2.037,
    "args": {
      "External id": 2480370,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 8073, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.26)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656455589.017, "dur": 17388.416,
    "args": {
      "External id": 2480371,"Record function id": 0, "Ev Idx": 8074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656455682.491, "dur": 6.493,
    "args": {
      "External id": 2480372,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 8075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656455692.254, "dur": 1.115,
    "args": {
      "External id": 2480373,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 8076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656455701.580, "dur": 1.116,
    "args": {
      "External id": 2480374,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 8077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656455706.416, "dur": 0.615,
    "args": {
      "External id": 2480375,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 8078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656455708.500, "dur": 0.816,
    "args": {
      "External id": 2480376,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 8079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656455710.700, "dur": 1.093,
    "args": {
      "External id": 2480377,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 8080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656455713.270, "dur": 0.853,
    "args": {
      "External id": 2480378,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 8081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656455717.580, "dur": 1.784,
    "args": {
      "External id": 2480379,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 8082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656455720.642, "dur": 1.029,
    "args": {
      "External id": 2480380,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 8083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656455722.893, "dur": 0.779,
    "args": {
      "External id": 2480381,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 8084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656455742.471, "dur": 17198.547,
    "args": {
      "External id": 2480382,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 8085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656455757.227, "dur": 17177.327,
    "args": {
      "External id": 2480383,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 8086
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656455775.133, "dur": 14.767,
    "args": {
      "External id": 2480384,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656455794.554, "dur": 17110.696,
    "args": {
      "External id": 2480385,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 8088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656455796.962, "dur": 17107.768,
    "args": {
      "External id": 2480386,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 8089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656455802.324, "dur": 5.555,
    "args": {
      "External id": 2480387,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656455809.386, "dur": 17092.640,
    "args": {
      "External id": 2480388,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 8091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656473135.382, "dur": 32.981,
    "args": {
      "External id": 2480389,"Sequence number": 24740473, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 8092
    }
  },
  {
    "ph": "s", "id": 175, "pid": 1336759, "tid": 1336759, "ts": 1514656473135.382,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514656473155.626, "dur": 7.737,
    "args": {
      "External id": 2480390,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 8093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656473158.599, "dur": 4.405,
    "args": {
      "External id": 2480391,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1336759, "tid": 1336759,
    "ts": 1514656473227.955, "dur": 71.675,
    "args": {
      "External id": 2480392,"Record function id": 0, "Ev Idx": 8095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656473301.242, "dur": 1018.126,
    "args": {
      "External id": 2480393,"Record function id": 0, "Ev Idx": 8096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656473337.641, "dur": 968.900,
    "args": {
      "External id": 2480394,"Sequence number": 24740474, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [32, 1], [32, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], "Input Dims": [[16, 4096, 2048], [2048], [2048, 2048], [2048, 2048], [2048, 2048], [8192, 32], [8192, 32], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], "Ev Idx": 8097
    }
  },
  {
    "ph": "s", "id": 174, "pid": 1336759, "tid": 1336759, "ts": 1514656473337.641,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656473399.453, "dur": 40.463,
    "args": {
      "External id": 2480395,"kernel_hash": "cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 8098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656473451.702, "dur": 105.183,
    "args": {
      "External id": 2480396,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 8099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656473569.226, "dur": 42.514,
    "args": {
      "External id": 2480397,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 8100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656473620.063, "dur": 30.751,
    "args": {
      "External id": 2480398,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 8101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656473673.760, "dur": 25.848,
    "args": {
      "External id": 2480399,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 8102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656473718.234, "dur": 14.236,
    "args": {
      "External id": 2480400,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 8103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656473749.550, "dur": 120.490,
    "args": {
      "External id": 2480401,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 8104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656473796.647, "dur": 10.645,
    "args": {
      "External id": 2480402,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 8105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656473801.337, "dur": 5.195,
    "args": {
      "External id": 2480403,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656473809.952, "dur": 5.442,
    "args": {
      "External id": 2480404,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656473816.549, "dur": 1.795,
    "args": {
      "External id": 2480405,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656473820.508, "dur": 2.963,
    "args": {
      "External id": 2480406,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656473880.628, "dur": 42.785,
    "args": {
      "External id": 2480407,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 8110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 1336759, "tid": 1336759,
    "ts": 1514656473952.315, "dur": 25.012,
    "args": {
      "External id": 2480408,"kernel_hash": "c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/3l/c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 8111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474027.312, "dur": 46.069,
    "args": {
      "External id": 2480409,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 8112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474082.452, "dur": 34.513,
    "args": {
      "External id": 2480410,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 8113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474138.769, "dur": 23.991,
    "args": {
      "External id": 2480411,"kernel_hash": "ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/kq/ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [16, 4096, 5632], []], "Ev Idx": 8114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474168.501, "dur": 34.390,
    "args": {
      "External id": 2480412,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 8115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474222.855, "dur": 17.229,
    "args": {
      "External id": 2480413,"kernel_hash": "cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u6/cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [65536, 2048], []], "Ev Idx": 8116
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.26)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474380.010, "dur": 70.781,
    "args": {
      "External id": 2480414,"Record function id": 0, "Ev Idx": 8117
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::cast_forward_inputs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474522.607, "dur": 64.686,
    "args": {
      "External id": 2480415,"Record function id": 0, "Ev Idx": 8118
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::pre_forward (model.layers.27)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474597.001, "dur": 18590.781,
    "args": {
      "External id": 2480416,"Record function id": 0, "Ev Idx": 8119
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather (model.layers.27)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474606.673, "dur": 865.937,
    "args": {
      "External id": 2480417,"Record function id": 0, "Ev Idx": 8120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474685.370, "dur": 9.131,
    "args": {
      "External id": 2480418,"Record function id": 0, "Concrete Inputs": ["[6423040]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474707.136, "dur": 38.760,
    "args": {
      "External id": 2480419,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 8122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474712.094, "dur": 4.620,
    "args": {
      "External id": 2480420,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474724.847, "dur": 0.723,
    "args": {
      "External id": 2480421,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "256"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474726.841, "dur": 0.679,
    "args": {
      "External id": 2480422,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "524544"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474728.554, "dur": 0.458,
    "args": {
      "External id": 2480423,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1048832"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474731.759, "dur": 0.510,
    "args": {
      "External id": 2480424,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "1573120"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474733.237, "dur": 0.742,
    "args": {
      "External id": 2480425,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "2097408"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474735.055, "dur": 2.209,
    "args": {
      "External id": 2480426,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "2097664"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474738.324, "dur": 0.654,
    "args": {
      "External id": 2480427,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "3539456"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474739.718, "dur": 0.583,
    "args": {
      "External id": 2480428,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "4981248"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474757.180, "dur": 39.123,
    "args": {
      "External id": 2480429,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 8132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::all_gather_copy_in", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474828.478, "dur": 100.832,
    "args": {
      "External id": 2480430,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "6423040", "8", "6", "15", ""], "Input type": ["TensorList", "ScalarList", "Scalar", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [], [], [], [], [], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [], [], [], [], [], []], "Ev Idx": 8133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474839.004, "dur": 4.667,
    "args": {
      "External id": 2480431,"Record function id": 0, "Concrete Inputs": ["[51384320]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::narrow", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474848.481, "dur": 9.832,
    "args": {
      "External id": 2480432,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "6423040"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 8135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474852.608, "dur": 5.307,
    "args": {
      "External id": 2480433,"Record function id": 0, "Concrete Inputs": ["", "0", "38538240", "44961280", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[51384320], [], [], [], []], "Ev Idx": 8136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474855.975, "dur": 0.837,
    "args": {
      "External id": 2480434,"Record function id": 0, "Concrete Inputs": ["", "[6423040]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[51384320], [], [], []], "Ev Idx": 8137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474864.831, "dur": 24.545,
    "args": {
      "External id": 2480435,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[6423040], [], []], "Ev Idx": 8138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474866.421, "dur": 2.396,
    "args": {
      "External id": 2480436,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "38538240"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474869.684, "dur": 0.524,
    "args": {
      "External id": 2480437,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "38538496"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474871.330, "dur": 0.603,
    "args": {
      "External id": 2480438,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39062784"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474874.704, "dur": 0.588,
    "args": {
      "External id": 2480439,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "39587072"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474876.038, "dur": 0.582,
    "args": {
      "External id": 2480440,"Record function id": 0, "Concrete Inputs": ["", "[524288]", "[1]", "40111360"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474877.451, "dur": 0.624,
    "args": {
      "External id": 2480441,"Record function id": 0, "Concrete Inputs": ["", "[256]", "[1]", "40635648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474881.180, "dur": 0.378,
    "args": {
      "External id": 2480442,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "40635904"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474882.568, "dur": 0.564,
    "args": {
      "External id": 2480443,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "42077696"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474883.801, "dur": 1.837,
    "args": {
      "External id": 2480444,"Record function id": 0, "Concrete Inputs": ["", "[1441792]", "[1]", "43519488"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[6423040], [], [], []], "Ev Idx": 8147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656474900.569, "dur": 21.259,
    "args": {
      "External id": 2480445,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["TensorList", "TensorList", "Scalar"], "Input Strides": [[[1], [1], [1], [1], [1], [1], [1], [1], [1]], [[1], [1], [1], [1], [1], [1], [1], [1], [1]], []], "Input Dims": [[[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], [[256], [524288], [524288], [524288], [524288], [256], [1441792], [1441792], [1441792]], []], "Ev Idx": 8148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::_allgather_base_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656475015.499, "dur": 364.002,
    "args": {
      "External id": 2480446,"Record function id": 0, "Concrete Inputs": ["", "", "", "False", "-1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "Scalar", "Scalar"], "Input Strides": [[1], [1], [], [], []], "Input Dims": [[51384320], [6423040], [], [], []], "Ev Idx": 8149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656475046.894, "dur": 328.308,
    "args": {
      "External id": 2480447,"Record function id": 0, "Collective name": "_allgather_base", "Process Group Description": "default_pg", "dtype": "BFloat16", "Rank": 6, "Input Strides": [[1], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 51384320, "Process Group Name": "0", "Input type": ["c10::BFloat16", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[6423040], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 8150, "In msg nelems": 6423040
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:_all_gather_base", "pid": 1336759, "tid": 1336759,
    "ts": 1514656475058.012, "dur": 311.831,
    "args": {
      "External id": 2480448,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[1]], "Input Dims": [[6423040]], "Ev Idx": 8151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514656475400.521, "dur": 1.894,
    "args": {
      "External id": 2480449,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 8152, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::all_gather_copy_out (model.layers.27)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656475493.594, "dur": 17460.760,
    "args": {
      "External id": 2480450,"Record function id": 0, "Ev Idx": 8153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656475624.773, "dur": 7.543,
    "args": {
      "External id": 2480451,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[51384320], []], "Ev Idx": 8154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656475635.937, "dur": 0.986,
    "args": {
      "External id": 2480452,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 8155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656475638.716, "dur": 1.537,
    "args": {
      "External id": 2480453,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 8156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656475646.702, "dur": 0.860,
    "args": {
      "External id": 2480454,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 8157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656475648.863, "dur": 0.727,
    "args": {
      "External id": 2480455,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 8158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656475650.699, "dur": 0.968,
    "args": {
      "External id": 2480456,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[4194304], []], "Ev Idx": 8159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656475653.143, "dur": 1.078,
    "args": {
      "External id": 2480457,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[2048], []], "Ev Idx": 8160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656475657.180, "dur": 2.128,
    "args": {
      "External id": 2480458,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 8161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656475660.733, "dur": 1.075,
    "args": {
      "External id": 2480459,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 8162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656475663.034, "dur": 0.919,
    "args": {
      "External id": 2480460,"Record function id": 0, "Concrete Inputs": ["", "[8, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[11534336], []], "Ev Idx": 8163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "fsdp::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656475683.327, "dur": 17230.920,
    "args": {
      "External id": 2480461,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 8164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::split_with_sizes_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656475699.009, "dur": 17208.929,
    "args": {
      "External id": 2480462,"Record function id": 0, "Concrete Inputs": ["", "[256, 524288, 524288, 524288, 524288, 256, 1441792, 1441792, 1441792]", "1", ""], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "TensorList"], "Input Strides": [[6423040, 1], [], [], [[256, 1], [524288, 1], [524288, 1], [524288, 1], [524288, 1], [256, 1], [1441792, 1], [1441792, 1], [1441792, 1]]], "Input Dims": [[8, 6423040], [], [], [[8, 256], [8, 524288], [8, 524288], [8, 524288], [8, 524288], [8, 256], [8, 1441792], [8, 1441792], [8, 1441792]]], "Ev Idx": 8165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656475720.388, "dur": 12.536,
    "args": {
      "External id": 2480463,"Record function id": 0, "Concrete Inputs": ["[2133]", "4", "", "", "True", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656475737.851, "dur": 17140.173,
    "args": {
      "External id": 2480464,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", "False", ""], "Input type": ["long int", "", "", "", "", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], [], []], "Ev Idx": 8167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656475740.166, "dur": 17137.343,
    "args": {
      "External id": 2480465,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "True", ""], "Input type": ["long int", "", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[2133], [], [], [], [], [], []], "Ev Idx": 8168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656475745.248, "dur": 5.179,
    "args": {
      "External id": 2480466,"Record function id": 0, "Concrete Inputs": ["[2133]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656475752.034, "dur": 17122.632,
    "args": {
      "External id": 2480467,"Record function id": 0, "Concrete Inputs": ["", "", "True"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[2133], [2133], []], "Ev Idx": 8170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RegisterPostBackwardFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656493127.754, "dur": 34.697,
    "args": {
      "External id": 2480468,"Sequence number": 24740475, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048]], "Ev Idx": 8171
    }
  },
  {
    "ph": "s", "id": 173, "pid": 1336759, "tid": 1336759, "ts": 1514656493127.754,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514656493149.261, "dur": 8.101,
    "args": {
      "External id": 2480469,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1]], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048]], "Ev Idx": 8172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656493152.090, "dur": 4.873,
    "args": {
      "External id": 2480470,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1336759, "tid": 1336759,
    "ts": 1514656493225.580, "dur": 70.913,
    "args": {
      "External id": 2480471,"Record function id": 0, "Ev Idx": 8174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 1/1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656493297.913, "dur": 1023.418,
    "args": {
      "External id": 2480472,"Record function id": 0, "Ev Idx": 8175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656493334.337, "dur": 974.201,
    "args": {
      "External id": 2480473,"Sequence number": 24740476, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "", "", "", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1], [2048, 1], [2048, 1], [2048, 1], [32, 1], [32, 1], [2048, 1], [1], [2048, 1], [2048, 1], [5632, 1]], "Input Dims": [[16, 4096, 2048], [2048], [2048, 2048], [2048, 2048], [2048, 2048], [8192, 32], [8192, 32], [2048, 2048], [2048], [5632, 2048], [5632, 2048], [2048, 5632]], "Ev Idx": 8176
    }
  },
  {
    "ph": "s", "id": 172, "pid": 1336759, "tid": 1336759, "ts": 1514656493334.337,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656493398.595, "dur": 40.096,
    "args": {
      "External id": 2480474,"kernel_hash": "cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 8177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656493451.161, "dur": 103.950,
    "args": {
      "External id": 2480475,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 8178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656493567.321, "dur": 41.724,
    "args": {
      "External id": 2480476,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 8179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656493617.886, "dur": 31.086,
    "args": {
      "External id": 2480477,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 8180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656493673.002, "dur": 24.225,
    "args": {
      "External id": 2480478,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 8181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "rotary_embedding_kernel_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656493715.705, "dur": 14.412,
    "args": {
      "External id": 2480479,"kernel_hash": "c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "0", "16", "4096", "32", "64", "32", "8192", "32", "64", "False", "False", "False", "False"], "kernel_file": "/tmp/torchinductor_cvm/5n/c5nleexjfe7ckjbhvsxfsribge47peqkyivjpkls2h4yztzme4px.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 8182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656493747.282, "dur": 120.655,
    "args": {
      "External id": 2480480,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 8183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656493795.485, "dur": 10.931,
    "args": {
      "External id": 2480481,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 8184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656493800.214, "dur": 5.485,
    "args": {
      "External id": 2480482,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656493808.867, "dur": 4.616,
    "args": {
      "External id": 2480483,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656493814.834, "dur": 1.413,
    "args": {
      "External id": 2480484,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656493818.657, "dur": 2.793,
    "args": {
      "External id": 2480485,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656493878.246, "dur": 43.693,
    "args": {
      "External id": 2480486,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [2048, 1]], "Input Dims": [[65536, 2048], [2048, 2048], [65536, 2048]], "Ev Idx": 8189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_2", "pid": 1336759, "tid": 1336759,
    "ts": 1514656493949.932, "dur": 26.435,
    "args": {
      "External id": 2480487,"kernel_hash": "c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "True", "True", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/3l/c3la5vptoeqypxfxfvfv4bdw7i4xiukijyga2gm6nuqgqnu7xukn.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [2048, 1], [2048, 1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536, 2048], [2048], [65536, 2048], [65536, 2048], [65536], [], [], [], [], [], [], [], [], []], "Ev Idx": 8190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656494023.333, "dur": 45.767,
    "args": {
      "External id": 2480488,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 8191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656494079.082, "dur": 34.528,
    "args": {
      "External id": 2480489,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048], [5632, 1]], "Input Dims": [[65536, 2048], [2048, 5632], [65536, 5632]], "Ev Idx": 8192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused__to_copy_mul_silu_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656494137.041, "dur": 27.340,
    "args": {
      "External id": 2480490,"kernel_hash": "ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg", "grid": "grid(369098752,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "369098752"], "kernel_file": "/tmp/torchinductor_cvm/kq/ckqncfcfp3lv2zt52bvm43teupuhix6t4d4srlqovp6jq4mh7ayg.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[5632, 1], [5632, 1], [23068672, 5632, 1], []], "Input Dims": [[65536, 5632], [65536, 5632], [16, 4096, 5632], []], "Ev Idx": 8193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656494170.253, "dur": 33.387,
    "args": {
      "External id": 2480491,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632], [2048, 1]], "Input Dims": [[65536, 5632], [5632, 2048], [65536, 2048]], "Ev Idx": 8194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "triton_poi_fused_add_1", "pid": 1336759, "tid": 1336759,
    "ts": 1514656494224.651, "dur": 17.531,
    "args": {
      "External id": 2480492,"kernel_hash": "cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p", "grid": "grid(134217728,)", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "134217728"], "kernel_file": "/tmp/torchinductor_cvm/u6/cu627l3x7nht56wplsa2rqn4a3r4dce6couhabvxwttkx3auhd6p.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [65536, 2048], []], "Ev Idx": 8195
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward (model.layers.27)", "pid": 1336759, "tid": 1336759,
    "ts": 1514656494383.135, "dur": 31.060,
    "args": {
      "External id": 2480493,"Record function id": 0, "Ev Idx": 8196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656494561.353, "dur": 280.174,
    "args": {
      "External id": 2480494,"Sequence number": 24740477, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "9.9999999999999995e-07", "False", "False", "True"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 1], [1], [], [], [], [], [], []], "Input Dims": [[16, 4096, 2048], [2048], [], [], [], [], [], []], "Ev Idx": 8197
    }
  },
  {
    "ph": "s", "id": 171, "pid": 1336759, "tid": 1336759, "ts": 1514656494561.353,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656494594.392, "dur": 8.778,
    "args": {
      "External id": 2480495,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656494596.369, "dur": 6.429,
    "args": {
      "External id": 2480496,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656494612.685, "dur": 11.368,
    "args": {
      "External id": 2480497,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], [], []], "Ev Idx": 8200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656494615.343, "dur": 8.148,
    "args": {
      "External id": 2480498,"Record function id": 0, "Concrete Inputs": ["[65536, 2048]", "[2048, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656494632.782, "dur": 4.972,
    "args": {
      "External id": 2480499,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656494818.303, "dur": 9.265,
    "args": {
      "External id": 2480500,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656494824.141, "dur": 3.196,
    "args": {
      "External id": 2480501,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656494868.362, "dur": 170.577,
    "args": {
      "External id": 2480502,"Sequence number": 24740478, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [2048, 2048], []], "Ev Idx": 8205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656494871.251, "dur": 14.770,
    "args": {
      "External id": 2480503,"Sequence number": 24740478, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 8206
    }
  },
  {
    "ph": "s", "id": 170, "pid": 1336759, "tid": 1336759, "ts": 1514656494871.251,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656494876.979, "dur": 7.435,
    "args": {
      "External id": 2480504,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 8207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656494881.594, "dur": 2.420,
    "args": {
      "External id": 2480505,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 8208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656494888.696, "dur": 149.757,
    "args": {
      "External id": 2480506,"Sequence number": 24740479, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1, 2048]], "Input Dims": [[16, 4096, 2048], [2048, 2048]], "Ev Idx": 8209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656494891.103, "dur": 4.723,
    "args": {
      "External id": 2480507,"Sequence number": 24740479, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656494891.927, "dur": 3.771,
    "args": {
      "External id": 2480508,"Sequence number": 24740479, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8211
    }
  },
  {
    "ph": "s", "id": 169, "pid": 1336759, "tid": 1336759, "ts": 1514656494891.927,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656494899.753, "dur": 127.322,
    "args": {
      "External id": 2480509,"Sequence number": 24740480, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 8212
    }
  },
  {
    "ph": "s", "id": 168, "pid": 1336759, "tid": 1336759, "ts": 1514656494899.753,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495032.398, "dur": 4.770,
    "args": {
      "External id": 2480510,"Sequence number": 24740481, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8213
    }
  },
  {
    "ph": "s", "id": 167, "pid": 1336759, "tid": 1336759, "ts": 1514656495032.398,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495050.982, "dur": 75.676,
    "args": {
      "External id": 2480511,"Sequence number": 24740482, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [2048, 2048], []], "Ev Idx": 8214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495051.864, "dur": 7.376,
    "args": {
      "External id": 2480512,"Sequence number": 24740482, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 8215
    }
  },
  {
    "ph": "s", "id": 166, "pid": 1336759, "tid": 1336759, "ts": 1514656495051.864,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495053.733, "dur": 4.356,
    "args": {
      "External id": 2480513,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 8216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495056.837, "dur": 1.027,
    "args": {
      "External id": 2480514,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 8217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495060.022, "dur": 66.323,
    "args": {
      "External id": 2480515,"Sequence number": 24740483, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1, 2048]], "Input Dims": [[16, 4096, 2048], [2048, 2048]], "Ev Idx": 8218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495061.295, "dur": 4.952,
    "args": {
      "External id": 2480516,"Sequence number": 24740483, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495061.789, "dur": 4.301,
    "args": {
      "External id": 2480517,"Sequence number": 24740483, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8220
    }
  },
  {
    "ph": "s", "id": 165, "pid": 1336759, "tid": 1336759, "ts": 1514656495061.789,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495067.051, "dur": 53.262,
    "args": {
      "External id": 2480518,"Sequence number": 24740484, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 8221
    }
  },
  {
    "ph": "s", "id": 164, "pid": 1336759, "tid": 1336759, "ts": 1514656495067.051,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495121.789, "dur": 4.152,
    "args": {
      "External id": 2480519,"Sequence number": 24740485, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8222
    }
  },
  {
    "ph": "s", "id": 163, "pid": 1336759, "tid": 1336759, "ts": 1514656495121.789,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495134.294, "dur": 60.145,
    "args": {
      "External id": 2480520,"Sequence number": 24740486, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [2048, 2048], []], "Ev Idx": 8223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495134.800, "dur": 5.413,
    "args": {
      "External id": 2480521,"Sequence number": 24740486, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 8224
    }
  },
  {
    "ph": "s", "id": 162, "pid": 1336759, "tid": 1336759, "ts": 1514656495134.800,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495136.718, "dur": 2.266,
    "args": {
      "External id": 2480522,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 8225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495138.165, "dur": 0.679,
    "args": {
      "External id": 2480523,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 8226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495142.774, "dur": 51.357,
    "args": {
      "External id": 2480524,"Sequence number": 24740487, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1, 2048]], "Input Dims": [[16, 4096, 2048], [2048, 2048]], "Ev Idx": 8227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495143.537, "dur": 4.489,
    "args": {
      "External id": 2480525,"Sequence number": 24740487, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495144.547, "dur": 3.343,
    "args": {
      "External id": 2480526,"Sequence number": 24740487, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8229
    }
  },
  {
    "ph": "s", "id": 161, "pid": 1336759, "tid": 1336759, "ts": 1514656495144.547,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495148.644, "dur": 38.632,
    "args": {
      "External id": 2480527,"Sequence number": 24740488, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 8230
    }
  },
  {
    "ph": "s", "id": 160, "pid": 1336759, "tid": 1336759, "ts": 1514656495148.644,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495188.775, "dur": 5.028,
    "args": {
      "External id": 2480528,"Sequence number": 24740489, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8231
    }
  },
  {
    "ph": "s", "id": 159, "pid": 1336759, "tid": 1336759, "ts": 1514656495188.775,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495217.823, "dur": 3.931,
    "args": {
      "External id": 2480529,"Sequence number": 24740490, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495218.844, "dur": 2.765,
    "args": {
      "External id": 2480530,"Sequence number": 24740490, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8233
    }
  },
  {
    "ph": "s", "id": 158, "pid": 1336759, "tid": 1336759, "ts": 1514656495218.844,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495228.138, "dur": 3.076,
    "args": {
      "External id": 2480531,"Sequence number": 24740491, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495229.265, "dur": 1.817,
    "args": {
      "External id": 2480532,"Sequence number": 24740491, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8235
    }
  },
  {
    "ph": "s", "id": 157, "pid": 1336759, "tid": 1336759, "ts": 1514656495229.265,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495237.005, "dur": 4.210,
    "args": {
      "External id": 2480533,"Sequence number": 24740492, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495237.914, "dur": 3.162,
    "args": {
      "External id": 2480534,"Sequence number": 24740492, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8237
    }
  },
  {
    "ph": "s", "id": 156, "pid": 1336759, "tid": 1336759, "ts": 1514656495237.914,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495276.757, "dur": 172.400,
    "args": {
      "External id": 2480535,"Sequence number": 24740493, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "False", "False", "0", "", "8192"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [], [], [], [], []], "Ev Idx": 8238
    }
  },
  {
    "ph": "s", "id": 155, "pid": 1336759, "tid": 1336759, "ts": 1514656495276.757,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495298.056, "dur": 8.593,
    "args": {
      "External id": 2480536,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 8239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495300.767, "dur": 5.371,
    "args": {
      "External id": 2480537,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495462.176, "dur": 127.105,
    "args": {
      "External id": 2480538,"Sequence number": 24740494, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "False", "False", "0", "", "8192"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [], [], [], [], []], "Ev Idx": 8241
    }
  },
  {
    "ph": "s", "id": 154, "pid": 1336759, "tid": 1336759, "ts": 1514656495462.176,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495475.810, "dur": 7.226,
    "args": {
      "External id": 2480539,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 8242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495478.103, "dur": 4.466,
    "args": {
      "External id": 2480540,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FlashAttnFunc", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495620.296, "dur": 179.924,
    "args": {
      "External id": 2480541,"Sequence number": 24740495, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "", "True", "", "0.", "", "False", "False", "True"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "", "Scalar", "", "Scalar", "", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], []], "Ev Idx": 8244
    }
  },
  {
    "ph": "s", "id": 153, "pid": 1336759, "tid": 1336759, "ts": 1514656495620.296,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495649.496, "dur": 122.383,
    "args": {
      "External id": 2480542,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 8245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495699.323, "dur": 7.810,
    "args": {
      "External id": 2480543,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 8246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495701.993, "dur": 4.476,
    "args": {
      "External id": 2480544,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495709.725, "dur": 3.848,
    "args": {
      "External id": 2480545,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495714.781, "dur": 1.394,
    "args": {
      "External id": 2480546,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495720.361, "dur": 3.258,
    "args": {
      "External id": 2480547,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495786.213, "dur": 5.223,
    "args": {
      "External id": 2480548,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 8251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495804.947, "dur": 6.007,
    "args": {
      "External id": 2480549,"Sequence number": 24740496, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 8252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495806.268, "dur": 4.560,
    "args": {
      "External id": 2480550,"Sequence number": 24740496, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 8253
    }
  },
  {
    "ph": "s", "id": 152, "pid": 1336759, "tid": 1336759, "ts": 1514656495806.268,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495827.992, "dur": 118.208,
    "args": {
      "External id": 2480551,"Sequence number": 24740497, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [2048, 2048], []], "Ev Idx": 8254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495830.662, "dur": 7.566,
    "args": {
      "External id": 2480552,"Sequence number": 24740497, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 8255
    }
  },
  {
    "ph": "s", "id": 151, "pid": 1336759, "tid": 1336759, "ts": 1514656495830.662,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495833.161, "dur": 4.161,
    "args": {
      "External id": 2480553,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 8256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495835.409, "dur": 1.639,
    "args": {
      "External id": 2480554,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 8257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495839.416, "dur": 106.467,
    "args": {
      "External id": 2480555,"Sequence number": 24740498, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1, 2048]], "Input Dims": [[16, 4096, 2048], [2048, 2048]], "Ev Idx": 8258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495841.346, "dur": 7.797,
    "args": {
      "External id": 2480556,"Sequence number": 24740498, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495843.902, "dur": 5.086,
    "args": {
      "External id": 2480557,"Sequence number": 24740498, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8260
    }
  },
  {
    "ph": "s", "id": 150, "pid": 1336759, "tid": 1336759, "ts": 1514656495843.902,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495850.309, "dur": 87.457,
    "args": {
      "External id": 2480558,"Sequence number": 24740499, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 8261
    }
  },
  {
    "ph": "s", "id": 149, "pid": 1336759, "tid": 1336759, "ts": 1514656495850.309,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495940.240, "dur": 4.912,
    "args": {
      "External id": 2480559,"Sequence number": 24740500, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8262
    }
  },
  {
    "ph": "s", "id": 148, "pid": 1336759, "tid": 1336759, "ts": 1514656495940.240,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656495977.467, "dur": 264.123,
    "args": {
      "External id": 2480560,"Sequence number": 24740501, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "9.9999999999999995e-07", "True", "False", "True"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 1], [1], [], [8388608, 2048, 1], [], [], [], []], "Input Dims": [[16, 4096, 2048], [2048], [], [16, 4096, 2048], [], [], [], []], "Ev Idx": 8263
    }
  },
  {
    "ph": "s", "id": 147, "pid": 1336759, "tid": 1336759, "ts": 1514656495977.467,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496038.795, "dur": 3.489,
    "args": {
      "External id": 2480561,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496039.599, "dur": 2.428,
    "args": {
      "External id": 2480562,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496048.572, "dur": 3.265,
    "args": {
      "External id": 2480563,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [2048, 1]], "Input Dims": [[16, 4096, 2048], [65536, 2048]], "Ev Idx": 8266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496049.616, "dur": 2.115,
    "args": {
      "External id": 2480564,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496050.327, "dur": 1.313,
    "args": {
      "External id": 2480565,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496059.208, "dur": 9.402,
    "args": {
      "External id": 2480566,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], [], []], "Ev Idx": 8269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496062.974, "dur": 5.308,
    "args": {
      "External id": 2480567,"Record function id": 0, "Concrete Inputs": ["[65536, 2048]", "[2048, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496074.495, "dur": 3.687,
    "args": {
      "External id": 2480568,"Record function id": 0, "Concrete Inputs": ["[65536, 2048]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496081.918, "dur": 2.726,
    "args": {
      "External id": 2480569,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496218.084, "dur": 3.111,
    "args": {
      "External id": 2480570,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496219.245, "dur": 1.796,
    "args": {
      "External id": 2480571,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496223.829, "dur": 2.090,
    "args": {
      "External id": 2480572,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496224.791, "dur": 0.975,
    "args": {
      "External id": 2480573,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496260.276, "dur": 95.217,
    "args": {
      "External id": 2480574,"Sequence number": 24740502, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [5632, 2048], []], "Ev Idx": 8277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496261.350, "dur": 7.595,
    "args": {
      "External id": 2480575,"Sequence number": 24740502, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 8278
    }
  },
  {
    "ph": "s", "id": 146, "pid": 1336759, "tid": 1336759, "ts": 1514656496261.350,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496263.775, "dur": 3.983,
    "args": {
      "External id": 2480576,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[5632, 2048], [], []], "Ev Idx": 8279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496265.680, "dur": 1.819,
    "args": {
      "External id": 2480577,"Record function id": 0, "Concrete Inputs": ["", "[2048, 5632]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[5632, 2048], [], [], []], "Ev Idx": 8280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496269.783, "dur": 85.399,
    "args": {
      "External id": 2480578,"Sequence number": 24740503, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1, 2048]], "Input Dims": [[16, 4096, 2048], [2048, 5632]], "Ev Idx": 8281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496272.828, "dur": 3.809,
    "args": {
      "External id": 2480579,"Sequence number": 24740503, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496273.758, "dur": 2.749,
    "args": {
      "External id": 2480580,"Sequence number": 24740503, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8283
    }
  },
  {
    "ph": "s", "id": 145, "pid": 1336759, "tid": 1336759, "ts": 1514656496273.758,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496277.536, "dur": 72.071,
    "args": {
      "External id": 2480581,"Sequence number": 24740504, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[65536, 2048], [2048, 5632]], "Ev Idx": 8284
    }
  },
  {
    "ph": "s", "id": 144, "pid": 1336759, "tid": 1336759, "ts": 1514656496277.536,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496351.615, "dur": 3.014,
    "args": {
      "External id": 2480582,"Sequence number": 24740505, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 1], []], "Input Dims": [[65536, 5632], []], "Ev Idx": 8285
    }
  },
  {
    "ph": "s", "id": 143, "pid": 1336759, "tid": 1336759, "ts": 1514656496351.615,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496364.080, "dur": 68.246,
    "args": {
      "External id": 2480583,"Sequence number": 24740506, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [5632, 2048], []], "Ev Idx": 8286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496366.629, "dur": 5.069,
    "args": {
      "External id": 2480584,"Sequence number": 24740506, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 8287
    }
  },
  {
    "ph": "s", "id": 142, "pid": 1336759, "tid": 1336759, "ts": 1514656496366.629,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496367.893, "dur": 2.856,
    "args": {
      "External id": 2480585,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[5632, 2048], [], []], "Ev Idx": 8288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496369.303, "dur": 1.289,
    "args": {
      "External id": 2480586,"Record function id": 0, "Concrete Inputs": ["", "[2048, 5632]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[5632, 2048], [], [], []], "Ev Idx": 8289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496372.346, "dur": 59.754,
    "args": {
      "External id": 2480587,"Sequence number": 24740507, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1, 2048]], "Input Dims": [[16, 4096, 2048], [2048, 5632]], "Ev Idx": 8290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496373.466, "dur": 7.340,
    "args": {
      "External id": 2480588,"Sequence number": 24740507, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496375.948, "dur": 4.677,
    "args": {
      "External id": 2480589,"Sequence number": 24740507, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8292
    }
  },
  {
    "ph": "s", "id": 141, "pid": 1336759, "tid": 1336759, "ts": 1514656496375.948,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496381.485, "dur": 44.491,
    "args": {
      "External id": 2480590,"Sequence number": 24740508, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[65536, 2048], [2048, 5632]], "Ev Idx": 8293
    }
  },
  {
    "ph": "s", "id": 140, "pid": 1336759, "tid": 1336759, "ts": 1514656496381.485,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496427.517, "dur": 4.247,
    "args": {
      "External id": 2480591,"Sequence number": 24740509, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 1], []], "Input Dims": [[65536, 5632], []], "Ev Idx": 8294
    }
  },
  {
    "ph": "s", "id": 139, "pid": 1336759, "tid": 1336759, "ts": 1514656496427.517,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SwiGLULinearFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496455.075, "dur": 177.954,
    "args": {
      "External id": 2480592,"Sequence number": 24740510, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[23068672, 5632, 1], [23068672, 5632, 1], [5632, 1], []], "Input Dims": [[16, 4096, 5632], [16, 4096, 5632], [2048, 5632], []], "Ev Idx": 8295
    }
  },
  {
    "ph": "s", "id": 138, "pid": 1336759, "tid": 1336759, "ts": 1514656496455.075,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496496.734, "dur": 4.712,
    "args": {
      "External id": 2480593,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 5632]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496554.086, "dur": 64.444,
    "args": {
      "External id": 2480594,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[23068672, 5632, 1], [5632, 1], []], "Input Dims": [[16, 4096, 5632], [2048, 5632], []], "Ev Idx": 8297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496554.967, "dur": 6.900,
    "args": {
      "External id": 2480595,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 8298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496556.734, "dur": 4.062,
    "args": {
      "External id": 2480596,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[5632, 1], [], []], "Input Dims": [[2048, 5632], [], []], "Ev Idx": 8299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496558.857, "dur": 1.617,
    "args": {
      "External id": 2480597,"Record function id": 0, "Concrete Inputs": ["", "[5632, 2048]", "[1, 5632]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[5632, 1], [], [], []], "Input Dims": [[2048, 5632], [], [], []], "Ev Idx": 8300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496562.652, "dur": 55.522,
    "args": {
      "External id": 2480598,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[23068672, 5632, 1], [1, 5632]], "Input Dims": [[16, 4096, 5632], [5632, 2048]], "Ev Idx": 8301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496565.921, "dur": 2.638,
    "args": {
      "External id": 2480599,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 8302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496566.793, "dur": 1.670,
    "args": {
      "External id": 2480600,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 8303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496569.341, "dur": 45.343,
    "args": {
      "External id": 2480601,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632]], "Input Dims": [[65536, 5632], [5632, 2048]], "Ev Idx": 8304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496616.220, "dur": 1.319,
    "args": {
      "External id": 2480602,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496642.439, "dur": 27.177,
    "args": {
      "External id": 2480603,"Sequence number": 24740511, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 8306
    }
  },
  {
    "ph": "s", "id": 137, "pid": 1336759, "tid": 1336759, "ts": 1514656496642.439,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496703.215, "dur": 169.216,
    "args": {
      "External id": 2480604,"Sequence number": 24740512, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "9.9999999999999995e-07", "False", "False", "True"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 1], [1], [], [], [], [], [], []], "Input Dims": [[16, 4096, 2048], [2048], [], [], [], [], [], []], "Ev Idx": 8307
    }
  },
  {
    "ph": "s", "id": 136, "pid": 1336759, "tid": 1336759, "ts": 1514656496703.215,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496722.058, "dur": 4.665,
    "args": {
      "External id": 2480605,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496724.619, "dur": 1.979,
    "args": {
      "External id": 2480606,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496733.860, "dur": 6.762,
    "args": {
      "External id": 2480607,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], [], []], "Ev Idx": 8310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496736.148, "dur": 4.112,
    "args": {
      "External id": 2480608,"Record function id": 0, "Concrete Inputs": ["[65536, 2048]", "[2048, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496746.087, "dur": 3.271,
    "args": {
      "External id": 2480609,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496859.474, "dur": 2.881,
    "args": {
      "External id": 2480610,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496860.448, "dur": 1.729,
    "args": {
      "External id": 2480611,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496890.031, "dur": 77.681,
    "args": {
      "External id": 2480612,"Sequence number": 24740513, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [2048, 2048], []], "Ev Idx": 8315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496890.937, "dur": 8.171,
    "args": {
      "External id": 2480613,"Sequence number": 24740513, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 8316
    }
  },
  {
    "ph": "s", "id": 135, "pid": 1336759, "tid": 1336759, "ts": 1514656496890.937,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496894.599, "dur": 3.400,
    "args": {
      "External id": 2480614,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 8317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496896.463, "dur": 1.364,
    "args": {
      "External id": 2480615,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 8318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496899.948, "dur": 67.497,
    "args": {
      "External id": 2480616,"Sequence number": 24740514, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1, 2048]], "Input Dims": [[16, 4096, 2048], [2048, 2048]], "Ev Idx": 8319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496901.359, "dur": 4.884,
    "args": {
      "External id": 2480617,"Sequence number": 24740514, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496903.543, "dur": 2.588,
    "args": {
      "External id": 2480618,"Sequence number": 24740514, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8321
    }
  },
  {
    "ph": "s", "id": 134, "pid": 1336759, "tid": 1336759, "ts": 1514656496903.543,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496906.919, "dur": 55.502,
    "args": {
      "External id": 2480619,"Sequence number": 24740515, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 8322
    }
  },
  {
    "ph": "s", "id": 133, "pid": 1336759, "tid": 1336759, "ts": 1514656496906.919,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496964.099, "dur": 2.807,
    "args": {
      "External id": 2480620,"Sequence number": 24740516, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8323
    }
  },
  {
    "ph": "s", "id": 132, "pid": 1336759, "tid": 1336759, "ts": 1514656496964.099,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496975.285, "dur": 107.677,
    "args": {
      "External id": 2480621,"Sequence number": 24740517, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [2048, 2048], []], "Ev Idx": 8324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496975.972, "dur": 43.706,
    "args": {
      "External id": 2480622,"Sequence number": 24740517, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 8325
    }
  },
  {
    "ph": "s", "id": 131, "pid": 1336759, "tid": 1336759, "ts": 1514656496975.972,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496981.201, "dur": 2.326,
    "args": {
      "External id": 2480623,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 8326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656496982.689, "dur": 0.688,
    "args": {
      "External id": 2480624,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 8327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497021.502, "dur": 61.194,
    "args": {
      "External id": 2480625,"Sequence number": 24740518, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1, 2048]], "Input Dims": [[16, 4096, 2048], [2048, 2048]], "Ev Idx": 8328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497022.891, "dur": 6.002,
    "args": {
      "External id": 2480626,"Sequence number": 24740518, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497024.034, "dur": 4.688,
    "args": {
      "External id": 2480627,"Sequence number": 24740518, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8330
    }
  },
  {
    "ph": "s", "id": 130, "pid": 1336759, "tid": 1336759, "ts": 1514656497024.034,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497029.678, "dur": 48.670,
    "args": {
      "External id": 2480628,"Sequence number": 24740519, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 8331
    }
  },
  {
    "ph": "s", "id": 129, "pid": 1336759, "tid": 1336759, "ts": 1514656497029.678,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497080.256, "dur": 1.932,
    "args": {
      "External id": 2480629,"Sequence number": 24740520, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8332
    }
  },
  {
    "ph": "s", "id": 128, "pid": 1336759, "tid": 1336759, "ts": 1514656497080.256,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497091.644, "dur": 60.731,
    "args": {
      "External id": 2480630,"Sequence number": 24740521, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [2048, 2048], []], "Ev Idx": 8333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497092.416, "dur": 8.384,
    "args": {
      "External id": 2480631,"Sequence number": 24740521, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 8334
    }
  },
  {
    "ph": "s", "id": 127, "pid": 1336759, "tid": 1336759, "ts": 1514656497092.416,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497093.685, "dur": 6.051,
    "args": {
      "External id": 2480632,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 8335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497097.129, "dur": 2.354,
    "args": {
      "External id": 2480633,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 8336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497101.583, "dur": 50.472,
    "args": {
      "External id": 2480634,"Sequence number": 24740522, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1, 2048]], "Input Dims": [[16, 4096, 2048], [2048, 2048]], "Ev Idx": 8337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497102.497, "dur": 4.035,
    "args": {
      "External id": 2480635,"Sequence number": 24740522, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497103.191, "dur": 3.003,
    "args": {
      "External id": 2480636,"Sequence number": 24740522, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8339
    }
  },
  {
    "ph": "s", "id": 126, "pid": 1336759, "tid": 1336759, "ts": 1514656497103.191,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497108.593, "dur": 37.253,
    "args": {
      "External id": 2480637,"Sequence number": 24740523, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 8340
    }
  },
  {
    "ph": "s", "id": 125, "pid": 1336759, "tid": 1336759, "ts": 1514656497108.593,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497147.580, "dur": 4.051,
    "args": {
      "External id": 2480638,"Sequence number": 24740524, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8341
    }
  },
  {
    "ph": "s", "id": 124, "pid": 1336759, "tid": 1336759, "ts": 1514656497147.580,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497169.518, "dur": 3.408,
    "args": {
      "External id": 2480639,"Sequence number": 24740525, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497170.272, "dur": 2.514,
    "args": {
      "External id": 2480640,"Sequence number": 24740525, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8343
    }
  },
  {
    "ph": "s", "id": 123, "pid": 1336759, "tid": 1336759, "ts": 1514656497170.272,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497210.208, "dur": 6.240,
    "args": {
      "External id": 2480641,"Sequence number": 24740526, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497213.031, "dur": 3.280,
    "args": {
      "External id": 2480642,"Sequence number": 24740526, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8345
    }
  },
  {
    "ph": "s", "id": 122, "pid": 1336759, "tid": 1336759, "ts": 1514656497213.031,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497220.879, "dur": 3.924,
    "args": {
      "External id": 2480643,"Sequence number": 24740527, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497221.481, "dur": 3.163,
    "args": {
      "External id": 2480644,"Sequence number": 24740527, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8347
    }
  },
  {
    "ph": "s", "id": 121, "pid": 1336759, "tid": 1336759, "ts": 1514656497221.481,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497253.636, "dur": 148.860,
    "args": {
      "External id": 2480645,"Sequence number": 24740528, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "False", "False", "0", "", "8192"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [], [], [], [], []], "Ev Idx": 8348
    }
  },
  {
    "ph": "s", "id": 120, "pid": 1336759, "tid": 1336759, "ts": 1514656497253.636,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497272.747, "dur": 7.287,
    "args": {
      "External id": 2480646,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 8349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497274.912, "dur": 4.615,
    "args": {
      "External id": 2480647,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497416.100, "dur": 100.053,
    "args": {
      "External id": 2480648,"Sequence number": 24740529, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "False", "False", "0", "", "8192"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [], [], [], [], []], "Ev Idx": 8351
    }
  },
  {
    "ph": "s", "id": 119, "pid": 1336759, "tid": 1336759, "ts": 1514656497416.100,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497429.914, "dur": 6.659,
    "args": {
      "External id": 2480649,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 8352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497431.955, "dur": 4.200,
    "args": {
      "External id": 2480650,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FlashAttnFunc", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497562.195, "dur": 173.483,
    "args": {
      "External id": 2480651,"Sequence number": 24740530, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "", "True", "", "0.", "", "False", "False", "True"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "", "Scalar", "", "Scalar", "", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], []], "Ev Idx": 8354
    }
  },
  {
    "ph": "s", "id": 118, "pid": 1336759, "tid": 1336759, "ts": 1514656497562.195,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497590.601, "dur": 119.129,
    "args": {
      "External id": 2480652,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 8355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497635.769, "dur": 8.110,
    "args": {
      "External id": 2480653,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 8356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497637.871, "dur": 5.363,
    "args": {
      "External id": 2480654,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497646.350, "dur": 3.984,
    "args": {
      "External id": 2480655,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497651.490, "dur": 2.686,
    "args": {
      "External id": 2480656,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497656.575, "dur": 2.781,
    "args": {
      "External id": 2480657,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497720.513, "dur": 5.020,
    "args": {
      "External id": 2480658,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 8361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497740.361, "dur": 7.020,
    "args": {
      "External id": 2480659,"Sequence number": 24740531, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 8362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497741.567, "dur": 5.651,
    "args": {
      "External id": 2480660,"Sequence number": 24740531, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 8363
    }
  },
  {
    "ph": "s", "id": 117, "pid": 1336759, "tid": 1336759, "ts": 1514656497741.567,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497762.971, "dur": 108.999,
    "args": {
      "External id": 2480661,"Sequence number": 24740532, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [2048, 2048], []], "Ev Idx": 8364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497764.174, "dur": 10.717,
    "args": {
      "External id": 2480662,"Sequence number": 24740532, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 8365
    }
  },
  {
    "ph": "s", "id": 116, "pid": 1336759, "tid": 1336759, "ts": 1514656497764.174,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497766.488, "dur": 7.226,
    "args": {
      "External id": 2480663,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 8366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497771.444, "dur": 1.979,
    "args": {
      "External id": 2480664,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 8367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497775.977, "dur": 95.676,
    "args": {
      "External id": 2480665,"Sequence number": 24740533, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1, 2048]], "Input Dims": [[16, 4096, 2048], [2048, 2048]], "Ev Idx": 8368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497779.257, "dur": 4.547,
    "args": {
      "External id": 2480666,"Sequence number": 24740533, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497780.204, "dur": 3.429,
    "args": {
      "External id": 2480667,"Sequence number": 24740533, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8370
    }
  },
  {
    "ph": "s", "id": 115, "pid": 1336759, "tid": 1336759, "ts": 1514656497780.204,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497784.958, "dur": 78.245,
    "args": {
      "External id": 2480668,"Sequence number": 24740534, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 8371
    }
  },
  {
    "ph": "s", "id": 114, "pid": 1336759, "tid": 1336759, "ts": 1514656497784.958,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497865.835, "dur": 5.038,
    "args": {
      "External id": 2480669,"Sequence number": 24740535, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8372
    }
  },
  {
    "ph": "s", "id": 113, "pid": 1336759, "tid": 1336759, "ts": 1514656497865.835,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497902.739, "dur": 249.806,
    "args": {
      "External id": 2480670,"Sequence number": 24740536, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "9.9999999999999995e-07", "True", "False", "True"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 1], [1], [], [8388608, 2048, 1], [], [], [], []], "Input Dims": [[16, 4096, 2048], [2048], [], [16, 4096, 2048], [], [], [], []], "Ev Idx": 8373
    }
  },
  {
    "ph": "s", "id": 112, "pid": 1336759, "tid": 1336759, "ts": 1514656497902.739,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497923.494, "dur": 2.762,
    "args": {
      "External id": 2480671,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497924.452, "dur": 1.671,
    "args": {
      "External id": 2480672,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497930.260, "dur": 4.501,
    "args": {
      "External id": 2480673,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [2048, 1]], "Input Dims": [[16, 4096, 2048], [65536, 2048]], "Ev Idx": 8376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497931.589, "dur": 3.067,
    "args": {
      "External id": 2480674,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497933.559, "dur": 1.000,
    "args": {
      "External id": 2480675,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497941.639, "dur": 5.847,
    "args": {
      "External id": 2480676,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], [], []], "Ev Idx": 8379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497943.306, "dur": 3.858,
    "args": {
      "External id": 2480677,"Record function id": 0, "Concrete Inputs": ["[65536, 2048]", "[2048, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497953.241, "dur": 3.350,
    "args": {
      "External id": 2480678,"Record function id": 0, "Concrete Inputs": ["[65536, 2048]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656497959.885, "dur": 3.601,
    "args": {
      "External id": 2480679,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498126.560, "dur": 3.939,
    "args": {
      "External id": 2480680,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498127.398, "dur": 2.757,
    "args": {
      "External id": 2480681,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498135.046, "dur": 2.045,
    "args": {
      "External id": 2480682,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498135.847, "dur": 1.148,
    "args": {
      "External id": 2480683,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498170.320, "dur": 99.146,
    "args": {
      "External id": 2480684,"Sequence number": 24740537, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [5632, 2048], []], "Ev Idx": 8387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498171.621, "dur": 7.627,
    "args": {
      "External id": 2480685,"Sequence number": 24740537, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 8388
    }
  },
  {
    "ph": "s", "id": 111, "pid": 1336759, "tid": 1336759, "ts": 1514656498171.621,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498174.297, "dur": 3.945,
    "args": {
      "External id": 2480686,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[5632, 2048], [], []], "Ev Idx": 8389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498176.411, "dur": 1.559,
    "args": {
      "External id": 2480687,"Record function id": 0, "Concrete Inputs": ["", "[2048, 5632]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[5632, 2048], [], [], []], "Ev Idx": 8390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498181.819, "dur": 87.345,
    "args": {
      "External id": 2480688,"Sequence number": 24740538, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1, 2048]], "Input Dims": [[16, 4096, 2048], [2048, 5632]], "Ev Idx": 8391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498183.337, "dur": 4.141,
    "args": {
      "External id": 2480689,"Sequence number": 24740538, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498184.050, "dur": 3.300,
    "args": {
      "External id": 2480690,"Sequence number": 24740538, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8393
    }
  },
  {
    "ph": "s", "id": 110, "pid": 1336759, "tid": 1336759, "ts": 1514656498184.050,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498188.409, "dur": 73.584,
    "args": {
      "External id": 2480691,"Sequence number": 24740539, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[65536, 2048], [2048, 5632]], "Ev Idx": 8394
    }
  },
  {
    "ph": "s", "id": 109, "pid": 1336759, "tid": 1336759, "ts": 1514656498188.409,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498264.126, "dur": 4.431,
    "args": {
      "External id": 2480692,"Sequence number": 24740540, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 1], []], "Input Dims": [[65536, 5632], []], "Ev Idx": 8395
    }
  },
  {
    "ph": "s", "id": 108, "pid": 1336759, "tid": 1336759, "ts": 1514656498264.126,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498277.984, "dur": 66.892,
    "args": {
      "External id": 2480693,"Sequence number": 24740541, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [5632, 2048], []], "Ev Idx": 8396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498278.776, "dur": 5.268,
    "args": {
      "External id": 2480694,"Sequence number": 24740541, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 8397
    }
  },
  {
    "ph": "s", "id": 107, "pid": 1336759, "tid": 1336759, "ts": 1514656498278.776,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498280.001, "dur": 3.058,
    "args": {
      "External id": 2480695,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[5632, 2048], [], []], "Ev Idx": 8398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498281.752, "dur": 1.148,
    "args": {
      "External id": 2480696,"Record function id": 0, "Concrete Inputs": ["", "[2048, 5632]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[5632, 2048], [], [], []], "Ev Idx": 8399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498284.554, "dur": 60.046,
    "args": {
      "External id": 2480697,"Sequence number": 24740542, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1, 2048]], "Input Dims": [[16, 4096, 2048], [2048, 5632]], "Ev Idx": 8400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498287.484, "dur": 5.594,
    "args": {
      "External id": 2480698,"Sequence number": 24740542, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498288.245, "dur": 4.684,
    "args": {
      "External id": 2480699,"Sequence number": 24740542, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8402
    }
  },
  {
    "ph": "s", "id": 106, "pid": 1336759, "tid": 1336759, "ts": 1514656498288.245,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498293.797, "dur": 44.917,
    "args": {
      "External id": 2480700,"Sequence number": 24740543, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[65536, 2048], [2048, 5632]], "Ev Idx": 8403
    }
  },
  {
    "ph": "s", "id": 105, "pid": 1336759, "tid": 1336759, "ts": 1514656498293.797,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498340.471, "dur": 3.805,
    "args": {
      "External id": 2480701,"Sequence number": 24740544, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 1], []], "Input Dims": [[65536, 5632], []], "Ev Idx": 8404
    }
  },
  {
    "ph": "s", "id": 104, "pid": 1336759, "tid": 1336759, "ts": 1514656498340.471,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SwiGLULinearFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498366.268, "dur": 136.757,
    "args": {
      "External id": 2480702,"Sequence number": 24740545, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[23068672, 5632, 1], [23068672, 5632, 1], [5632, 1], []], "Input Dims": [[16, 4096, 5632], [16, 4096, 5632], [2048, 5632], []], "Ev Idx": 8405
    }
  },
  {
    "ph": "s", "id": 103, "pid": 1336759, "tid": 1336759, "ts": 1514656498366.268,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498401.760, "dur": 3.965,
    "args": {
      "External id": 2480703,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 5632]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498435.449, "dur": 56.106,
    "args": {
      "External id": 2480704,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[23068672, 5632, 1], [5632, 1], []], "Input Dims": [[16, 4096, 5632], [2048, 5632], []], "Ev Idx": 8407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498436.108, "dur": 4.525,
    "args": {
      "External id": 2480705,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 8408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498437.261, "dur": 2.583,
    "args": {
      "External id": 2480706,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[5632, 1], [], []], "Input Dims": [[2048, 5632], [], []], "Ev Idx": 8409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498438.647, "dur": 1.028,
    "args": {
      "External id": 2480707,"Record function id": 0, "Concrete Inputs": ["", "[5632, 2048]", "[1, 5632]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[5632, 1], [], [], []], "Input Dims": [[2048, 5632], [], [], []], "Ev Idx": 8410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498442.956, "dur": 48.321,
    "args": {
      "External id": 2480708,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[23068672, 5632, 1], [1, 5632]], "Input Dims": [[16, 4096, 5632], [5632, 2048]], "Ev Idx": 8411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498444.096, "dur": 2.208,
    "args": {
      "External id": 2480709,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 8412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498444.879, "dur": 1.305,
    "args": {
      "External id": 2480710,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 8413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498446.839, "dur": 40.643,
    "args": {
      "External id": 2480711,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632]], "Input Dims": [[65536, 5632], [5632, 2048]], "Ev Idx": 8414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498489.702, "dur": 0.933,
    "args": {
      "External id": 2480712,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498510.338, "dur": 41.005,
    "args": {
      "External id": 2480713,"Sequence number": 24740546, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 8416
    }
  },
  {
    "ph": "s", "id": 102, "pid": 1336759, "tid": 1336759, "ts": 1514656498510.338,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498587.582, "dur": 172.866,
    "args": {
      "External id": 2480714,"Sequence number": 24740547, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "9.9999999999999995e-07", "False", "False", "True"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 1], [1], [], [], [], [], [], []], "Input Dims": [[16, 4096, 2048], [2048], [], [], [], [], [], []], "Ev Idx": 8417
    }
  },
  {
    "ph": "s", "id": 101, "pid": 1336759, "tid": 1336759, "ts": 1514656498587.582,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498608.493, "dur": 3.678,
    "args": {
      "External id": 2480715,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498609.507, "dur": 2.482,
    "args": {
      "External id": 2480716,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498619.823, "dur": 7.004,
    "args": {
      "External id": 2480717,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], [], []], "Ev Idx": 8420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498622.075, "dur": 4.327,
    "args": {
      "External id": 2480718,"Record function id": 0, "Concrete Inputs": ["[65536, 2048]", "[2048, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498632.322, "dur": 3.694,
    "args": {
      "External id": 2480719,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498744.882, "dur": 3.027,
    "args": {
      "External id": 2480720,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498745.744, "dur": 1.992,
    "args": {
      "External id": 2480721,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498778.523, "dur": 82.147,
    "args": {
      "External id": 2480722,"Sequence number": 24740548, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [2048, 2048], []], "Ev Idx": 8425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498779.281, "dur": 5.818,
    "args": {
      "External id": 2480723,"Sequence number": 24740548, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 8426
    }
  },
  {
    "ph": "s", "id": 100, "pid": 1336759, "tid": 1336759, "ts": 1514656498779.281,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498780.985, "dur": 3.156,
    "args": {
      "External id": 2480724,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 8427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498782.820, "dur": 1.171,
    "args": {
      "External id": 2480725,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 8428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498785.981, "dur": 74.456,
    "args": {
      "External id": 2480726,"Sequence number": 24740549, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1, 2048]], "Input Dims": [[16, 4096, 2048], [2048, 2048]], "Ev Idx": 8429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498789.265, "dur": 3.635,
    "args": {
      "External id": 2480727,"Sequence number": 24740549, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498790.002, "dur": 2.767,
    "args": {
      "External id": 2480728,"Sequence number": 24740549, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8431
    }
  },
  {
    "ph": "s", "id": 99, "pid": 1336759, "tid": 1336759, "ts": 1514656498790.002,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498793.890, "dur": 59.564,
    "args": {
      "External id": 2480729,"Sequence number": 24740550, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 8432
    }
  },
  {
    "ph": "s", "id": 98, "pid": 1336759, "tid": 1336759, "ts": 1514656498793.890,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498855.557, "dur": 4.298,
    "args": {
      "External id": 2480730,"Sequence number": 24740551, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8433
    }
  },
  {
    "ph": "s", "id": 97, "pid": 1336759, "tid": 1336759, "ts": 1514656498855.557,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498867.840, "dur": 58.265,
    "args": {
      "External id": 2480731,"Sequence number": 24740552, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [2048, 2048], []], "Ev Idx": 8434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498870.270, "dur": 5.006,
    "args": {
      "External id": 2480732,"Sequence number": 24740552, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 8435
    }
  },
  {
    "ph": "s", "id": 96, "pid": 1336759, "tid": 1336759, "ts": 1514656498870.270,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498871.888, "dur": 2.395,
    "args": {
      "External id": 2480733,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 8436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498873.384, "dur": 0.709,
    "args": {
      "External id": 2480734,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 8437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498875.722, "dur": 50.192,
    "args": {
      "External id": 2480735,"Sequence number": 24740553, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1, 2048]], "Input Dims": [[16, 4096, 2048], [2048, 2048]], "Ev Idx": 8438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498877.057, "dur": 4.871,
    "args": {
      "External id": 2480736,"Sequence number": 24740553, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498879.444, "dur": 2.352,
    "args": {
      "External id": 2480737,"Sequence number": 24740553, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8440
    }
  },
  {
    "ph": "s", "id": 95, "pid": 1336759, "tid": 1336759, "ts": 1514656498879.444,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498882.582, "dur": 39.639,
    "args": {
      "External id": 2480738,"Sequence number": 24740554, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 8441
    }
  },
  {
    "ph": "s", "id": 94, "pid": 1336759, "tid": 1336759, "ts": 1514656498882.582,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498923.685, "dur": 1.912,
    "args": {
      "External id": 2480739,"Sequence number": 24740555, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8442
    }
  },
  {
    "ph": "s", "id": 93, "pid": 1336759, "tid": 1336759, "ts": 1514656498923.685,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498933.497, "dur": 97.393,
    "args": {
      "External id": 2480740,"Sequence number": 24740556, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [2048, 2048], []], "Ev Idx": 8443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498934.222, "dur": 8.025,
    "args": {
      "External id": 2480741,"Sequence number": 24740556, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 8444
    }
  },
  {
    "ph": "s", "id": 92, "pid": 1336759, "tid": 1336759, "ts": 1514656498934.222,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498937.459, "dur": 3.731,
    "args": {
      "External id": 2480742,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 8445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498938.600, "dur": 2.366,
    "args": {
      "External id": 2480743,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 8446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498942.864, "dur": 87.799,
    "args": {
      "External id": 2480744,"Sequence number": 24740557, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1, 2048]], "Input Dims": [[16, 4096, 2048], [2048, 2048]], "Ev Idx": 8447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498943.474, "dur": 6.203,
    "args": {
      "External id": 2480745,"Sequence number": 24740557, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498946.338, "dur": 3.096,
    "args": {
      "External id": 2480746,"Sequence number": 24740557, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8449
    }
  },
  {
    "ph": "s", "id": 91, "pid": 1336759, "tid": 1336759, "ts": 1514656498946.338,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656498950.359, "dur": 72.570,
    "args": {
      "External id": 2480747,"Sequence number": 24740558, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 8450
    }
  },
  {
    "ph": "s", "id": 90, "pid": 1336759, "tid": 1336759, "ts": 1514656498950.359,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499025.965, "dur": 4.069,
    "args": {
      "External id": 2480748,"Sequence number": 24740559, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8451
    }
  },
  {
    "ph": "s", "id": 89, "pid": 1336759, "tid": 1336759, "ts": 1514656499025.965,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499048.998, "dur": 5.771,
    "args": {
      "External id": 2480749,"Sequence number": 24740560, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499049.905, "dur": 4.729,
    "args": {
      "External id": 2480750,"Sequence number": 24740560, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8453
    }
  },
  {
    "ph": "s", "id": 88, "pid": 1336759, "tid": 1336759, "ts": 1514656499049.905,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499061.133, "dur": 3.063,
    "args": {
      "External id": 2480751,"Sequence number": 24740561, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499062.161, "dur": 1.913,
    "args": {
      "External id": 2480752,"Sequence number": 24740561, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8455
    }
  },
  {
    "ph": "s", "id": 87, "pid": 1336759, "tid": 1336759, "ts": 1514656499062.161,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499068.193, "dur": 3.125,
    "args": {
      "External id": 2480753,"Sequence number": 24740562, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499069.340, "dur": 1.860,
    "args": {
      "External id": 2480754,"Sequence number": 24740562, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8457
    }
  },
  {
    "ph": "s", "id": 86, "pid": 1336759, "tid": 1336759, "ts": 1514656499069.340,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499098.979, "dur": 144.313,
    "args": {
      "External id": 2480755,"Sequence number": 24740563, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "False", "False", "0", "", "8192"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [], [], [], [], []], "Ev Idx": 8458
    }
  },
  {
    "ph": "s", "id": 85, "pid": 1336759, "tid": 1336759, "ts": 1514656499098.979,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499117.938, "dur": 10.098,
    "args": {
      "External id": 2480756,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 8459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499120.378, "dur": 7.175,
    "args": {
      "External id": 2480757,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499253.877, "dur": 95.134,
    "args": {
      "External id": 2480758,"Sequence number": 24740564, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "False", "False", "0", "", "8192"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [], [], [], [], []], "Ev Idx": 8461
    }
  },
  {
    "ph": "s", "id": 84, "pid": 1336759, "tid": 1336759, "ts": 1514656499253.877,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499266.910, "dur": 6.511,
    "args": {
      "External id": 2480759,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 8462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499268.924, "dur": 4.166,
    "args": {
      "External id": 2480760,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FlashAttnFunc", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499374.316, "dur": 182.910,
    "args": {
      "External id": 2480761,"Sequence number": 24740565, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "", "True", "", "0.", "", "False", "False", "True"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "", "Scalar", "", "Scalar", "", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], []], "Ev Idx": 8464
    }
  },
  {
    "ph": "s", "id": 83, "pid": 1336759, "tid": 1336759, "ts": 1514656499374.316,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499398.398, "dur": 114.762,
    "args": {
      "External id": 2480762,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 8465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499442.198, "dur": 6.826,
    "args": {
      "External id": 2480763,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 8466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499444.729, "dur": 3.833,
    "args": {
      "External id": 2480764,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499451.395, "dur": 4.353,
    "args": {
      "External id": 2480765,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499456.728, "dur": 1.386,
    "args": {
      "External id": 2480766,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499460.786, "dur": 2.820,
    "args": {
      "External id": 2480767,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499523.611, "dur": 21.692,
    "args": {
      "External id": 2480768,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 8471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499563.389, "dur": 5.761,
    "args": {
      "External id": 2480769,"Sequence number": 24740566, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 8472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499564.757, "dur": 4.253,
    "args": {
      "External id": 2480770,"Sequence number": 24740566, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 8473
    }
  },
  {
    "ph": "s", "id": 82, "pid": 1336759, "tid": 1336759, "ts": 1514656499564.757,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499581.638, "dur": 110.967,
    "args": {
      "External id": 2480771,"Sequence number": 24740567, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [2048, 2048], []], "Ev Idx": 8474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499582.826, "dur": 12.529,
    "args": {
      "External id": 2480772,"Sequence number": 24740567, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 8475
    }
  },
  {
    "ph": "s", "id": 81, "pid": 1336759, "tid": 1336759, "ts": 1514656499582.826,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499587.797, "dur": 6.420,
    "args": {
      "External id": 2480773,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 8476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499592.168, "dur": 1.794,
    "args": {
      "External id": 2480774,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 8477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499596.497, "dur": 95.698,
    "args": {
      "External id": 2480775,"Sequence number": 24740568, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1, 2048]], "Input Dims": [[16, 4096, 2048], [2048, 2048]], "Ev Idx": 8478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499598.408, "dur": 2.842,
    "args": {
      "External id": 2480776,"Sequence number": 24740568, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499599.004, "dur": 2.118,
    "args": {
      "External id": 2480777,"Sequence number": 24740568, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8480
    }
  },
  {
    "ph": "s", "id": 80, "pid": 1336759, "tid": 1336759, "ts": 1514656499599.004,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499602.264, "dur": 81.771,
    "args": {
      "External id": 2480778,"Sequence number": 24740569, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 8481
    }
  },
  {
    "ph": "s", "id": 79, "pid": 1336759, "tid": 1336759, "ts": 1514656499602.264,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499686.521, "dur": 4.969,
    "args": {
      "External id": 2480779,"Sequence number": 24740570, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8482
    }
  },
  {
    "ph": "s", "id": 78, "pid": 1336759, "tid": 1336759, "ts": 1514656499686.521,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499725.599, "dur": 201.083,
    "args": {
      "External id": 2480780,"Sequence number": 24740571, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "9.9999999999999995e-07", "True", "False", "True"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 1], [1], [], [8388608, 2048, 1], [], [], [], []], "Input Dims": [[16, 4096, 2048], [2048], [], [16, 4096, 2048], [], [], [], []], "Ev Idx": 8483
    }
  },
  {
    "ph": "s", "id": 77, "pid": 1336759, "tid": 1336759, "ts": 1514656499725.599,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499744.202, "dur": 2.338,
    "args": {
      "External id": 2480781,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499744.898, "dur": 1.493,
    "args": {
      "External id": 2480782,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499750.334, "dur": 5.102,
    "args": {
      "External id": 2480783,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [2048, 1]], "Input Dims": [[16, 4096, 2048], [65536, 2048]], "Ev Idx": 8486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499753.173, "dur": 2.152,
    "args": {
      "External id": 2480784,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499754.179, "dur": 1.043,
    "args": {
      "External id": 2480785,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499762.568, "dur": 7.015,
    "args": {
      "External id": 2480786,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], [], []], "Ev Idx": 8489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499764.815, "dur": 4.486,
    "args": {
      "External id": 2480787,"Record function id": 0, "Concrete Inputs": ["[65536, 2048]", "[2048, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499775.534, "dur": 2.733,
    "args": {
      "External id": 2480788,"Record function id": 0, "Concrete Inputs": ["[65536, 2048]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499781.699, "dur": 3.011,
    "args": {
      "External id": 2480789,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499907.259, "dur": 3.289,
    "args": {
      "External id": 2480790,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499908.285, "dur": 2.077,
    "args": {
      "External id": 2480791,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499912.704, "dur": 1.879,
    "args": {
      "External id": 2480792,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499913.614, "dur": 0.861,
    "args": {
      "External id": 2480793,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499943.755, "dur": 142.362,
    "args": {
      "External id": 2480794,"Sequence number": 24740572, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [5632, 2048], []], "Ev Idx": 8497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499951.706, "dur": 8.431,
    "args": {
      "External id": 2480795,"Sequence number": 24740572, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 8498
    }
  },
  {
    "ph": "s", "id": 76, "pid": 1336759, "tid": 1336759, "ts": 1514656499951.706,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499955.451, "dur": 3.612,
    "args": {
      "External id": 2480796,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[5632, 2048], [], []], "Ev Idx": 8499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499957.550, "dur": 1.272,
    "args": {
      "External id": 2480797,"Record function id": 0, "Concrete Inputs": ["", "[2048, 5632]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[5632, 2048], [], [], []], "Ev Idx": 8500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499961.122, "dur": 124.662,
    "args": {
      "External id": 2480798,"Sequence number": 24740573, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1, 2048]], "Input Dims": [[16, 4096, 2048], [2048, 5632]], "Ev Idx": 8501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499962.471, "dur": 5.668,
    "args": {
      "External id": 2480799,"Sequence number": 24740573, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499963.148, "dur": 4.837,
    "args": {
      "External id": 2480800,"Sequence number": 24740573, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8503
    }
  },
  {
    "ph": "s", "id": 75, "pid": 1336759, "tid": 1336759, "ts": 1514656499963.148,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656499968.739, "dur": 105.739,
    "args": {
      "External id": 2480801,"Sequence number": 24740574, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[65536, 2048], [2048, 5632]], "Ev Idx": 8504
    }
  },
  {
    "ph": "s", "id": 74, "pid": 1336759, "tid": 1336759, "ts": 1514656499968.739,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500078.301, "dur": 6.821,
    "args": {
      "External id": 2480802,"Sequence number": 24740575, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 1], []], "Input Dims": [[65536, 5632], []], "Ev Idx": 8505
    }
  },
  {
    "ph": "s", "id": 73, "pid": 1336759, "tid": 1336759, "ts": 1514656500078.301,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500096.206, "dur": 73.277,
    "args": {
      "External id": 2480803,"Sequence number": 24740576, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [5632, 2048], []], "Ev Idx": 8506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500096.981, "dur": 7.295,
    "args": {
      "External id": 2480804,"Sequence number": 24740576, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 8507
    }
  },
  {
    "ph": "s", "id": 72, "pid": 1336759, "tid": 1336759, "ts": 1514656500096.981,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500098.463, "dur": 4.550,
    "args": {
      "External id": 2480805,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[5632, 2048], [], []], "Ev Idx": 8508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500101.808, "dur": 1.035,
    "args": {
      "External id": 2480806,"Record function id": 0, "Concrete Inputs": ["", "[2048, 5632]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[5632, 2048], [], [], []], "Ev Idx": 8509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500104.878, "dur": 64.311,
    "args": {
      "External id": 2480807,"Sequence number": 24740577, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1, 2048]], "Input Dims": [[16, 4096, 2048], [2048, 5632]], "Ev Idx": 8510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500105.959, "dur": 3.901,
    "args": {
      "External id": 2480808,"Sequence number": 24740577, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500106.541, "dur": 3.187,
    "args": {
      "External id": 2480809,"Sequence number": 24740577, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8512
    }
  },
  {
    "ph": "s", "id": 71, "pid": 1336759, "tid": 1336759, "ts": 1514656500106.541,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500112.109, "dur": 52.864,
    "args": {
      "External id": 2480810,"Sequence number": 24740578, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[65536, 2048], [2048, 5632]], "Ev Idx": 8513
    }
  },
  {
    "ph": "s", "id": 70, "pid": 1336759, "tid": 1336759, "ts": 1514656500112.109,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500166.557, "dur": 2.303,
    "args": {
      "External id": 2480811,"Sequence number": 24740579, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 1], []], "Input Dims": [[65536, 5632], []], "Ev Idx": 8514
    }
  },
  {
    "ph": "s", "id": 69, "pid": 1336759, "tid": 1336759, "ts": 1514656500166.557,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SwiGLULinearFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500189.986, "dur": 138.675,
    "args": {
      "External id": 2480812,"Sequence number": 24740580, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[23068672, 5632, 1], [23068672, 5632, 1], [5632, 1], []], "Input Dims": [[16, 4096, 5632], [16, 4096, 5632], [2048, 5632], []], "Ev Idx": 8515
    }
  },
  {
    "ph": "s", "id": 68, "pid": 1336759, "tid": 1336759, "ts": 1514656500189.986,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500225.418, "dur": 3.791,
    "args": {
      "External id": 2480813,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 5632]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500260.182, "dur": 55.849,
    "args": {
      "External id": 2480814,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[23068672, 5632, 1], [5632, 1], []], "Input Dims": [[16, 4096, 5632], [2048, 5632], []], "Ev Idx": 8517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500260.857, "dur": 4.963,
    "args": {
      "External id": 2480815,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 8518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500261.888, "dur": 3.256,
    "args": {
      "External id": 2480816,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[5632, 1], [], []], "Input Dims": [[2048, 5632], [], []], "Ev Idx": 8519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500264.191, "dur": 0.766,
    "args": {
      "External id": 2480817,"Record function id": 0, "Concrete Inputs": ["", "[5632, 2048]", "[1, 5632]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[5632, 1], [], [], []], "Input Dims": [[2048, 5632], [], [], []], "Ev Idx": 8520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500266.683, "dur": 49.051,
    "args": {
      "External id": 2480818,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[23068672, 5632, 1], [1, 5632]], "Input Dims": [[16, 4096, 5632], [5632, 2048]], "Ev Idx": 8521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500267.882, "dur": 2.799,
    "args": {
      "External id": 2480819,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 8522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500268.982, "dur": 1.586,
    "args": {
      "External id": 2480820,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 8523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500273.195, "dur": 39.221,
    "args": {
      "External id": 2480821,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632]], "Input Dims": [[65536, 5632], [5632, 2048]], "Ev Idx": 8524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500314.131, "dur": 1.099,
    "args": {
      "External id": 2480822,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500336.528, "dur": 23.062,
    "args": {
      "External id": 2480823,"Sequence number": 24740581, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 8526
    }
  },
  {
    "ph": "s", "id": 67, "pid": 1336759, "tid": 1336759, "ts": 1514656500336.528,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500391.066, "dur": 208.338,
    "args": {
      "External id": 2480824,"Sequence number": 24740582, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "9.9999999999999995e-07", "False", "False", "True"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 1], [1], [], [], [], [], [], []], "Input Dims": [[16, 4096, 2048], [2048], [], [], [], [], [], []], "Ev Idx": 8527
    }
  },
  {
    "ph": "s", "id": 66, "pid": 1336759, "tid": 1336759, "ts": 1514656500391.066,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500408.950, "dur": 2.922,
    "args": {
      "External id": 2480825,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500409.859, "dur": 1.827,
    "args": {
      "External id": 2480826,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500421.299, "dur": 10.515,
    "args": {
      "External id": 2480827,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], [], []], "Ev Idx": 8530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500427.205, "dur": 4.145,
    "args": {
      "External id": 2480828,"Record function id": 0, "Concrete Inputs": ["[65536, 2048]", "[2048, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500440.553, "dur": 5.260,
    "args": {
      "External id": 2480829,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500583.457, "dur": 3.978,
    "args": {
      "External id": 2480830,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500584.514, "dur": 2.637,
    "args": {
      "External id": 2480831,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500618.726, "dur": 83.697,
    "args": {
      "External id": 2480832,"Sequence number": 24740583, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [2048, 2048], []], "Ev Idx": 8535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500619.960, "dur": 8.064,
    "args": {
      "External id": 2480833,"Sequence number": 24740583, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 8536
    }
  },
  {
    "ph": "s", "id": 65, "pid": 1336759, "tid": 1336759, "ts": 1514656500619.960,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500622.067, "dur": 4.730,
    "args": {
      "External id": 2480834,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 8537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500625.276, "dur": 1.365,
    "args": {
      "External id": 2480835,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 8538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500628.879, "dur": 73.182,
    "args": {
      "External id": 2480836,"Sequence number": 24740584, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1, 2048]], "Input Dims": [[16, 4096, 2048], [2048, 2048]], "Ev Idx": 8539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500630.377, "dur": 3.270,
    "args": {
      "External id": 2480837,"Sequence number": 24740584, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500630.938, "dur": 2.579,
    "args": {
      "External id": 2480838,"Sequence number": 24740584, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8541
    }
  },
  {
    "ph": "s", "id": 64, "pid": 1336759, "tid": 1336759, "ts": 1514656500630.938,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500636.139, "dur": 60.818,
    "args": {
      "External id": 2480839,"Sequence number": 24740585, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 8542
    }
  },
  {
    "ph": "s", "id": 63, "pid": 1336759, "tid": 1336759, "ts": 1514656500636.139,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500698.506, "dur": 2.887,
    "args": {
      "External id": 2480840,"Sequence number": 24740586, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8543
    }
  },
  {
    "ph": "s", "id": 62, "pid": 1336759, "tid": 1336759, "ts": 1514656500698.506,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500710.470, "dur": 61.991,
    "args": {
      "External id": 2480841,"Sequence number": 24740587, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [2048, 2048], []], "Ev Idx": 8544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500711.066, "dur": 8.082,
    "args": {
      "External id": 2480842,"Sequence number": 24740587, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 8545
    }
  },
  {
    "ph": "s", "id": 61, "pid": 1336759, "tid": 1336759, "ts": 1514656500711.066,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500712.433, "dur": 5.630,
    "args": {
      "External id": 2480843,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 8546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500715.718, "dur": 2.166,
    "args": {
      "External id": 2480844,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 8547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500719.692, "dur": 52.459,
    "args": {
      "External id": 2480845,"Sequence number": 24740588, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1, 2048]], "Input Dims": [[16, 4096, 2048], [2048, 2048]], "Ev Idx": 8548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500720.631, "dur": 3.858,
    "args": {
      "External id": 2480846,"Sequence number": 24740588, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500721.662, "dur": 2.598,
    "args": {
      "External id": 2480847,"Sequence number": 24740588, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8550
    }
  },
  {
    "ph": "s", "id": 60, "pid": 1336759, "tid": 1336759, "ts": 1514656500721.662,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500725.286, "dur": 39.182,
    "args": {
      "External id": 2480848,"Sequence number": 24740589, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 8551
    }
  },
  {
    "ph": "s", "id": 59, "pid": 1336759, "tid": 1336759, "ts": 1514656500725.286,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500766.123, "dur": 5.660,
    "args": {
      "External id": 2480849,"Sequence number": 24740590, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8552
    }
  },
  {
    "ph": "s", "id": 58, "pid": 1336759, "tid": 1336759, "ts": 1514656500766.123,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500778.610, "dur": 56.132,
    "args": {
      "External id": 2480850,"Sequence number": 24740591, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [2048, 2048], []], "Ev Idx": 8553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500779.258, "dur": 4.617,
    "args": {
      "External id": 2480851,"Sequence number": 24740591, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 8554
    }
  },
  {
    "ph": "s", "id": 57, "pid": 1336759, "tid": 1336759, "ts": 1514656500779.258,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500780.423, "dur": 2.255,
    "args": {
      "External id": 2480852,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 8555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500781.616, "dur": 0.769,
    "args": {
      "External id": 2480853,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 8556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500786.603, "dur": 47.889,
    "args": {
      "External id": 2480854,"Sequence number": 24740592, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1, 2048]], "Input Dims": [[16, 4096, 2048], [2048, 2048]], "Ev Idx": 8557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500787.754, "dur": 3.552,
    "args": {
      "External id": 2480855,"Sequence number": 24740592, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500788.258, "dur": 2.918,
    "args": {
      "External id": 2480856,"Sequence number": 24740592, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8559
    }
  },
  {
    "ph": "s", "id": 56, "pid": 1336759, "tid": 1336759, "ts": 1514656500788.258,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500791.796, "dur": 37.438,
    "args": {
      "External id": 2480857,"Sequence number": 24740593, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 8560
    }
  },
  {
    "ph": "s", "id": 55, "pid": 1336759, "tid": 1336759, "ts": 1514656500791.796,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500830.665, "dur": 3.470,
    "args": {
      "External id": 2480858,"Sequence number": 24740594, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8561
    }
  },
  {
    "ph": "s", "id": 54, "pid": 1336759, "tid": 1336759, "ts": 1514656500830.665,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500850.658, "dur": 3.662,
    "args": {
      "External id": 2480859,"Sequence number": 24740595, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500851.713, "dur": 2.479,
    "args": {
      "External id": 2480860,"Sequence number": 24740595, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8563
    }
  },
  {
    "ph": "s", "id": 53, "pid": 1336759, "tid": 1336759, "ts": 1514656500851.713,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500860.613, "dur": 3.054,
    "args": {
      "External id": 2480861,"Sequence number": 24740596, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500861.436, "dur": 2.078,
    "args": {
      "External id": 2480862,"Sequence number": 24740596, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8565
    }
  },
  {
    "ph": "s", "id": 52, "pid": 1336759, "tid": 1336759, "ts": 1514656500861.436,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500869.004, "dur": 4.698,
    "args": {
      "External id": 2480863,"Sequence number": 24740597, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500869.868, "dur": 3.676,
    "args": {
      "External id": 2480864,"Sequence number": 24740597, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 32, 64]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8567
    }
  },
  {
    "ph": "s", "id": 51, "pid": 1336759, "tid": 1336759, "ts": 1514656500869.868,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500901.812, "dur": 189.537,
    "args": {
      "External id": 2480865,"Sequence number": 24740598, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "False", "False", "0", "", "8192"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [], [], [], [], []], "Ev Idx": 8568
    }
  },
  {
    "ph": "s", "id": 50, "pid": 1336759, "tid": 1336759, "ts": 1514656500901.812,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500919.744, "dur": 7.607,
    "args": {
      "External id": 2480866,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 8569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656500921.877, "dur": 5.010,
    "args": {
      "External id": 2480867,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "RotaryEmbeddingFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501105.615, "dur": 108.872,
    "args": {
      "External id": 2480868,"Sequence number": 24740599, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "False", "False", "0", "", "8192"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [32, 1], [32, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [8192, 32], [8192, 32], [], [], [], [], []], "Ev Idx": 8571
    }
  },
  {
    "ph": "s", "id": 49, "pid": 1336759, "tid": 1336759, "ts": 1514656501105.615,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501121.079, "dur": 8.110,
    "args": {
      "External id": 2480869,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 8572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501123.463, "dur": 5.281,
    "args": {
      "External id": 2480870,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FlashAttnFunc", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501240.505, "dur": 162.512,
    "args": {
      "External id": 2480871,"Sequence number": 24740600, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "", "True", "", "0.", "", "False", "False", "True"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "", "Scalar", "", "Scalar", "", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], [], []], "Ev Idx": 8574
    }
  },
  {
    "ph": "s", "id": 48, "pid": 1336759, "tid": 1336759, "ts": 1514656501240.505,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "flash_attn::_flash_attn_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501264.873, "dur": 115.087,
    "args": {
      "External id": 2480872,"Record function id": 0, "Concrete Inputs": ["", "", "", "0.", "0.125", "True", "-1", "-1", "0.", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [8388608, 2048, 64, 1], [], [], [], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [16, 4096, 32, 64], [16, 4096, 32, 64], [], [], [], [], [], [], [], []], "Ev Idx": 8575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501310.790, "dur": 6.724,
    "args": {
      "External id": 2480873,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", ""], "Input type": ["c10::BFloat16", "", "", "", "", ""], "Input Strides": [[8388608, 2048, 64, 1], [], [], [], [], []], "Input Dims": [[16, 4096, 32, 64], [], [], [], [], []], "Ev Idx": 8576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501312.745, "dur": 4.332,
    "args": {
      "External id": 2480874,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 32, 64]", "[8388608, 2048, 64, 1]", "15", "0", "", ""], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501320.057, "dur": 3.862,
    "args": {
      "External id": 2480875,"Record function id": 0, "Concrete Inputs": ["[16, 32, 4096]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501325.149, "dur": 2.800,
    "args": {
      "External id": 2480876,"Record function id": 0, "Concrete Inputs": ["[0]", "15", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501330.313, "dur": 3.294,
    "args": {
      "External id": 2480877,"Record function id": 0, "Concrete Inputs": ["[2]", "4", "", "", "", ""], "Input type": ["ScalarList", "Scalar", "", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::alias", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501390.423, "dur": 4.409,
    "args": {
      "External id": 2480878,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[8388608, 2048, 64, 1]], "Input Dims": [[16, 4096, 32, 64]], "Ev Idx": 8581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501408.029, "dur": 5.796,
    "args": {
      "External id": 2480879,"Sequence number": 24740601, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 8582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501409.618, "dur": 4.041,
    "args": {
      "External id": 2480880,"Sequence number": 24740601, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, -1]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 64, 1], []], "Input Dims": [[16, 4096, 32, 64], []], "Ev Idx": 8583
    }
  },
  {
    "ph": "s", "id": 47, "pid": 1336759, "tid": 1336759, "ts": 1514656501409.618,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501424.635, "dur": 95.542,
    "args": {
      "External id": 2480881,"Sequence number": 24740602, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [2048, 2048], []], "Ev Idx": 8584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501427.555, "dur": 6.667,
    "args": {
      "External id": 2480882,"Sequence number": 24740602, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[2048, 2048]], "Ev Idx": 8585
    }
  },
  {
    "ph": "s", "id": 46, "pid": 1336759, "tid": 1336759, "ts": 1514656501427.555,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501429.715, "dur": 3.633,
    "args": {
      "External id": 2480883,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[2048, 2048], [], []], "Ev Idx": 8586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501431.652, "dur": 1.466,
    "args": {
      "External id": 2480884,"Record function id": 0, "Concrete Inputs": ["", "[2048, 2048]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[2048, 2048], [], [], []], "Ev Idx": 8587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501435.532, "dur": 84.326,
    "args": {
      "External id": 2480885,"Sequence number": 24740603, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1, 2048]], "Input Dims": [[16, 4096, 2048], [2048, 2048]], "Ev Idx": 8588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501437.294, "dur": 4.534,
    "args": {
      "External id": 2480886,"Sequence number": 24740603, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501439.957, "dur": 1.741,
    "args": {
      "External id": 2480887,"Sequence number": 24740603, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8590
    }
  },
  {
    "ph": "s", "id": 45, "pid": 1336759, "tid": 1336759, "ts": 1514656501439.957,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501442.592, "dur": 71.016,
    "args": {
      "External id": 2480888,"Sequence number": 24740604, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[65536, 2048], [2048, 2048]], "Ev Idx": 8591
    }
  },
  {
    "ph": "s", "id": 44, "pid": 1336759, "tid": 1336759, "ts": 1514656501442.592,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501515.973, "dur": 3.267,
    "args": {
      "External id": 2480889,"Sequence number": 24740605, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8592
    }
  },
  {
    "ph": "s", "id": 43, "pid": 1336759, "tid": 1336759, "ts": 1514656501515.973,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "LayerNormFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501569.881, "dur": 207.918,
    "args": {
      "External id": 2480890,"Sequence number": 24740606, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "9.9999999999999995e-07", "True", "False", "True"], "Input type": ["c10::BFloat16", "c10::BFloat16", "", "c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8388608, 2048, 1], [1], [], [8388608, 2048, 1], [], [], [], []], "Input Dims": [[16, 4096, 2048], [2048], [], [16, 4096, 2048], [], [], [], []], "Ev Idx": 8593
    }
  },
  {
    "ph": "s", "id": 42, "pid": 1336759, "tid": 1336759, "ts": 1514656501569.881,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501590.874, "dur": 3.763,
    "args": {
      "External id": 2480891,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501592.083, "dur": 2.233,
    "args": {
      "External id": 2480892,"Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501598.429, "dur": 3.248,
    "args": {
      "External id": 2480893,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [2048, 1]], "Input Dims": [[16, 4096, 2048], [65536, 2048]], "Ev Idx": 8596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501599.953, "dur": 1.615,
    "args": {
      "External id": 2480894,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501600.388, "dur": 1.062,
    "args": {
      "External id": 2480895,"Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501608.710, "dur": 9.193,
    "args": {
      "External id": 2480896,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], [], []], "Ev Idx": 8599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501612.424, "dur": 5.191,
    "args": {
      "External id": 2480897,"Record function id": 0, "Concrete Inputs": ["[65536, 2048]", "[2048, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501624.034, "dur": 3.210,
    "args": {
      "External id": 2480898,"Record function id": 0, "Concrete Inputs": ["[65536, 2048]", "15", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501630.368, "dur": 3.966,
    "args": {
      "External id": 2480899,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501757.615, "dur": 3.111,
    "args": {
      "External id": 2480900,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501758.660, "dur": 1.888,
    "args": {
      "External id": 2480901,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501763.243, "dur": 1.633,
    "args": {
      "External id": 2480902,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501763.825, "dur": 0.944,
    "args": {
      "External id": 2480903,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501794.813, "dur": 91.002,
    "args": {
      "External id": 2480904,"Sequence number": 24740607, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [5632, 2048], []], "Ev Idx": 8607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501796.167, "dur": 8.768,
    "args": {
      "External id": 2480905,"Sequence number": 24740607, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 8608
    }
  },
  {
    "ph": "s", "id": 41, "pid": 1336759, "tid": 1336759, "ts": 1514656501796.167,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501798.820, "dur": 5.165,
    "args": {
      "External id": 2480906,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[5632, 2048], [], []], "Ev Idx": 8609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501800.856, "dur": 2.880,
    "args": {
      "External id": 2480907,"Record function id": 0, "Concrete Inputs": ["", "[2048, 5632]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[5632, 2048], [], [], []], "Ev Idx": 8610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501805.926, "dur": 79.520,
    "args": {
      "External id": 2480908,"Sequence number": 24740608, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1, 2048]], "Input Dims": [[16, 4096, 2048], [2048, 5632]], "Ev Idx": 8611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501807.657, "dur": 4.615,
    "args": {
      "External id": 2480909,"Sequence number": 24740608, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501809.926, "dur": 2.162,
    "args": {
      "External id": 2480910,"Sequence number": 24740608, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8613
    }
  },
  {
    "ph": "s", "id": 40, "pid": 1336759, "tid": 1336759, "ts": 1514656501809.926,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501813.125, "dur": 64.155,
    "args": {
      "External id": 2480911,"Sequence number": 24740609, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[65536, 2048], [2048, 5632]], "Ev Idx": 8614
    }
  },
  {
    "ph": "s", "id": 39, "pid": 1336759, "tid": 1336759, "ts": 1514656501813.125,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501879.493, "dur": 5.252,
    "args": {
      "External id": 2480912,"Sequence number": 24740610, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 1], []], "Input Dims": [[65536, 5632], []], "Ev Idx": 8615
    }
  },
  {
    "ph": "s", "id": 38, "pid": 1336759, "tid": 1336759, "ts": 1514656501879.493,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501893.425, "dur": 67.528,
    "args": {
      "External id": 2480913,"Sequence number": 24740611, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[8388608, 2048, 1], [2048, 1], []], "Input Dims": [[16, 4096, 2048], [5632, 2048], []], "Ev Idx": 8616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501894.035, "dur": 8.209,
    "args": {
      "External id": 2480914,"Sequence number": 24740611, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[5632, 2048]], "Ev Idx": 8617
    }
  },
  {
    "ph": "s", "id": 37, "pid": 1336759, "tid": 1336759, "ts": 1514656501894.035,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501898.807, "dur": 2.454,
    "args": {
      "External id": 2480915,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[5632, 2048], [], []], "Ev Idx": 8618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501900.295, "dur": 0.803,
    "args": {
      "External id": 2480916,"Record function id": 0, "Concrete Inputs": ["", "[2048, 5632]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[5632, 2048], [], [], []], "Ev Idx": 8619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501902.856, "dur": 57.826,
    "args": {
      "External id": 2480917,"Sequence number": 24740612, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[8388608, 2048, 1], [1, 2048]], "Input Dims": [[16, 4096, 2048], [2048, 5632]], "Ev Idx": 8620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501904.080, "dur": 5.867,
    "args": {
      "External id": 2480918,"Sequence number": 24740612, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501906.304, "dur": 3.518,
    "args": {
      "External id": 2480919,"Sequence number": 24740612, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[65536, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8622
    }
  },
  {
    "ph": "s", "id": 36, "pid": 1336759, "tid": 1336759, "ts": 1514656501906.304,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501910.711, "dur": 44.999,
    "args": {
      "External id": 2480920,"Sequence number": 24740613, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[65536, 2048], [2048, 5632]], "Ev Idx": 8623
    }
  },
  {
    "ph": "s", "id": 35, "pid": 1336759, "tid": 1336759, "ts": 1514656501910.711,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501957.762, "dur": 2.527,
    "args": {
      "External id": 2480921,"Sequence number": 24740614, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[5632, 1], []], "Input Dims": [[65536, 5632], []], "Ev Idx": 8624
    }
  },
  {
    "ph": "s", "id": 34, "pid": 1336759, "tid": 1336759, "ts": 1514656501957.762,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "SwiGLULinearFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656501979.424, "dur": 188.765,
    "args": {
      "External id": 2480922,"Sequence number": 24740615, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[23068672, 5632, 1], [23068672, 5632, 1], [5632, 1], []], "Input Dims": [[16, 4096, 5632], [16, 4096, 5632], [2048, 5632], []], "Ev Idx": 8625
    }
  },
  {
    "ph": "s", "id": 33, "pid": 1336759, "tid": 1336759, "ts": 1514656501979.424,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502054.720, "dur": 5.515,
    "args": {
      "External id": 2480923,"Record function id": 0, "Concrete Inputs": ["[16, 4096, 5632]", "15", "", "", "", "0"], "Input type": ["ScalarList", "Scalar", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502092.023, "dur": 64.066,
    "args": {
      "External id": 2480924,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[23068672, 5632, 1], [5632, 1], []], "Input Dims": [[16, 4096, 5632], [2048, 5632], []], "Ev Idx": 8627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502093.064, "dur": 7.261,
    "args": {
      "External id": 2480925,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[5632, 1]], "Input Dims": [[2048, 5632]], "Ev Idx": 8628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502094.648, "dur": 4.644,
    "args": {
      "External id": 2480926,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[5632, 1], [], []], "Input Dims": [[2048, 5632], [], []], "Ev Idx": 8629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502096.459, "dur": 2.613,
    "args": {
      "External id": 2480927,"Record function id": 0, "Concrete Inputs": ["", "[5632, 2048]", "[1, 5632]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[5632, 1], [], [], []], "Input Dims": [[2048, 5632], [], [], []], "Ev Idx": 8630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502101.186, "dur": 54.623,
    "args": {
      "External id": 2480928,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[23068672, 5632, 1], [1, 5632]], "Input Dims": [[16, 4096, 5632], [5632, 2048]], "Ev Idx": 8631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reshape", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502102.733, "dur": 4.261,
    "args": {
      "External id": 2480929,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 8632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502105.085, "dur": 1.667,
    "args": {
      "External id": 2480930,"Record function id": 0, "Concrete Inputs": ["", "[65536, 5632]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[23068672, 5632, 1], []], "Input Dims": [[16, 4096, 5632], []], "Ev Idx": 8633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502107.557, "dur": 44.218,
    "args": {
      "External id": 2480931,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[5632, 1], [1, 5632]], "Input Dims": [[65536, 5632], [5632, 2048]], "Ev Idx": 8634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_unsafe_view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502153.909, "dur": 1.214,
    "args": {
      "External id": 2480932,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502176.473, "dur": 23.172,
    "args": {
      "External id": 2480933,"Sequence number": 24740616, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[8388608, 2048, 1], [8388608, 2048, 1], []], "Input Dims": [[16, 4096, 2048], [16, 4096, 2048], []], "Ev Idx": 8636
    }
  },
  {
    "ph": "s", "id": 32, "pid": 1336759, "tid": 1336759, "ts": 1514656502176.473,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::stack", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502216.885, "dur": 38.102,
    "args": {
      "External id": 2480934,"Sequence number": 24740617, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "-2"], "Input type": ["TensorList", "Scalar"], "Input Strides": [[[8388608, 2048, 1], [8388608, 2048, 1], [8388608, 2048, 1], [8388608, 2048, 1]], []], "Input Dims": [[[16, 4096, 2048], [16, 4096, 2048], [16, 4096, 2048], [16, 4096, 2048]], []], "Ev Idx": 8637
    }
  },
  {
    "ph": "s", "id": 31, "pid": 1336759, "tid": 1336759, "ts": 1514656502216.885,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502225.599, "dur": 25.425,
    "args": {
      "External id": 2480935,"Record function id": 0, "Concrete Inputs": ["", "2"], "Input type": ["TensorList", "Scalar"], "Input Strides": [[[8388608, 2048, 1], [8388608, 2048, 1], [8388608, 2048, 1], [8388608, 2048, 1]], []], "Input Dims": [[[16, 4096, 2048], [16, 4096, 2048], [16, 4096, 2048], [16, 4096, 2048]], []], "Ev Idx": 8638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502252.489, "dur": 1.108,
    "args": {
      "External id": 2480936,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[33554432, 8192, 1], []], "Input Dims": [[16, 4096, 8192], []], "Ev Idx": 8639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "TorchDynamo Cache Lookup", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502288.603, "dur": 34.979,
    "args": {
      "External id": 2480937,"Record function id": 0, "Ev Idx": 8640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Torch-Compiled Region: 2/0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502324.619, "dur": 174.959,
    "args": {
      "External id": 2480938,"Record function id": 0, "Ev Idx": 8641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "CompiledFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502359.264, "dur": 132.335,
    "args": {
      "External id": 2480939,"Sequence number": 24740618, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1], [33554432, 8192, 2048, 1]], "Input Dims": [[2048], [16, 4096, 4, 2048]], "Ev Idx": 8642
    }
  },
  {
    "ph": "s", "id": 30, "pid": 1336759, "tid": 1336759, "ts": 1514656502359.264,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "layer_norm_fwd_kernel_0", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502423.169, "dur": 33.498,
    "args": {
      "External id": 2480940,"kernel_hash": "cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj", "Record function id": 0, "stream": 0, "Concrete Inputs": ["", "", "", "", "2048", "1", "9.9999999999999995e-07", "True", "2048", "False", "False", "True", "False"], "kernel_file": "/tmp/torchinductor_cvm/xq/cxq73k7hzeap6epyl22bkojjyhqsfjuzklwbe2qcgpfenip5hwvj.py", "kernel_backend": "triton", "Input type": ["c10::BFloat16", "c10::BFloat16", "c10::BFloat16", "float", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], [1], [1], [], [], [], [], [], [], [], [], []], "Input Dims": [[262144, 2048], [262144, 2048], [2048], [262144], [], [], [], [], [], [], [], [], []], "Ev Idx": 8643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502598.892, "dur": 39.945,
    "args": {
      "External id": 2480941,"Record function id": 0, "Concrete Inputs": ["[1]", "", "", "", "False"], "Input type": ["ScalarList", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 8644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502601.753, "dur": 6.286,
    "args": {
      "External id": 2480942,"Record function id": 0, "Concrete Inputs": ["[1]", "", "", "", "False", ""], "Input type": ["ScalarList", "", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502611.189, "dur": 27.346,
    "args": {
      "External id": 2480943,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 8646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502614.656, "dur": 23.082,
    "args": {
      "External id": 2480944,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[1], []], "Ev Idx": 8647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502643.458, "dur": 18.111,
    "args": {
      "External id": 2480945,"Record function id": 0, "Concrete Inputs": ["[1]", "", "", "", "False"], "Input type": ["ScalarList", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 8648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502644.618, "dur": 2.567,
    "args": {
      "External id": 2480946,"Record function id": 0, "Concrete Inputs": ["[1]", "", "", "", "False", ""], "Input type": ["ScalarList", "", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502649.785, "dur": 11.512,
    "args": {
      "External id": 2480947,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 8650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502650.479, "dur": 10.397,
    "args": {
      "External id": 2480948,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[1], []], "Ev Idx": 8651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502664.270, "dur": 14.506,
    "args": {
      "External id": 2480949,"Record function id": 0, "Concrete Inputs": ["[1]", "", "", "", "False"], "Input type": ["ScalarList", "", "", "", "Scalar"], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 8652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502664.928, "dur": 2.347,
    "args": {
      "External id": 2480950,"Record function id": 0, "Concrete Inputs": ["[1]", "", "", "", "False", ""], "Input type": ["ScalarList", "", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502668.067, "dur": 10.446,
    "args": {
      "External id": 2480951,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[1]], "Ev Idx": 8654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502668.683, "dur": 9.320,
    "args": {
      "External id": 2480952,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[1], []], "Ev Idx": 8655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502688.649, "dur": 0.592,
    "args": {
      "External id": 2480953,"Sequence number": 24740619, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "4", "0", "", "", "False", "False", ""], "Input type": ["long int", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[8192, 1], [], [], [], [], [], [], []], "Input Dims": [[16, 8192], [], [], [], [], [], [], []], "Ev Idx": 8656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unfold", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502696.822, "dur": 10.158,
    "args": {
      "External id": 2480954,"Sequence number": 24740619, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "5", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], [], []], "Input Dims": [[16, 8192], [], [], []], "Ev Idx": 8657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502703.229, "dur": 1.950,
    "args": {
      "External id": 2480955,"Record function id": 0, "Concrete Inputs": ["", "[16, 8188, 5]", "[8192, 1, 1]", ""], "Input type": ["long int", "ScalarList", "ScalarList", ""], "Input Strides": [[8192, 1], [], [], []], "Input Dims": [[16, 8192], [], [], []], "Ev Idx": 8658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502713.323, "dur": 7.025,
    "args": {
      "External id": 2480956,"Sequence number": 24740619, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], [], []], "Input Dims": [[16, 8188, 5], [], [], [], []], "Ev Idx": 8659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502717.315, "dur": 1.340,
    "args": {
      "External id": 2480957,"Record function id": 0, "Concrete Inputs": ["", "[16, 8188, 5]", "[8192, 1, 1]", "0"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[16, 8188, 5], [], [], []], "Ev Idx": 8660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502721.579, "dur": 3.299,
    "args": {
      "External id": 2480958,"Sequence number": 24740619, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], [], []], "Input Dims": [[16, 8188, 5], [], [], [], []], "Ev Idx": 8661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502723.350, "dur": 0.775,
    "args": {
      "External id": 2480959,"Record function id": 0, "Concrete Inputs": ["", "[16, 8188, 5]", "[8192, 1, 1]", "0"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[16, 8188, 5], [], [], []], "Ev Idx": 8662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502726.021, "dur": 2.876,
    "args": {
      "External id": 2480960,"Sequence number": 24740619, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2", "1", "9223372036854775807", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], [], []], "Input Dims": [[16, 8188, 5], [], [], [], []], "Ev Idx": 8663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502727.617, "dur": 0.760,
    "args": {
      "External id": 2480961,"Record function id": 0, "Concrete Inputs": ["", "[16, 8188, 4]", "[8192, 1, 1]", "1"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[16, 8188, 5], [], [], []], "Ev Idx": 8664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502731.551, "dur": 2.675,
    "args": {
      "External id": 2480962,"Sequence number": 24740619, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], [], []], "Input Dims": [[16, 8188, 4], [], [], [], []], "Ev Idx": 8665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502733.020, "dur": 0.640,
    "args": {
      "External id": 2480963,"Record function id": 0, "Concrete Inputs": ["", "[16, 8188, 4]", "[8192, 1, 1]", "1"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[16, 8188, 4], [], [], []], "Ev Idx": 8666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502735.120, "dur": 3.230,
    "args": {
      "External id": 2480964,"Sequence number": 24740619, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "4096", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], [], []], "Input Dims": [[16, 8188, 4], [], [], [], []], "Ev Idx": 8667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502736.686, "dur": 0.969,
    "args": {
      "External id": 2480965,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4]", "[8192, 1, 1]", "1"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[16, 8188, 4], [], [], []], "Ev Idx": 8668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502739.337, "dur": 2.669,
    "args": {
      "External id": 2480966,"Sequence number": 24740619, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2", "0", "9223372036854775807", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], [], []], "Input Dims": [[16, 4096, 4], [], [], [], []], "Ev Idx": 8669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502740.937, "dur": 0.538,
    "args": {
      "External id": 2480967,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4]", "[8192, 1, 1]", "1"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[16, 4096, 4], [], [], []], "Ev Idx": 8670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502745.777, "dur": 5.318,
    "args": {
      "External id": 2480968,"Sequence number": 24740619, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "2"], "Input type": ["long int", "Scalar", "Scalar"], "Input Strides": [[8192, 1, 1], [], []], "Input Dims": [[16, 4096, 4], [], []], "Ev Idx": 8671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502749.479, "dur": 0.831,
    "args": {
      "External id": 2480969,"Record function id": 0, "Concrete Inputs": ["", "[16, 4, 4096]", "[8192, 1, 1]", ""], "Input type": ["long int", "ScalarList", "ScalarList", ""], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[16, 4096, 4], [], [], []], "Ev Idx": 8672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502754.899, "dur": 3.019,
    "args": {
      "External id": 2480970,"Sequence number": 24740619, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], [], []], "Input Dims": [[16, 4, 4096], [], [], [], []], "Ev Idx": 8673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502756.708, "dur": 0.627,
    "args": {
      "External id": 2480971,"Record function id": 0, "Concrete Inputs": ["", "[16, 4, 4096]", "[8192, 1, 1]", "1"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[16, 4, 4096], [], [], []], "Ev Idx": 8674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502763.208, "dur": 7.119,
    "args": {
      "External id": 2480972,"Sequence number": 24740619, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0"], "Input type": ["long int", "Scalar", "Scalar"], "Input Strides": [[8192, 1, 1], [], []], "Input Dims": [[16, 4, 4096], [], []], "Ev Idx": 8675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502768.400, "dur": 0.725,
    "args": {
      "External id": 2480973,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096]", "[8192, 1]", "1"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[16, 4, 4096], [], [], []], "Ev Idx": 8676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502771.467, "dur": 2.534,
    "args": {
      "External id": 2480974,"Sequence number": 24740619, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], [], [], []], "Input Dims": [[16, 4096], [], [], [], []], "Ev Idx": 8677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502772.886, "dur": 0.523,
    "args": {
      "External id": 2480975,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096]", "[8192, 1]", "1"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1], [], [], []], "Input Dims": [[16, 4096], [], [], []], "Ev Idx": 8678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502776.338, "dur": 7.750,
    "args": {
      "External id": 2480976,"Sequence number": 24740619, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], [], []], "Ev Idx": 8679
    }
  },
  {
    "ph": "s", "id": 29, "pid": 1336759, "tid": 1336759, "ts": 1514656502776.338,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502781.733, "dur": 0.911,
    "args": {
      "External id": 2480977,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 2048]", "[33554432, 8192, 2048, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], []], "Ev Idx": 8680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502785.123, "dur": 4.616,
    "args": {
      "External id": 2480978,"Sequence number": 24740620, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], [], []], "Ev Idx": 8681
    }
  },
  {
    "ph": "s", "id": 28, "pid": 1336759, "tid": 1336759, "ts": 1514656502785.123,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502788.258, "dur": 0.803,
    "args": {
      "External id": 2480979,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 2048]", "[33554432, 8192, 2048, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], []], "Ev Idx": 8682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502790.690, "dur": 6.231,
    "args": {
      "External id": 2480980,"Sequence number": 24740621, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2", "0"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], []], "Input Dims": [[16, 4096, 4, 2048], [], []], "Ev Idx": 8683
    }
  },
  {
    "ph": "s", "id": 27, "pid": 1336759, "tid": 1336759, "ts": 1514656502790.690,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502795.405, "dur": 0.689,
    "args": {
      "External id": 2480981,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]", "[33554432, 8192, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], []], "Ev Idx": 8684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502797.993, "dur": 5.584,
    "args": {
      "External id": 2480982,"Sequence number": 24740622, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 1], [], [], [], []], "Input Dims": [[16, 4096, 2048], [], [], [], []], "Ev Idx": 8685
    }
  },
  {
    "ph": "s", "id": 26, "pid": 1336759, "tid": 1336759, "ts": 1514656502797.993,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502800.564, "dur": 2.206,
    "args": {
      "External id": 2480983,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]", "[33554432, 8192, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 1], [], [], []], "Input Dims": [[16, 4096, 2048], [], [], []], "Ev Idx": 8686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502807.643, "dur": 33.022,
    "args": {
      "External id": 2480984,"Sequence number": 24740623, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["long int", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[16, 4096], []], "Ev Idx": 8687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502809.650, "dur": 30.812,
    "args": {
      "External id": 2480985,"Sequence number": 24740623, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["long int", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[16, 4096], []], "Ev Idx": 8688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502812.157, "dur": 5.435,
    "args": {
      "External id": 2480986,"Record function id": 0, "Concrete Inputs": ["", "4", "0", "", "", "0"], "Input type": ["long int", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[8192, 1], [], [], [], [], []], "Input Dims": [[16, 4096], [], [], [], [], []], "Ev Idx": 8689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502814.205, "dur": 2.856,
    "args": {
      "External id": 2480987,"Record function id": 0, "Concrete Inputs": ["[16, 4096]", "4", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502818.840, "dur": 21.136,
    "args": {
      "External id": 2480988,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[4096, 1], [8192, 1], []], "Input Dims": [[16, 4096], [16, 4096], []], "Ev Idx": 8691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502863.219, "dur": 4.295,
    "args": {
      "External id": 2480989,"Sequence number": 24740623, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[33554432, 8192, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 8692
    }
  },
  {
    "ph": "s", "id": 25, "pid": 1336759, "tid": 1336759, "ts": 1514656502863.219,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502869.770, "dur": 1.299,
    "args": {
      "External id": 2480990,"Sequence number": 24740624, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["long int", "ScalarList"], "Input Strides": [[4096, 1], []], "Input Dims": [[16, 4096], []], "Ev Idx": 8693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FusedLinearCrossEntropyFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502897.746, "dur": 73306.950,
    "args": {
      "External id": 2480991,"Sequence number": 24740624, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "-100", "0.", "1.", "8"], "Input type": ["c10::BFloat16", "long int", "c10::BFloat16", "", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [1], [2048, 1], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536], [32000, 2048], [], [], [], [], []], "Ev Idx": 8694
    }
  },
  {
    "ph": "s", "id": 24, "pid": 1336759, "tid": 1336759, "ts": 1514656502897.746,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502910.296, "dur": 25.852,
    "args": {
      "External id": 2480992,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502911.014, "dur": 24.915,
    "args": {
      "External id": 2480993,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502912.259, "dur": 6.119,
    "args": {
      "External id": 2480994,"Record function id": 0, "Concrete Inputs": ["", "15", "0", "", "", "0"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[8192, 1], [], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], [], []], "Ev Idx": 8697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502913.689, "dur": 4.285,
    "args": {
      "External id": 2480995,"Record function id": 0, "Concrete Inputs": ["[65536, 2048]", "15", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502919.451, "dur": 16.106,
    "args": {
      "External id": 2480996,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [8192, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 8699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502954.283, "dur": 23.312,
    "args": {
      "External id": 2480997,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], [], []], "Ev Idx": 8700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502955.407, "dur": 5.676,
    "args": {
      "External id": 2480998,"Record function id": 0, "Concrete Inputs": ["", "15", "0", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], [], []], "Ev Idx": 8701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502957.143, "dur": 3.590,
    "args": {
      "External id": 2480999,"Record function id": 0, "Concrete Inputs": ["[65536, 2048]", "[2048, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502961.888, "dur": 15.490,
    "args": {
      "External id": 2481000,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 8703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502962.906, "dur": 13.929,
    "args": {
      "External id": 2481001,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 8704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502981.198, "dur": 69.096,
    "args": {
      "External id": 2481002,"Record function id": 0, "Concrete Inputs": ["", "6", "", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[32000, 2048], [], [], [], [], []], "Ev Idx": 8705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656502982.082, "dur": 45.745,
    "args": {
      "External id": 2481003,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[32000, 2048], [], [], [], [], []], "Ev Idx": 8706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656503023.086, "dur": 4.220,
    "args": {
      "External id": 2481004,"Record function id": 0, "Concrete Inputs": ["[32000, 2048]", "[2048, 1]", "6", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656503036.272, "dur": 13.827,
    "args": {
      "External id": 2481005,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 8708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656503036.989, "dur": 12.729,
    "args": {
      "External id": 2481006,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[2048, 1], []], "Input Dims": [[32000, 2048], []], "Ev Idx": 8709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 1336759, "tid": 1336759,
    "ts": 1514656503056.651, "dur": 21.122,
    "args": {
      "External id": 2481007,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 8710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656503061.507, "dur": 4.579,
    "args": {
      "External id": 2481008,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656503066.746, "dur": 10.740,
    "args": {
      "External id": 2481009,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[65536]], "Ev Idx": 8712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656503067.347, "dur": 9.575,
    "args": {
      "External id": 2481010,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 8713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::ne", "pid": 1336759, "tid": 1336759,
    "ts": 1514656503087.002, "dur": 24.989,
    "args": {
      "External id": 2481011,"Record function id": 0, "Concrete Inputs": ["", "-100"], "Input type": ["long int", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 8714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656503116.068, "dur": 60.074,
    "args": {
      "External id": 2481012,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["bool", ""], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 8715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656503119.637, "dur": 56.121,
    "args": {
      "External id": 2481013,"Record function id": 0, "Concrete Inputs": ["", "[]", "False", ""], "Input type": ["bool", "ScalarList", "Scalar", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656503125.921, "dur": 1.680,
    "args": {
      "External id": 2481014,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["long int", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 8717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656503129.005, "dur": 28.796,
    "args": {
      "External id": 2481015,"Record function id": 0, "Concrete Inputs": ["", "4", "False", "False", ""], "Input type": ["bool", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656503131.124, "dur": 26.429,
    "args": {
      "External id": 2481016,"Record function id": 0, "Concrete Inputs": ["", "4", "", "", "", "False", ""], "Input type": ["bool", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[65536], [], [], [], [], [], []], "Ev Idx": 8719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656503136.080, "dur": 2.389,
    "args": {
      "External id": 2481017,"Record function id": 0, "Concrete Inputs": ["[65536]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656503139.487, "dur": 17.700,
    "args": {
      "External id": 2481018,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[65536], [65536], []], "Ev Idx": 8721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1336759, "tid": 1336759,
    "ts": 1514656503181.514, "dur": 67484.768,
    "args": {
      "External id": 2481019,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 8722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1336759, "tid": 1336759,
    "ts": 1514656503183.560, "dur": 67481.520,
    "args": {
      "External id": 2481020,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 8723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656570679.815, "dur": 8.794,
    "args": {
      "External id": 2481021,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 8724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656570685.525, "dur": 1.111,
    "args": {
      "External id": 2481022,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 8725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656570694.076, "dur": 108.223,
    "args": {
      "External id": 2481023,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 8726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656570695.656, "dur": 9.171,
    "args": {
      "External id": 2481024,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 8727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656570697.941, "dur": 6.103,
    "args": {
      "External id": 2481025,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 8728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656570703.070, "dur": 0.751,
    "args": {
      "External id": 2481026,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 8729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656570706.122, "dur": 95.407,
    "args": {
      "External id": 2481027,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 8730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656570709.808, "dur": 90.856,
    "args": {
      "External id": 2481028,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 8731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656570806.233, "dur": 4.407,
    "args": {
      "External id": 2481029,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656570808.549, "dur": 0.750,
    "args": {
      "External id": 2481030,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "0"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656570818.842, "dur": 3.534,
    "args": {
      "External id": 2481031,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 8734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656570833.498, "dur": 7.436,
    "args": {
      "External id": 2481032,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 8735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656570836.148, "dur": 4.463,
    "args": {
      "External id": 2481033,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656570979.034, "dur": 270.192,
    "args": {
      "External id": 2481034,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571021.667, "dur": 3.866,
    "args": {
      "External id": 2481035,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571028.792, "dur": 219.749,
    "args": {
      "External id": 2481036,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 8739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571032.198, "dur": 0.753,
    "args": {
      "External id": 2481037,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571036.110, "dur": 30.473,
    "args": {
      "External id": 2481038,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571069.123, "dur": 5.853,
    "args": {
      "External id": 2481039,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 8742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571073.706, "dur": 0.878,
    "args": {
      "External id": 2481040,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 8743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571076.698, "dur": 27.053,
    "args": {
      "External id": 2481041,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571077.720, "dur": 1.296,
    "args": {
      "External id": 2481042,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571083.094, "dur": 20.344,
    "args": {
      "External id": 2481043,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 8746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571087.455, "dur": 3.412,
    "args": {
      "External id": 2481044,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571106.414, "dur": 24.386,
    "args": {
      "External id": 2481045,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571133.520, "dur": 19.237,
    "args": {
      "External id": 2481046,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571157.084, "dur": 15.988,
    "args": {
      "External id": 2481047,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 8750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571175.987, "dur": 16.379,
    "args": {
      "External id": 2481048,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571194.937, "dur": 22.259,
    "args": {
      "External id": 2481049,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 8752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571197.183, "dur": 1.375,
    "args": {
      "External id": 2481050,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571202.178, "dur": 0.580,
    "args": {
      "External id": 2481051,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 8754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571220.062, "dur": 14.954,
    "args": {
      "External id": 2481052,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571237.077, "dur": 10.344,
    "args": {
      "External id": 2481053,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571257.471, "dur": 2.840,
    "args": {
      "External id": 2481054,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571267.865, "dur": 4.288,
    "args": {
      "External id": 2481055,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571270.331, "dur": 0.829,
    "args": {
      "External id": 2481056,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571360.575, "dur": 72.077,
    "args": {
      "External id": 2481057,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 8760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571438.850, "dur": 6.836,
    "args": {
      "External id": 2481058,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 8761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571443.062, "dur": 1.322,
    "args": {
      "External id": 2481059,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 8762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571447.413, "dur": 30.249,
    "args": {
      "External id": 2481060,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 8763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571483.597, "dur": 6.073,
    "args": {
      "External id": 2481061,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 8764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571485.486, "dur": 3.532,
    "args": {
      "External id": 2481062,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 8765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571487.242, "dur": 1.565,
    "args": {
      "External id": 2481063,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 8766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571493.397, "dur": 65.404,
    "args": {
      "External id": 2481064,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 8767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571496.629, "dur": 60.987,
    "args": {
      "External id": 2481065,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 8768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571565.789, "dur": 18.515,
    "args": {
      "External id": 2481066,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 8769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571591.435, "dur": 5.041,
    "args": {
      "External id": 2481067,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 8770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571594.119, "dur": 1.073,
    "args": {
      "External id": 2481068,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "16777216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 8771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571601.063, "dur": 53.618,
    "args": {
      "External id": 2481069,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 8772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571602.137, "dur": 5.902,
    "args": {
      "External id": 2481070,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 8773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571603.445, "dur": 3.898,
    "args": {
      "External id": 2481071,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 8774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571606.322, "dur": 0.885,
    "args": {
      "External id": 2481072,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 8775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571608.628, "dur": 45.687,
    "args": {
      "External id": 2481073,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 8776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571609.452, "dur": 44.240,
    "args": {
      "External id": 2481074,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 8777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571658.896, "dur": 4.048,
    "args": {
      "External id": 2481075,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571661.067, "dur": 0.849,
    "args": {
      "External id": 2481076,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "8192"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571669.329, "dur": 2.037,
    "args": {
      "External id": 2481077,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 8780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571680.980, "dur": 7.743,
    "args": {
      "External id": 2481078,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 8781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571682.693, "dur": 5.723,
    "args": {
      "External id": 2481079,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571785.508, "dur": 179.824,
    "args": {
      "External id": 2481080,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571787.850, "dur": 2.410,
    "args": {
      "External id": 2481081,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571791.702, "dur": 173.088,
    "args": {
      "External id": 2481082,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 8785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571793.460, "dur": 0.298,
    "args": {
      "External id": 2481083,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571796.695, "dur": 23.351,
    "args": {
      "External id": 2481084,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571821.936, "dur": 3.382,
    "args": {
      "External id": 2481085,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 8788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571824.249, "dur": 0.804,
    "args": {
      "External id": 2481086,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 8789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571826.332, "dur": 22.975,
    "args": {
      "External id": 2481087,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571827.439, "dur": 1.801,
    "args": {
      "External id": 2481088,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571830.388, "dur": 18.559,
    "args": {
      "External id": 2481089,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 8792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571834.818, "dur": 2.503,
    "args": {
      "External id": 2481090,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571850.873, "dur": 21.164,
    "args": {
      "External id": 2481091,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571873.417, "dur": 12.774,
    "args": {
      "External id": 2481092,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571889.283, "dur": 13.284,
    "args": {
      "External id": 2481093,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 8796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571905.883, "dur": 11.544,
    "args": {
      "External id": 2481094,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571919.437, "dur": 19.785,
    "args": {
      "External id": 2481095,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 8798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571921.548, "dur": 1.549,
    "args": {
      "External id": 2481096,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571925.001, "dur": 0.922,
    "args": {
      "External id": 2481097,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 8800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571940.932, "dur": 11.448,
    "args": {
      "External id": 2481098,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571953.387, "dur": 10.430,
    "args": {
      "External id": 2481099,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656571971.883, "dur": 1.975,
    "args": {
      "External id": 2481100,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572019.832, "dur": 4.386,
    "args": {
      "External id": 2481101,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572022.211, "dur": 0.654,
    "args": {
      "External id": 2481102,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "8192"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572098.923, "dur": 56.632,
    "args": {
      "External id": 2481103,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 8806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572161.318, "dur": 5.091,
    "args": {
      "External id": 2481104,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 8807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572163.813, "dur": 1.378,
    "args": {
      "External id": 2481105,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "16777216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 8808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572168.022, "dur": 26.212,
    "args": {
      "External id": 2481106,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 8809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572198.949, "dur": 6.836,
    "args": {
      "External id": 2481107,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 8810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572200.702, "dur": 4.445,
    "args": {
      "External id": 2481108,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 8811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572204.131, "dur": 0.775,
    "args": {
      "External id": 2481109,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 8812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572208.428, "dur": 40.164,
    "args": {
      "External id": 2481110,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 8813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572209.716, "dur": 38.169,
    "args": {
      "External id": 2481111,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 8814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572252.628, "dur": 15.301,
    "args": {
      "External id": 2481112,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 8815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572273.611, "dur": 4.000,
    "args": {
      "External id": 2481113,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 8816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572275.892, "dur": 0.779,
    "args": {
      "External id": 2481114,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "33554432"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 8817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572283.157, "dur": 50.068,
    "args": {
      "External id": 2481115,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 8818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572284.195, "dur": 5.490,
    "args": {
      "External id": 2481116,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 8819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572285.081, "dur": 3.978,
    "args": {
      "External id": 2481117,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 8820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572286.422, "dur": 2.466,
    "args": {
      "External id": 2481118,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 8821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572290.524, "dur": 42.309,
    "args": {
      "External id": 2481119,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 8822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572291.305, "dur": 40.938,
    "args": {
      "External id": 2481120,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 8823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572337.132, "dur": 3.528,
    "args": {
      "External id": 2481121,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572339.023, "dur": 0.619,
    "args": {
      "External id": 2481122,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "16384"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572348.064, "dur": 1.558,
    "args": {
      "External id": 2481123,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 8826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572357.327, "dur": 6.002,
    "args": {
      "External id": 2481124,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 8827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572359.352, "dur": 3.665,
    "args": {
      "External id": 2481125,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572449.203, "dur": 193.930,
    "args": {
      "External id": 2481126,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572451.661, "dur": 2.060,
    "args": {
      "External id": 2481127,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572455.068, "dur": 187.574,
    "args": {
      "External id": 2481128,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 8831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572458.177, "dur": 0.470,
    "args": {
      "External id": 2481129,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572461.169, "dur": 19.511,
    "args": {
      "External id": 2481130,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572482.454, "dur": 2.741,
    "args": {
      "External id": 2481131,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 8834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572484.106, "dur": 0.814,
    "args": {
      "External id": 2481132,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 8835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572486.038, "dur": 22.446,
    "args": {
      "External id": 2481133,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572487.587, "dur": 3.226,
    "args": {
      "External id": 2481134,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572492.002, "dur": 16.106,
    "args": {
      "External id": 2481135,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 8838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572494.318, "dur": 3.123,
    "args": {
      "External id": 2481136,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572509.872, "dur": 36.124,
    "args": {
      "External id": 2481137,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572548.335, "dur": 14.289,
    "args": {
      "External id": 2481138,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572565.744, "dur": 12.908,
    "args": {
      "External id": 2481139,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 8842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572581.429, "dur": 11.346,
    "args": {
      "External id": 2481140,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572594.488, "dur": 19.440,
    "args": {
      "External id": 2481141,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 8844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572596.601, "dur": 1.740,
    "args": {
      "External id": 2481142,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572600.209, "dur": 0.780,
    "args": {
      "External id": 2481143,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 8846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572615.263, "dur": 12.445,
    "args": {
      "External id": 2481144,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572628.554, "dur": 11.306,
    "args": {
      "External id": 2481145,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572650.670, "dur": 2.295,
    "args": {
      "External id": 2481146,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572661.858, "dur": 3.423,
    "args": {
      "External id": 2481147,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572663.984, "dur": 0.542,
    "args": {
      "External id": 2481148,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "16384"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572732.658, "dur": 78.720,
    "args": {
      "External id": 2481149,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 8852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572817.096, "dur": 4.709,
    "args": {
      "External id": 2481150,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 8853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572819.816, "dur": 0.973,
    "args": {
      "External id": 2481151,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "33554432"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 8854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572823.500, "dur": 25.713,
    "args": {
      "External id": 2481152,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 8855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572891.304, "dur": 9.018,
    "args": {
      "External id": 2481153,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 8856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572895.060, "dur": 4.567,
    "args": {
      "External id": 2481154,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 8857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572896.963, "dur": 2.394,
    "args": {
      "External id": 2481155,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 8858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572902.893, "dur": 42.079,
    "args": {
      "External id": 2481156,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 8859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572903.931, "dur": 40.521,
    "args": {
      "External id": 2481157,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 8860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572948.932, "dur": 13.546,
    "args": {
      "External id": 2481158,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 8861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572968.162, "dur": 5.325,
    "args": {
      "External id": 2481159,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 8862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572971.868, "dur": 0.875,
    "args": {
      "External id": 2481160,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "50331648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 8863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572977.329, "dur": 89.885,
    "args": {
      "External id": 2481161,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 8864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572978.077, "dur": 39.391,
    "args": {
      "External id": 2481162,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 8865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572978.823, "dur": 37.731,
    "args": {
      "External id": 2481163,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 8866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656572980.006, "dur": 35.899,
    "args": {
      "External id": 2481164,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 8867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573018.921, "dur": 47.778,
    "args": {
      "External id": 2481165,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 8868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573019.756, "dur": 46.451,
    "args": {
      "External id": 2481166,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 8869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573074.336, "dur": 4.015,
    "args": {
      "External id": 2481167,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573076.570, "dur": 0.644,
    "args": {
      "External id": 2481168,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "24576"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573085.206, "dur": 1.545,
    "args": {
      "External id": 2481169,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 8872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573094.598, "dur": 6.021,
    "args": {
      "External id": 2481170,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 8873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573096.204, "dur": 4.150,
    "args": {
      "External id": 2481171,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573189.302, "dur": 171.416,
    "args": {
      "External id": 2481172,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573191.572, "dur": 2.226,
    "args": {
      "External id": 2481173,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573196.986, "dur": 163.208,
    "args": {
      "External id": 2481174,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 8877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573198.464, "dur": 0.308,
    "args": {
      "External id": 2481175,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573199.957, "dur": 21.162,
    "args": {
      "External id": 2481176,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573222.549, "dur": 4.895,
    "args": {
      "External id": 2481177,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 8880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573224.474, "dur": 2.646,
    "args": {
      "External id": 2481178,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 8881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573228.257, "dur": 21.772,
    "args": {
      "External id": 2481179,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573229.564, "dur": 1.597,
    "args": {
      "External id": 2481180,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573233.865, "dur": 15.848,
    "args": {
      "External id": 2481181,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 8884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573236.659, "dur": 2.438,
    "args": {
      "External id": 2481182,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573251.302, "dur": 20.326,
    "args": {
      "External id": 2481183,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573273.124, "dur": 12.784,
    "args": {
      "External id": 2481184,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573289.118, "dur": 12.706,
    "args": {
      "External id": 2481185,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 8888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573302.980, "dur": 12.053,
    "args": {
      "External id": 2481186,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573316.524, "dur": 17.122,
    "args": {
      "External id": 2481187,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 8890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573318.393, "dur": 1.038,
    "args": {
      "External id": 2481188,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573321.093, "dur": 0.718,
    "args": {
      "External id": 2481189,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 8892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573336.752, "dur": 11.312,
    "args": {
      "External id": 2481190,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573348.938, "dur": 10.497,
    "args": {
      "External id": 2481191,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573366.565, "dur": 1.572,
    "args": {
      "External id": 2481192,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573375.541, "dur": 2.895,
    "args": {
      "External id": 2481193,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573377.239, "dur": 0.359,
    "args": {
      "External id": 2481194,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "24576"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573438.920, "dur": 46.626,
    "args": {
      "External id": 2481195,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 8898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573490.354, "dur": 8.078,
    "args": {
      "External id": 2481196,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 8899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573494.524, "dur": 2.699,
    "args": {
      "External id": 2481197,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "50331648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 8900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573499.897, "dur": 22.936,
    "args": {
      "External id": 2481198,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 8901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573543.728, "dur": 5.926,
    "args": {
      "External id": 2481199,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 8902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573545.379, "dur": 3.463,
    "args": {
      "External id": 2481200,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 8903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573547.368, "dur": 0.868,
    "args": {
      "External id": 2481201,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 8904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573552.882, "dur": 43.324,
    "args": {
      "External id": 2481202,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 8905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573554.088, "dur": 41.564,
    "args": {
      "External id": 2481203,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 8906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573601.824, "dur": 15.117,
    "args": {
      "External id": 2481204,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 8907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573622.783, "dur": 3.901,
    "args": {
      "External id": 2481205,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 8908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573625.052, "dur": 0.773,
    "args": {
      "External id": 2481206,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "67108864"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 8909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573630.382, "dur": 46.573,
    "args": {
      "External id": 2481207,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 8910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573631.288, "dur": 4.752,
    "args": {
      "External id": 2481208,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 8911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573631.846, "dur": 3.666,
    "args": {
      "External id": 2481209,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 8912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573634.737, "dur": 0.640,
    "args": {
      "External id": 2481210,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 8913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573636.608, "dur": 40.000,
    "args": {
      "External id": 2481211,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 8914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573637.240, "dur": 38.907,
    "args": {
      "External id": 2481212,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 8915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573681.290, "dur": 3.554,
    "args": {
      "External id": 2481213,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573683.099, "dur": 0.737,
    "args": {
      "External id": 2481214,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "32768"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573690.589, "dur": 1.374,
    "args": {
      "External id": 2481215,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 8918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573699.158, "dur": 7.555,
    "args": {
      "External id": 2481216,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 8919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573702.541, "dur": 3.859,
    "args": {
      "External id": 2481217,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573784.162, "dur": 164.280,
    "args": {
      "External id": 2481218,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573785.794, "dur": 3.649,
    "args": {
      "External id": 2481219,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573790.566, "dur": 157.294,
    "args": {
      "External id": 2481220,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 8923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573792.169, "dur": 0.300,
    "args": {
      "External id": 2481221,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573793.223, "dur": 19.781,
    "args": {
      "External id": 2481222,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573814.541, "dur": 4.557,
    "args": {
      "External id": 2481223,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 8926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573818.116, "dur": 0.638,
    "args": {
      "External id": 2481224,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 8927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573820.152, "dur": 20.872,
    "args": {
      "External id": 2481225,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573822.941, "dur": 1.449,
    "args": {
      "External id": 2481226,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573825.358, "dur": 15.354,
    "args": {
      "External id": 2481227,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 8930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573827.514, "dur": 2.363,
    "args": {
      "External id": 2481228,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573842.229, "dur": 18.324,
    "args": {
      "External id": 2481229,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573861.816, "dur": 12.458,
    "args": {
      "External id": 2481230,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573876.869, "dur": 12.352,
    "args": {
      "External id": 2481231,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 8934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573890.509, "dur": 10.689,
    "args": {
      "External id": 2481232,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573902.710, "dur": 20.836,
    "args": {
      "External id": 2481233,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 8936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573906.371, "dur": 1.423,
    "args": {
      "External id": 2481234,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573911.104, "dur": 0.570,
    "args": {
      "External id": 2481235,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 8938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573924.824, "dur": 10.882,
    "args": {
      "External id": 2481236,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573936.569, "dur": 10.417,
    "args": {
      "External id": 2481237,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573954.577, "dur": 1.822,
    "args": {
      "External id": 2481238,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573963.288, "dur": 3.005,
    "args": {
      "External id": 2481239,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656573965.045, "dur": 0.522,
    "args": {
      "External id": 2481240,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "32768"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574070.180, "dur": 54.922,
    "args": {
      "External id": 2481241,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 8944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574130.716, "dur": 7.496,
    "args": {
      "External id": 2481242,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 8945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574135.353, "dur": 1.250,
    "args": {
      "External id": 2481243,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "67108864"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 8946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574139.330, "dur": 23.441,
    "args": {
      "External id": 2481244,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 8947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574167.778, "dur": 5.457,
    "args": {
      "External id": 2481245,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 8948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574169.095, "dur": 3.346,
    "args": {
      "External id": 2481246,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 8949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574171.199, "dur": 1.001,
    "args": {
      "External id": 2481247,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 8950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574175.649, "dur": 40.311,
    "args": {
      "External id": 2481248,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 8951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574178.407, "dur": 37.082,
    "args": {
      "External id": 2481249,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 8952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574220.187, "dur": 14.844,
    "args": {
      "External id": 2481250,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 8953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574240.559, "dur": 3.416,
    "args": {
      "External id": 2481251,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 8954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574242.468, "dur": 0.710,
    "args": {
      "External id": 2481252,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "83886080"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 8955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574247.495, "dur": 47.980,
    "args": {
      "External id": 2481253,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 8956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574248.271, "dur": 5.620,
    "args": {
      "External id": 2481254,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 8957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574249.369, "dur": 3.950,
    "args": {
      "External id": 2481255,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 8958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574252.550, "dur": 0.601,
    "args": {
      "External id": 2481256,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 8959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574254.540, "dur": 40.535,
    "args": {
      "External id": 2481257,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 8960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574255.308, "dur": 39.291,
    "args": {
      "External id": 2481258,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 8961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574299.212, "dur": 3.897,
    "args": {
      "External id": 2481259,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574301.139, "dur": 0.856,
    "args": {
      "External id": 2481260,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "40960"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574316.057, "dur": 1.677,
    "args": {
      "External id": 2481261,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 8964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574329.162, "dur": 7.609,
    "args": {
      "External id": 2481262,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 8965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574330.750, "dur": 5.678,
    "args": {
      "External id": 2481263,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574420.933, "dur": 197.066,
    "args": {
      "External id": 2481264,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574423.218, "dur": 1.759,
    "args": {
      "External id": 2481265,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574426.474, "dur": 190.873,
    "args": {
      "External id": 2481266,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 8969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574435.137, "dur": 0.344,
    "args": {
      "External id": 2481267,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574439.854, "dur": 19.928,
    "args": {
      "External id": 2481268,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 8971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574461.528, "dur": 4.641,
    "args": {
      "External id": 2481269,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 8972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574463.663, "dur": 2.242,
    "args": {
      "External id": 2481270,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 8973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574466.991, "dur": 20.187,
    "args": {
      "External id": 2481271,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574468.240, "dur": 1.485,
    "args": {
      "External id": 2481272,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 8975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574470.963, "dur": 15.872,
    "args": {
      "External id": 2481273,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 8976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574473.365, "dur": 2.727,
    "args": {
      "External id": 2481274,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574488.593, "dur": 19.859,
    "args": {
      "External id": 2481275,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574509.568, "dur": 11.474,
    "args": {
      "External id": 2481276,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574523.438, "dur": 31.142,
    "args": {
      "External id": 2481277,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 8980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574556.699, "dur": 11.819,
    "args": {
      "External id": 2481278,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 8981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574572.281, "dur": 19.936,
    "args": {
      "External id": 2481279,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 8982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574574.055, "dur": 2.140,
    "args": {
      "External id": 2481280,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 8983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574578.186, "dur": 0.907,
    "args": {
      "External id": 2481281,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 8984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574593.689, "dur": 10.777,
    "args": {
      "External id": 2481282,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 8985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574605.389, "dur": 11.022,
    "args": {
      "External id": 2481283,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 8986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574627.017, "dur": 2.113,
    "args": {
      "External id": 2481284,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 8987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574638.528, "dur": 4.029,
    "args": {
      "External id": 2481285,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 8988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574641.270, "dur": 0.476,
    "args": {
      "External id": 2481286,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "40960"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 8989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574708.400, "dur": 52.233,
    "args": {
      "External id": 2481287,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 8990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574765.479, "dur": 5.477,
    "args": {
      "External id": 2481288,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 8991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574767.589, "dur": 2.308,
    "args": {
      "External id": 2481289,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "83886080"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 8992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574772.368, "dur": 22.708,
    "args": {
      "External id": 2481290,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 8993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574799.320, "dur": 6.535,
    "args": {
      "External id": 2481291,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 8994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574800.716, "dur": 4.554,
    "args": {
      "External id": 2481292,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 8995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574804.098, "dur": 1.015,
    "args": {
      "External id": 2481293,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 8996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574808.514, "dur": 38.333,
    "args": {
      "External id": 2481294,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 8997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574809.581, "dur": 36.663,
    "args": {
      "External id": 2481295,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 8998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574850.384, "dur": 14.099,
    "args": {
      "External id": 2481296,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 8999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574870.294, "dur": 3.795,
    "args": {
      "External id": 2481297,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574872.440, "dur": 0.887,
    "args": {
      "External id": 2481298,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "100663296"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574877.837, "dur": 46.592,
    "args": {
      "External id": 2481299,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 9002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574880.115, "dur": 3.670,
    "args": {
      "External id": 2481300,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 9003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574881.067, "dur": 2.122,
    "args": {
      "External id": 2481301,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 9004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574882.394, "dur": 0.665,
    "args": {
      "External id": 2481302,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 9005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574884.258, "dur": 39.842,
    "args": {
      "External id": 2481303,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574884.761, "dur": 38.823,
    "args": {
      "External id": 2481304,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574928.691, "dur": 3.282,
    "args": {
      "External id": 2481305,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574930.369, "dur": 0.585,
    "args": {
      "External id": 2481306,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "49152"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574937.164, "dur": 1.480,
    "args": {
      "External id": 2481307,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 9010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574946.994, "dur": 5.451,
    "args": {
      "External id": 2481308,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 9011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656574948.603, "dur": 3.566,
    "args": {
      "External id": 2481309,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575072.830, "dur": 172.479,
    "args": {
      "External id": 2481310,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575075.406, "dur": 5.017,
    "args": {
      "External id": 2481311,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575081.698, "dur": 163.200,
    "args": {
      "External id": 2481312,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 9015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575083.312, "dur": 0.553,
    "args": {
      "External id": 2481313,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575086.703, "dur": 20.856,
    "args": {
      "External id": 2481314,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575108.908, "dur": 5.753,
    "args": {
      "External id": 2481315,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 9018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575113.278, "dur": 1.043,
    "args": {
      "External id": 2481316,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 9019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575115.670, "dur": 19.695,
    "args": {
      "External id": 2481317,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575116.530, "dur": 1.414,
    "args": {
      "External id": 2481318,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575118.983, "dur": 16.122,
    "args": {
      "External id": 2481319,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 9022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575121.862, "dur": 2.488,
    "args": {
      "External id": 2481320,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575136.622, "dur": 20.054,
    "args": {
      "External id": 2481321,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575158.040, "dur": 12.580,
    "args": {
      "External id": 2481322,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575173.449, "dur": 12.048,
    "args": {
      "External id": 2481323,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 9026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575188.534, "dur": 10.926,
    "args": {
      "External id": 2481324,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575201.027, "dur": 19.133,
    "args": {
      "External id": 2481325,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 9028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575202.829, "dur": 1.134,
    "args": {
      "External id": 2481326,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575205.830, "dur": 2.157,
    "args": {
      "External id": 2481327,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 9030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575221.445, "dur": 10.765,
    "args": {
      "External id": 2481328,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575233.043, "dur": 10.787,
    "args": {
      "External id": 2481329,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575251.672, "dur": 1.766,
    "args": {
      "External id": 2481330,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575263.073, "dur": 3.464,
    "args": {
      "External id": 2481331,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575265.382, "dur": 0.454,
    "args": {
      "External id": 2481332,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "49152"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575326.566, "dur": 51.079,
    "args": {
      "External id": 2481333,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 9036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575382.730, "dur": 3.863,
    "args": {
      "External id": 2481334,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575384.978, "dur": 0.542,
    "args": {
      "External id": 2481335,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "100663296"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575388.324, "dur": 26.700,
    "args": {
      "External id": 2481336,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 9039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575419.217, "dur": 6.992,
    "args": {
      "External id": 2481337,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 9040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575420.698, "dur": 4.802,
    "args": {
      "External id": 2481338,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 9041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575424.221, "dur": 1.110,
    "args": {
      "External id": 2481339,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 9042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575428.809, "dur": 37.283,
    "args": {
      "External id": 2481340,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575429.942, "dur": 35.553,
    "args": {
      "External id": 2481341,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575469.805, "dur": 13.043,
    "args": {
      "External id": 2481342,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 9045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575488.251, "dur": 2.909,
    "args": {
      "External id": 2481343,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575489.827, "dur": 0.635,
    "args": {
      "External id": 2481344,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "117440512"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575496.486, "dur": 60.886,
    "args": {
      "External id": 2481345,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 9048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575497.362, "dur": 3.497,
    "args": {
      "External id": 2481346,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 9049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575497.995, "dur": 2.304,
    "args": {
      "External id": 2481347,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 9050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575499.367, "dur": 0.780,
    "args": {
      "External id": 2481348,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 9051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575501.570, "dur": 55.269,
    "args": {
      "External id": 2481349,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575502.404, "dur": 53.528,
    "args": {
      "External id": 2481350,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575564.752, "dur": 3.956,
    "args": {
      "External id": 2481351,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575566.797, "dur": 0.660,
    "args": {
      "External id": 2481352,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "57344"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575574.926, "dur": 1.494,
    "args": {
      "External id": 2481353,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 9056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575583.448, "dur": 5.999,
    "args": {
      "External id": 2481354,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 9057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575584.954, "dur": 4.207,
    "args": {
      "External id": 2481355,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575666.380, "dur": 164.351,
    "args": {
      "External id": 2481356,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575669.857, "dur": 2.272,
    "args": {
      "External id": 2481357,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575673.419, "dur": 156.864,
    "args": {
      "External id": 2481358,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 9061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575674.954, "dur": 0.264,
    "args": {
      "External id": 2481359,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575676.077, "dur": 19.752,
    "args": {
      "External id": 2481360,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575697.008, "dur": 4.965,
    "args": {
      "External id": 2481361,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 9064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575700.846, "dur": 0.832,
    "args": {
      "External id": 2481362,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 9065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575702.866, "dur": 20.713,
    "args": {
      "External id": 2481363,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575703.741, "dur": 1.494,
    "args": {
      "External id": 2481364,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575706.119, "dur": 17.092,
    "args": {
      "External id": 2481365,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 9068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575710.384, "dur": 2.270,
    "args": {
      "External id": 2481366,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575724.929, "dur": 19.197,
    "args": {
      "External id": 2481367,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575745.154, "dur": 11.166,
    "args": {
      "External id": 2481368,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575758.768, "dur": 12.217,
    "args": {
      "External id": 2481369,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 9072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575772.095, "dur": 10.899,
    "args": {
      "External id": 2481370,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575784.830, "dur": 19.460,
    "args": {
      "External id": 2481371,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 9074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575788.593, "dur": 1.092,
    "args": {
      "External id": 2481372,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575791.663, "dur": 0.608,
    "args": {
      "External id": 2481373,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 9076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575807.057, "dur": 10.783,
    "args": {
      "External id": 2481374,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575818.872, "dur": 10.555,
    "args": {
      "External id": 2481375,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575836.348, "dur": 1.426,
    "args": {
      "External id": 2481376,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575844.656, "dur": 3.225,
    "args": {
      "External id": 2481377,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575846.702, "dur": 0.459,
    "args": {
      "External id": 2481378,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "57344"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575902.115, "dur": 50.065,
    "args": {
      "External id": 2481379,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 9082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575956.791, "dur": 3.949,
    "args": {
      "External id": 2481380,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575959.025, "dur": 0.747,
    "args": {
      "External id": 2481381,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "117440512"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656575963.883, "dur": 57.027,
    "args": {
      "External id": 2481382,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 9085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576027.524, "dur": 5.754,
    "args": {
      "External id": 2481383,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 9086
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576029.164, "dur": 3.305,
    "args": {
      "External id": 2481384,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 9087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576031.028, "dur": 1.200,
    "args": {
      "External id": 2481385,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 9088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576036.007, "dur": 43.399,
    "args": {
      "External id": 2481386,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576037.021, "dur": 41.700,
    "args": {
      "External id": 2481387,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576083.417, "dur": 14.218,
    "args": {
      "External id": 2481388,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 9091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576103.790, "dur": 21.876,
    "args": {
      "External id": 2481389,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 9092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576105.906, "dur": 19.387,
    "args": {
      "External id": 2481390,"Record function id": 0, "Concrete Inputs": ["", "[]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576110.879, "dur": 0.609,
    "args": {
      "External id": 2481391,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576131.413, "dur": 28.197,
    "args": {
      "External id": 2481392,"Record function id": 0, "Concrete Inputs": ["", "", "15", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[32000, 2048], [], [], [], [], []], "Ev Idx": 9095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576133.093, "dur": 26.300,
    "args": {
      "External id": 2481393,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], [], []], "Input Dims": [[32000, 2048], [], [], [], [], [], []], "Ev Idx": 9096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576137.310, "dur": 3.738,
    "args": {
      "External id": 2481394,"Record function id": 0, "Concrete Inputs": ["[32000, 2048]", "[2048, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576142.059, "dur": 16.868,
    "args": {
      "External id": 2481395,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "float", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 9098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576173.195, "dur": 4.633,
    "args": {
      "External id": 2481396,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 9099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576174.986, "dur": 2.599,
    "args": {
      "External id": 2481397,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 9100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576178.834, "dur": 2.779,
    "args": {
      "External id": 2481398,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 9101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576181.093, "dur": 0.457,
    "args": {
      "External id": 2481399,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 9102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576221.444, "dur": 23.454,
    "args": {
      "External id": 2481400,"Sequence number": 24740625, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1], [], []], "Ev Idx": 9103
    }
  },
  {
    "ph": "s", "id": 23, "pid": 1336759, "tid": 1336759, "ts": 1514656576221.444,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576250.911, "dur": 10.664,
    "args": {
      "External id": 2481401,"Sequence number": 24740626, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], [], []], "Input Dims": [[16, 4, 4096], [], [], [], []], "Ev Idx": 9104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576259.278, "dur": 1.077,
    "args": {
      "External id": 2481402,"Record function id": 0, "Concrete Inputs": ["", "[16, 4, 4096]", "[8192, 1, 1]", "1"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[16, 4, 4096], [], [], []], "Ev Idx": 9105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576264.065, "dur": 7.623,
    "args": {
      "External id": 2481403,"Sequence number": 24740626, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "1"], "Input type": ["long int", "Scalar", "Scalar"], "Input Strides": [[8192, 1, 1], [], []], "Input Dims": [[16, 4, 4096], [], []], "Ev Idx": 9106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576269.667, "dur": 0.881,
    "args": {
      "External id": 2481404,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096]", "[8192, 1]", "2"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[16, 4, 4096], [], [], []], "Ev Idx": 9107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576273.032, "dur": 2.937,
    "args": {
      "External id": 2481405,"Sequence number": 24740626, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], [], [], []], "Input Dims": [[16, 4096], [], [], [], []], "Ev Idx": 9108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576274.660, "dur": 0.736,
    "args": {
      "External id": 2481406,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096]", "[8192, 1]", "2"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1], [], [], []], "Input Dims": [[16, 4096], [], [], []], "Ev Idx": 9109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576279.959, "dur": 5.277,
    "args": {
      "External id": 2481407,"Sequence number": 24740626, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], [], []], "Ev Idx": 9110
    }
  },
  {
    "ph": "s", "id": 22, "pid": 1336759, "tid": 1336759, "ts": 1514656576279.959,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576283.174, "dur": 0.890,
    "args": {
      "External id": 2481408,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 2048]", "[33554432, 8192, 2048, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], []], "Ev Idx": 9111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576286.359, "dur": 6.536,
    "args": {
      "External id": 2481409,"Sequence number": 24740627, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], [], []], "Ev Idx": 9112
    }
  },
  {
    "ph": "s", "id": 21, "pid": 1336759, "tid": 1336759, "ts": 1514656576286.359,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576291.371, "dur": 0.768,
    "args": {
      "External id": 2481410,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 2048]", "[33554432, 8192, 2048, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], []], "Ev Idx": 9113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576293.926, "dur": 5.418,
    "args": {
      "External id": 2481411,"Sequence number": 24740628, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], []], "Input Dims": [[16, 4096, 4, 2048], [], []], "Ev Idx": 9114
    }
  },
  {
    "ph": "s", "id": 20, "pid": 1336759, "tid": 1336759, "ts": 1514656576293.926,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576297.549, "dur": 0.934,
    "args": {
      "External id": 2481412,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]", "[33554432, 8192, 1]", "2048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], []], "Ev Idx": 9115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576300.318, "dur": 4.293,
    "args": {
      "External id": 2481413,"Sequence number": 24740629, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 1], [], [], [], []], "Input Dims": [[16, 4096, 2048], [], [], [], []], "Ev Idx": 9116
    }
  },
  {
    "ph": "s", "id": 19, "pid": 1336759, "tid": 1336759, "ts": 1514656576300.318,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576302.880, "dur": 1.026,
    "args": {
      "External id": 2481414,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]", "[33554432, 8192, 1]", "2048"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 1], [], [], []], "Input Dims": [[16, 4096, 2048], [], [], []], "Ev Idx": 9117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576308.453, "dur": 29.292,
    "args": {
      "External id": 2481415,"Sequence number": 24740630, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["long int", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[16, 4096], []], "Ev Idx": 9118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576309.926, "dur": 27.600,
    "args": {
      "External id": 2481416,"Sequence number": 24740630, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["long int", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[16, 4096], []], "Ev Idx": 9119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576312.165, "dur": 9.155,
    "args": {
      "External id": 2481417,"Record function id": 0, "Concrete Inputs": ["", "4", "0", "", "", "0"], "Input type": ["long int", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[8192, 1], [], [], [], [], []], "Input Dims": [[16, 4096], [], [], [], [], []], "Ev Idx": 9120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576315.775, "dur": 4.958,
    "args": {
      "External id": 2481418,"Record function id": 0, "Concrete Inputs": ["[16, 4096]", "4", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576322.192, "dur": 14.824,
    "args": {
      "External id": 2481419,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[4096, 1], [8192, 1], []], "Input Dims": [[16, 4096], [16, 4096], []], "Ev Idx": 9122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576362.529, "dur": 4.297,
    "args": {
      "External id": 2481420,"Sequence number": 24740630, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[33554432, 8192, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 9123
    }
  },
  {
    "ph": "s", "id": 18, "pid": 1336759, "tid": 1336759, "ts": 1514656576362.529,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576369.070, "dur": 1.072,
    "args": {
      "External id": 2481421,"Sequence number": 24740631, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["long int", "ScalarList"], "Input Strides": [[4096, 1], []], "Input Dims": [[16, 4096], []], "Ev Idx": 9124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FusedLinearCrossEntropyFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576401.684, "dur": 42907.129,
    "args": {
      "External id": 2481422,"Sequence number": 24740631, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "-100", "0.", "1.", "8"], "Input type": ["c10::BFloat16", "long int", "c10::BFloat16", "", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [1], [2048, 1], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536], [32000, 2048], [], [], [], [], []], "Ev Idx": 9125
    }
  },
  {
    "ph": "s", "id": 17, "pid": 1336759, "tid": 1336759, "ts": 1514656576401.684,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576415.739, "dur": 25.515,
    "args": {
      "External id": 2481423,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 9126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576418.215, "dur": 22.824,
    "args": {
      "External id": 2481424,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 9127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576419.813, "dur": 5.270,
    "args": {
      "External id": 2481425,"Record function id": 0, "Concrete Inputs": ["", "15", "0", "", "", "0"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[8192, 1], [], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], [], []], "Ev Idx": 9128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576421.131, "dur": 3.354,
    "args": {
      "External id": 2481426,"Record function id": 0, "Concrete Inputs": ["[65536, 2048]", "15", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576425.666, "dur": 14.986,
    "args": {
      "External id": 2481427,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [8192, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 9130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576456.262, "dur": 23.192,
    "args": {
      "External id": 2481428,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], [], []], "Ev Idx": 9131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576457.119, "dur": 5.359,
    "args": {
      "External id": 2481429,"Record function id": 0, "Concrete Inputs": ["", "15", "0", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], [], []], "Ev Idx": 9132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576458.961, "dur": 3.215,
    "args": {
      "External id": 2481430,"Record function id": 0, "Concrete Inputs": ["[65536, 2048]", "[2048, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576463.667, "dur": 15.564,
    "args": {
      "External id": 2481431,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 9134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576465.390, "dur": 13.313,
    "args": {
      "External id": 2481432,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 9135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576482.691, "dur": 21.355,
    "args": {
      "External id": 2481433,"Record function id": 0, "Concrete Inputs": ["", "6", "", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[32000, 2048], [], [], [], [], []], "Ev Idx": 9136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576483.444, "dur": 3.897,
    "args": {
      "External id": 2481434,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[32000, 2048], [], [], [], [], []], "Ev Idx": 9137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576484.493, "dur": 2.573,
    "args": {
      "External id": 2481435,"Record function id": 0, "Concrete Inputs": ["[32000, 2048]", "[2048, 1]", "6", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576489.856, "dur": 13.978,
    "args": {
      "External id": 2481436,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 9139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576490.274, "dur": 13.253,
    "args": {
      "External id": 2481437,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[2048, 1], []], "Input Dims": [[32000, 2048], []], "Ev Idx": 9140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576509.498, "dur": 15.640,
    "args": {
      "External id": 2481438,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 9141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576510.947, "dur": 2.856,
    "args": {
      "External id": 2481439,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576514.321, "dur": 10.547,
    "args": {
      "External id": 2481440,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[65536]], "Ev Idx": 9143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576514.781, "dur": 9.768,
    "args": {
      "External id": 2481441,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 9144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::ne", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576546.898, "dur": 23.784,
    "args": {
      "External id": 2481442,"Record function id": 0, "Concrete Inputs": ["", "-100"], "Input type": ["long int", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 9145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576573.670, "dur": 62.498,
    "args": {
      "External id": 2481443,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["bool", ""], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 9146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576591.199, "dur": 44.451,
    "args": {
      "External id": 2481444,"Record function id": 0, "Concrete Inputs": ["", "[]", "False", ""], "Input type": ["bool", "ScalarList", "Scalar", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576596.419, "dur": 1.028,
    "args": {
      "External id": 2481445,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["long int", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576598.573, "dur": 22.971,
    "args": {
      "External id": 2481446,"Record function id": 0, "Concrete Inputs": ["", "4", "False", "False", ""], "Input type": ["bool", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576601.646, "dur": 19.662,
    "args": {
      "External id": 2481447,"Record function id": 0, "Concrete Inputs": ["", "4", "", "", "", "False", ""], "Input type": ["bool", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[65536], [], [], [], [], [], []], "Ev Idx": 9150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576603.983, "dur": 2.912,
    "args": {
      "External id": 2481448,"Record function id": 0, "Concrete Inputs": ["[65536]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576607.442, "dur": 13.415,
    "args": {
      "External id": 2481449,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[65536], [65536], []], "Ev Idx": 9152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576640.121, "dur": 37237.647,
    "args": {
      "External id": 2481450,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1336759, "tid": 1336759,
    "ts": 1514656576641.817, "dur": 37234.781,
    "args": {
      "External id": 2481451,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656613887.744, "dur": 6.198,
    "args": {
      "External id": 2481452,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656613891.218, "dur": 0.906,
    "args": {
      "External id": 2481453,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656613900.629, "dur": 132.216,
    "args": {
      "External id": 2481454,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 9157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656613902.089, "dur": 6.049,
    "args": {
      "External id": 2481455,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 9158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656613904.316, "dur": 2.993,
    "args": {
      "External id": 2481456,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 9159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656613906.157, "dur": 0.860,
    "args": {
      "External id": 2481457,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 9160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656613909.262, "dur": 122.879,
    "args": {
      "External id": 2481458,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656613910.718, "dur": 120.040,
    "args": {
      "External id": 2481459,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614038.057, "dur": 5.231,
    "args": {
      "External id": 2481460,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614040.638, "dur": 1.083,
    "args": {
      "External id": 2481461,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "0"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614052.552, "dur": 2.423,
    "args": {
      "External id": 2481462,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 9165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614063.789, "dur": 6.828,
    "args": {
      "External id": 2481463,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 9166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614065.835, "dur": 4.546,
    "args": {
      "External id": 2481464,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614198.385, "dur": 193.802,
    "args": {
      "External id": 2481465,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614201.167, "dur": 3.720,
    "args": {
      "External id": 2481466,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614208.102, "dur": 183.530,
    "args": {
      "External id": 2481467,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 9170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614209.853, "dur": 0.468,
    "args": {
      "External id": 2481468,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614211.731, "dur": 25.023,
    "args": {
      "External id": 2481469,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614238.708, "dur": 5.587,
    "args": {
      "External id": 2481470,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 9173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614243.186, "dur": 0.798,
    "args": {
      "External id": 2481471,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 9174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614245.424, "dur": 22.436,
    "args": {
      "External id": 2481472,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614246.564, "dur": 1.626,
    "args": {
      "External id": 2481473,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614249.668, "dur": 17.901,
    "args": {
      "External id": 2481474,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 9177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614253.414, "dur": 3.210,
    "args": {
      "External id": 2481475,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614271.280, "dur": 21.310,
    "args": {
      "External id": 2481476,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614294.388, "dur": 13.783,
    "args": {
      "External id": 2481477,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614311.442, "dur": 13.662,
    "args": {
      "External id": 2481478,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 9181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614326.667, "dur": 12.528,
    "args": {
      "External id": 2481479,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614341.210, "dur": 21.606,
    "args": {
      "External id": 2481480,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 9183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614344.133, "dur": 1.481,
    "args": {
      "External id": 2481481,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614347.514, "dur": 2.226,
    "args": {
      "External id": 2481482,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 9185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614366.110, "dur": 12.518,
    "args": {
      "External id": 2481483,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614380.066, "dur": 10.477,
    "args": {
      "External id": 2481484,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614399.189, "dur": 1.983,
    "args": {
      "External id": 2481485,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614407.050, "dur": 3.331,
    "args": {
      "External id": 2481486,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614408.967, "dur": 0.594,
    "args": {
      "External id": 2481487,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614484.201, "dur": 82.329,
    "args": {
      "External id": 2481488,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 9191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614573.543, "dur": 7.235,
    "args": {
      "External id": 2481489,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614576.803, "dur": 1.386,
    "args": {
      "External id": 2481490,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614584.102, "dur": 33.028,
    "args": {
      "External id": 2481491,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 9194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614623.183, "dur": 6.229,
    "args": {
      "External id": 2481492,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 9195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614624.835, "dur": 3.642,
    "args": {
      "External id": 2481493,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 9196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614626.587, "dur": 1.603,
    "args": {
      "External id": 2481494,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 9197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614632.402, "dur": 47.946,
    "args": {
      "External id": 2481495,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614633.557, "dur": 46.306,
    "args": {
      "External id": 2481496,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614684.483, "dur": 15.152,
    "args": {
      "External id": 2481497,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 9200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614707.439, "dur": 3.626,
    "args": {
      "External id": 2481498,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614709.463, "dur": 0.701,
    "args": {
      "External id": 2481499,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "16777216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614715.309, "dur": 55.084,
    "args": {
      "External id": 2481500,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 9203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614716.193, "dur": 3.770,
    "args": {
      "External id": 2481501,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 9204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614717.095, "dur": 2.321,
    "args": {
      "External id": 2481502,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 9205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614718.473, "dur": 0.796,
    "args": {
      "External id": 2481503,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 9206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614722.435, "dur": 47.627,
    "args": {
      "External id": 2481504,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614727.788, "dur": 41.711,
    "args": {
      "External id": 2481505,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614774.563, "dur": 3.910,
    "args": {
      "External id": 2481506,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614776.442, "dur": 0.705,
    "args": {
      "External id": 2481507,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "8192"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614784.595, "dur": 1.525,
    "args": {
      "External id": 2481508,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 9211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614794.299, "dur": 9.439,
    "args": {
      "External id": 2481509,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 9212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614797.784, "dur": 5.652,
    "args": {
      "External id": 2481510,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614895.213, "dur": 219.500,
    "args": {
      "External id": 2481511,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614897.280, "dur": 2.362,
    "args": {
      "External id": 2481512,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614901.358, "dur": 212.792,
    "args": {
      "External id": 2481513,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 9216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614902.784, "dur": 0.388,
    "args": {
      "External id": 2481514,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614904.296, "dur": 22.123,
    "args": {
      "External id": 2481515,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614929.812, "dur": 5.180,
    "args": {
      "External id": 2481516,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 9219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614933.940, "dur": 0.816,
    "args": {
      "External id": 2481517,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 9220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614937.572, "dur": 22.574,
    "args": {
      "External id": 2481518,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614938.806, "dur": 3.023,
    "args": {
      "External id": 2481519,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614943.164, "dur": 16.677,
    "args": {
      "External id": 2481520,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 9223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614945.688, "dur": 2.568,
    "args": {
      "External id": 2481521,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656614961.372, "dur": 55.831,
    "args": {
      "External id": 2481522,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615019.959, "dur": 15.224,
    "args": {
      "External id": 2481523,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615038.450, "dur": 13.364,
    "args": {
      "External id": 2481524,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 9227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615053.181, "dur": 11.991,
    "args": {
      "External id": 2481525,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615067.381, "dur": 21.560,
    "args": {
      "External id": 2481526,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 9229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615071.017, "dur": 1.808,
    "args": {
      "External id": 2481527,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615075.097, "dur": 0.684,
    "args": {
      "External id": 2481528,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 9231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615090.524, "dur": 11.249,
    "args": {
      "External id": 2481529,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615102.879, "dur": 10.272,
    "args": {
      "External id": 2481530,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615122.948, "dur": 2.469,
    "args": {
      "External id": 2481531,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615134.106, "dur": 3.881,
    "args": {
      "External id": 2481532,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615136.718, "dur": 0.487,
    "args": {
      "External id": 2481533,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "8192"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615209.242, "dur": 56.098,
    "args": {
      "External id": 2481534,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 9237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615272.347, "dur": 4.026,
    "args": {
      "External id": 2481535,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615274.604, "dur": 0.756,
    "args": {
      "External id": 2481536,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "16777216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615277.921, "dur": 24.845,
    "args": {
      "External id": 2481537,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 9240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615307.535, "dur": 5.091,
    "args": {
      "External id": 2481538,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 9241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615309.073, "dur": 2.922,
    "args": {
      "External id": 2481539,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 9242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615310.886, "dur": 0.887,
    "args": {
      "External id": 2481540,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 9243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615316.804, "dur": 40.041,
    "args": {
      "External id": 2481541,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615317.970, "dur": 38.268,
    "args": {
      "External id": 2481542,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615360.588, "dur": 13.882,
    "args": {
      "External id": 2481543,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 9246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615379.935, "dur": 3.309,
    "args": {
      "External id": 2481544,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615381.917, "dur": 0.485,
    "args": {
      "External id": 2481545,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "33554432"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615386.825, "dur": 50.609,
    "args": {
      "External id": 2481546,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 9249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615387.648, "dur": 7.269,
    "args": {
      "External id": 2481547,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 9250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615390.039, "dur": 4.178,
    "args": {
      "External id": 2481548,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 9251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615393.076, "dur": 1.000,
    "args": {
      "External id": 2481549,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 9252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615395.737, "dur": 41.264,
    "args": {
      "External id": 2481550,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615397.040, "dur": 39.361,
    "args": {
      "External id": 2481551,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615442.107, "dur": 3.298,
    "args": {
      "External id": 2481552,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615443.827, "dur": 0.393,
    "args": {
      "External id": 2481553,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "16384"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615451.039, "dur": 1.676,
    "args": {
      "External id": 2481554,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 9257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615461.943, "dur": 6.077,
    "args": {
      "External id": 2481555,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 9258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615463.889, "dur": 3.856,
    "args": {
      "External id": 2481556,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615567.080, "dur": 175.461,
    "args": {
      "External id": 2481557,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615570.004, "dur": 3.220,
    "args": {
      "External id": 2481558,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615574.598, "dur": 167.412,
    "args": {
      "External id": 2481559,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 9262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615577.847, "dur": 0.328,
    "args": {
      "External id": 2481560,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615579.451, "dur": 21.203,
    "args": {
      "External id": 2481561,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615602.069, "dur": 3.709,
    "args": {
      "External id": 2481562,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 9265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615604.182, "dur": 1.354,
    "args": {
      "External id": 2481563,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 9266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615608.666, "dur": 23.618,
    "args": {
      "External id": 2481564,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615610.996, "dur": 1.483,
    "args": {
      "External id": 2481565,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615613.682, "dur": 18.373,
    "args": {
      "External id": 2481566,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 9269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615616.598, "dur": 2.461,
    "args": {
      "External id": 2481567,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615633.860, "dur": 19.666,
    "args": {
      "External id": 2481568,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615654.853, "dur": 12.841,
    "args": {
      "External id": 2481569,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615670.566, "dur": 12.411,
    "args": {
      "External id": 2481570,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 9273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615684.376, "dur": 11.427,
    "args": {
      "External id": 2481571,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615697.702, "dur": 19.913,
    "args": {
      "External id": 2481572,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 9275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615700.928, "dur": 1.689,
    "args": {
      "External id": 2481573,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615704.597, "dur": 0.927,
    "args": {
      "External id": 2481574,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 9277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615718.920, "dur": 10.875,
    "args": {
      "External id": 2481575,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615731.066, "dur": 10.082,
    "args": {
      "External id": 2481576,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615749.109, "dur": 2.081,
    "args": {
      "External id": 2481577,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615759.194, "dur": 3.366,
    "args": {
      "External id": 2481578,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615761.383, "dur": 0.397,
    "args": {
      "External id": 2481579,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "16384"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615829.161, "dur": 54.744,
    "args": {
      "External id": 2481580,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 9283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615888.626, "dur": 6.265,
    "args": {
      "External id": 2481581,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615890.894, "dur": 2.938,
    "args": {
      "External id": 2481582,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "33554432"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615896.326, "dur": 26.083,
    "args": {
      "External id": 2481583,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 9286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615927.026, "dur": 6.430,
    "args": {
      "External id": 2481584,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 9287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615928.194, "dur": 4.642,
    "args": {
      "External id": 2481585,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 9288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615931.830, "dur": 0.802,
    "args": {
      "External id": 2481586,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 9289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615935.716, "dur": 83.412,
    "args": {
      "External id": 2481587,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656615936.772, "dur": 81.247,
    "args": {
      "External id": 2481588,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616024.817, "dur": 30.728,
    "args": {
      "External id": 2481589,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 9292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616061.775, "dur": 4.064,
    "args": {
      "External id": 2481590,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616064.014, "dur": 0.826,
    "args": {
      "External id": 2481591,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "50331648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616069.640, "dur": 66.247,
    "args": {
      "External id": 2481592,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 9295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616072.288, "dur": 4.093,
    "args": {
      "External id": 2481593,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 9296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616073.478, "dur": 2.341,
    "args": {
      "External id": 2481594,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 9297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616075.006, "dur": 0.663,
    "args": {
      "External id": 2481595,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 9298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616077.143, "dur": 58.337,
    "args": {
      "External id": 2481596,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616078.212, "dur": 56.734,
    "args": {
      "External id": 2481597,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616139.651, "dur": 3.479,
    "args": {
      "External id": 2481598,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616141.134, "dur": 0.723,
    "args": {
      "External id": 2481599,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "24576"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616150.894, "dur": 1.405,
    "args": {
      "External id": 2481600,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 9303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616159.414, "dur": 6.164,
    "args": {
      "External id": 2481601,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 9304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616161.005, "dur": 4.315,
    "args": {
      "External id": 2481602,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616249.474, "dur": 168.854,
    "args": {
      "External id": 2481603,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616251.774, "dur": 4.484,
    "args": {
      "External id": 2481604,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616258.764, "dur": 159.044,
    "args": {
      "External id": 2481605,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 9308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616260.531, "dur": 0.305,
    "args": {
      "External id": 2481606,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616261.795, "dur": 19.825,
    "args": {
      "External id": 2481607,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616282.922, "dur": 4.986,
    "args": {
      "External id": 2481608,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 9311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616286.803, "dur": 0.843,
    "args": {
      "External id": 2481609,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 9312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616288.807, "dur": 19.232,
    "args": {
      "External id": 2481610,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616289.798, "dur": 1.630,
    "args": {
      "External id": 2481611,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616292.417, "dur": 15.312,
    "args": {
      "External id": 2481612,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 9315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616294.856, "dur": 2.211,
    "args": {
      "External id": 2481613,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616311.286, "dur": 19.624,
    "args": {
      "External id": 2481614,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616331.957, "dur": 12.075,
    "args": {
      "External id": 2481615,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616346.477, "dur": 12.846,
    "args": {
      "External id": 2481616,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 9319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616360.596, "dur": 11.053,
    "args": {
      "External id": 2481617,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616373.657, "dur": 19.271,
    "args": {
      "External id": 2481618,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 9321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616375.577, "dur": 1.080,
    "args": {
      "External id": 2481619,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616378.490, "dur": 2.424,
    "args": {
      "External id": 2481620,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 9323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616394.158, "dur": 10.589,
    "args": {
      "External id": 2481621,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616407.394, "dur": 9.623,
    "args": {
      "External id": 2481622,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616424.672, "dur": 1.760,
    "args": {
      "External id": 2481623,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616434.194, "dur": 3.346,
    "args": {
      "External id": 2481624,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616436.273, "dur": 0.508,
    "args": {
      "External id": 2481625,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "24576"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616500.165, "dur": 66.279,
    "args": {
      "External id": 2481626,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 9329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616573.054, "dur": 5.631,
    "args": {
      "External id": 2481627,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616576.298, "dur": 1.051,
    "args": {
      "External id": 2481628,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "50331648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616580.271, "dur": 26.167,
    "args": {
      "External id": 2481629,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 9332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616613.347, "dur": 5.287,
    "args": {
      "External id": 2481630,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 9333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616614.595, "dur": 3.367,
    "args": {
      "External id": 2481631,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 9334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616616.541, "dur": 1.228,
    "args": {
      "External id": 2481632,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 9335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616621.341, "dur": 42.764,
    "args": {
      "External id": 2481633,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616622.429, "dur": 41.168,
    "args": {
      "External id": 2481634,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616668.399, "dur": 14.579,
    "args": {
      "External id": 2481635,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 9338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616688.999, "dur": 5.192,
    "args": {
      "External id": 2481636,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616692.576, "dur": 0.770,
    "args": {
      "External id": 2481637,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "67108864"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616697.828, "dur": 46.593,
    "args": {
      "External id": 2481638,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 9341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616698.611, "dur": 2.986,
    "args": {
      "External id": 2481639,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 9342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616699.392, "dur": 1.704,
    "args": {
      "External id": 2481640,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 9343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616700.352, "dur": 0.584,
    "args": {
      "External id": 2481641,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 9344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616702.059, "dur": 41.920,
    "args": {
      "External id": 2481642,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616704.488, "dur": 38.949,
    "args": {
      "External id": 2481643,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616748.757, "dur": 3.593,
    "args": {
      "External id": 2481644,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616750.642, "dur": 0.788,
    "args": {
      "External id": 2481645,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "32768"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616758.192, "dur": 1.351,
    "args": {
      "External id": 2481646,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 9349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616766.724, "dur": 6.458,
    "args": {
      "External id": 2481647,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 9350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616768.674, "dur": 4.202,
    "args": {
      "External id": 2481648,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616858.187, "dur": 244.038,
    "args": {
      "External id": 2481649,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616861.612, "dur": 1.952,
    "args": {
      "External id": 2481650,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616864.913, "dur": 236.740,
    "args": {
      "External id": 2481651,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 9354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616866.147, "dur": 0.464,
    "args": {
      "External id": 2481652,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616867.610, "dur": 19.073,
    "args": {
      "External id": 2481653,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616888.076, "dur": 4.813,
    "args": {
      "External id": 2481654,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 9357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616891.752, "dur": 0.900,
    "args": {
      "External id": 2481655,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 9358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616893.750, "dur": 21.776,
    "args": {
      "External id": 2481656,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616895.176, "dur": 1.546,
    "args": {
      "External id": 2481657,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616897.636, "dur": 17.675,
    "args": {
      "External id": 2481658,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 9361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616902.009, "dur": 2.713,
    "args": {
      "External id": 2481659,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616916.691, "dur": 17.979,
    "args": {
      "External id": 2481660,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616935.708, "dur": 12.383,
    "args": {
      "External id": 2481661,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616950.821, "dur": 11.952,
    "args": {
      "External id": 2481662,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 9365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616964.154, "dur": 10.854,
    "args": {
      "External id": 2481663,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656616976.606, "dur": 97.921,
    "args": {
      "External id": 2481664,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 9367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617051.530, "dur": 3.068,
    "args": {
      "External id": 2481665,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617057.646, "dur": 0.735,
    "args": {
      "External id": 2481666,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 9369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617077.935, "dur": 11.216,
    "args": {
      "External id": 2481667,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617090.254, "dur": 10.278,
    "args": {
      "External id": 2481668,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617110.432, "dur": 2.368,
    "args": {
      "External id": 2481669,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617121.554, "dur": 3.740,
    "args": {
      "External id": 2481670,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617123.961, "dur": 0.509,
    "args": {
      "External id": 2481671,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "32768"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617194.849, "dur": 56.909,
    "args": {
      "External id": 2481672,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 9375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617256.685, "dur": 7.033,
    "args": {
      "External id": 2481673,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617261.235, "dur": 1.448,
    "args": {
      "External id": 2481674,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "67108864"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617265.160, "dur": 23.151,
    "args": {
      "External id": 2481675,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 9378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617292.591, "dur": 5.978,
    "args": {
      "External id": 2481676,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 9379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617294.101, "dur": 3.746,
    "args": {
      "External id": 2481677,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 9380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617296.020, "dur": 1.597,
    "args": {
      "External id": 2481678,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 9381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617300.789, "dur": 38.779,
    "args": {
      "External id": 2481679,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617301.768, "dur": 37.217,
    "args": {
      "External id": 2481680,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617345.387, "dur": 14.177,
    "args": {
      "External id": 2481681,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 9384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617364.837, "dur": 3.023,
    "args": {
      "External id": 2481682,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617366.494, "dur": 0.685,
    "args": {
      "External id": 2481683,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "83886080"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617371.516, "dur": 45.965,
    "args": {
      "External id": 2481684,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 9387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617372.144, "dur": 5.062,
    "args": {
      "External id": 2481685,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 9388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617372.910, "dur": 3.742,
    "args": {
      "External id": 2481686,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 9389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617375.811, "dur": 0.687,
    "args": {
      "External id": 2481687,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 9390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617377.690, "dur": 39.440,
    "args": {
      "External id": 2481688,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617378.175, "dur": 38.320,
    "args": {
      "External id": 2481689,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617421.523, "dur": 3.616,
    "args": {
      "External id": 2481690,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617423.381, "dur": 0.611,
    "args": {
      "External id": 2481691,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "40960"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617430.457, "dur": 1.281,
    "args": {
      "External id": 2481692,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 9395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617440.478, "dur": 5.958,
    "args": {
      "External id": 2481693,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 9396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617442.270, "dur": 3.889,
    "args": {
      "External id": 2481694,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617522.584, "dur": 191.133,
    "args": {
      "External id": 2481695,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617526.262, "dur": 18.735,
    "args": {
      "External id": 2481696,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617547.198, "dur": 165.960,
    "args": {
      "External id": 2481697,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 9400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617548.622, "dur": 0.352,
    "args": {
      "External id": 2481698,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617549.905, "dur": 21.584,
    "args": {
      "External id": 2481699,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617572.801, "dur": 5.577,
    "args": {
      "External id": 2481700,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 9403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617577.169, "dur": 0.903,
    "args": {
      "External id": 2481701,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 9404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617580.836, "dur": 22.112,
    "args": {
      "External id": 2481702,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617581.882, "dur": 1.663,
    "args": {
      "External id": 2481703,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617584.488, "dur": 18.207,
    "args": {
      "External id": 2481704,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 9407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617587.055, "dur": 2.454,
    "args": {
      "External id": 2481705,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617604.214, "dur": 21.804,
    "args": {
      "External id": 2481706,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617627.100, "dur": 13.330,
    "args": {
      "External id": 2481707,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617643.141, "dur": 12.310,
    "args": {
      "External id": 2481708,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 9411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617656.762, "dur": 10.869,
    "args": {
      "External id": 2481709,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617669.338, "dur": 19.598,
    "args": {
      "External id": 2481710,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 9413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617672.910, "dur": 1.581,
    "args": {
      "External id": 2481711,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617676.462, "dur": 0.803,
    "args": {
      "External id": 2481712,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 9415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617690.014, "dur": 10.628,
    "args": {
      "External id": 2481713,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617701.541, "dur": 10.723,
    "args": {
      "External id": 2481714,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617720.817, "dur": 2.088,
    "args": {
      "External id": 2481715,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617731.553, "dur": 3.422,
    "args": {
      "External id": 2481716,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617733.717, "dur": 0.478,
    "args": {
      "External id": 2481717,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "40960"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617797.737, "dur": 51.102,
    "args": {
      "External id": 2481718,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 9421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617853.752, "dur": 3.708,
    "args": {
      "External id": 2481719,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617855.889, "dur": 0.667,
    "args": {
      "External id": 2481720,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "83886080"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617858.837, "dur": 22.382,
    "args": {
      "External id": 2481721,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 9424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617885.102, "dur": 6.881,
    "args": {
      "External id": 2481722,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 9425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617886.289, "dur": 5.111,
    "args": {
      "External id": 2481723,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 9426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617890.024, "dur": 1.192,
    "args": {
      "External id": 2481724,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 9427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617894.224, "dur": 38.148,
    "args": {
      "External id": 2481725,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617895.243, "dur": 36.649,
    "args": {
      "External id": 2481726,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617936.168, "dur": 13.412,
    "args": {
      "External id": 2481727,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 9430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617954.351, "dur": 3.452,
    "args": {
      "External id": 2481728,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617956.163, "dur": 0.883,
    "args": {
      "External id": 2481729,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "100663296"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617960.930, "dur": 84.713,
    "args": {
      "External id": 2481730,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 9433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617963.626, "dur": 3.250,
    "args": {
      "External id": 2481731,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 9434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617964.330, "dur": 2.026,
    "args": {
      "External id": 2481732,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 9435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617965.663, "dur": 0.557,
    "args": {
      "External id": 2481733,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 9436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617967.356, "dur": 77.722,
    "args": {
      "External id": 2481734,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656617967.789, "dur": 76.280,
    "args": {
      "External id": 2481735,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618052.166, "dur": 4.749,
    "args": {
      "External id": 2481736,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618054.864, "dur": 0.714,
    "args": {
      "External id": 2481737,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "49152"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618063.215, "dur": 1.551,
    "args": {
      "External id": 2481738,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 9441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618073.276, "dur": 7.919,
    "args": {
      "External id": 2481739,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 9442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618075.069, "dur": 5.830,
    "args": {
      "External id": 2481740,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618160.170, "dur": 171.075,
    "args": {
      "External id": 2481741,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618162.633, "dur": 2.134,
    "args": {
      "External id": 2481742,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618165.944, "dur": 164.857,
    "args": {
      "External id": 2481743,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 9446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618167.728, "dur": 0.445,
    "args": {
      "External id": 2481744,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618171.103, "dur": 22.203,
    "args": {
      "External id": 2481745,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618194.731, "dur": 2.847,
    "args": {
      "External id": 2481746,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 9449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618196.739, "dur": 0.616,
    "args": {
      "External id": 2481747,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 9450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618205.346, "dur": 21.568,
    "args": {
      "External id": 2481748,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618206.141, "dur": 2.893,
    "args": {
      "External id": 2481749,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618210.178, "dur": 16.085,
    "args": {
      "External id": 2481750,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 9453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618212.654, "dur": 2.533,
    "args": {
      "External id": 2481751,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618228.114, "dur": 18.176,
    "args": {
      "External id": 2481752,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618247.391, "dur": 12.315,
    "args": {
      "External id": 2481753,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618263.738, "dur": 11.617,
    "args": {
      "External id": 2481754,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 9457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618276.548, "dur": 10.925,
    "args": {
      "External id": 2481755,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618289.160, "dur": 17.093,
    "args": {
      "External id": 2481756,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 9459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618290.479, "dur": 1.171,
    "args": {
      "External id": 2481757,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618293.501, "dur": 0.780,
    "args": {
      "External id": 2481758,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 9461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618307.478, "dur": 10.643,
    "args": {
      "External id": 2481759,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618319.018, "dur": 10.943,
    "args": {
      "External id": 2481760,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618338.920, "dur": 1.539,
    "args": {
      "External id": 2481761,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618347.450, "dur": 3.463,
    "args": {
      "External id": 2481762,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618349.586, "dur": 0.514,
    "args": {
      "External id": 2481763,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "49152"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618408.608, "dur": 48.655,
    "args": {
      "External id": 2481764,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 9467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618461.948, "dur": 3.543,
    "args": {
      "External id": 2481765,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618463.838, "dur": 0.782,
    "args": {
      "External id": 2481766,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "100663296"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618469.329, "dur": 20.727,
    "args": {
      "External id": 2481767,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 9470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618496.270, "dur": 8.377,
    "args": {
      "External id": 2481768,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 9471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618497.525, "dur": 6.390,
    "args": {
      "External id": 2481769,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 9472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618501.231, "dur": 2.424,
    "args": {
      "External id": 2481770,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 9473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618506.945, "dur": 54.689,
    "args": {
      "External id": 2481771,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618507.923, "dur": 52.820,
    "args": {
      "External id": 2481772,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618567.171, "dur": 14.183,
    "args": {
      "External id": 2481773,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 9476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618587.220, "dur": 6.225,
    "args": {
      "External id": 2481774,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618591.528, "dur": 0.928,
    "args": {
      "External id": 2481775,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "117440512"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618597.132, "dur": 51.117,
    "args": {
      "External id": 2481776,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 9479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618597.994, "dur": 4.020,
    "args": {
      "External id": 2481777,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 9480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618598.885, "dur": 2.564,
    "args": {
      "External id": 2481778,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 9481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618600.338, "dur": 0.979,
    "args": {
      "External id": 2481779,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 9482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618602.563, "dur": 45.261,
    "args": {
      "External id": 2481780,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618605.245, "dur": 42.064,
    "args": {
      "External id": 2481781,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618651.947, "dur": 3.505,
    "args": {
      "External id": 2481782,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618653.729, "dur": 0.641,
    "args": {
      "External id": 2481783,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "57344"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618661.076, "dur": 1.455,
    "args": {
      "External id": 2481784,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 9487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618669.352, "dur": 7.473,
    "args": {
      "External id": 2481785,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 9488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618672.889, "dur": 3.657,
    "args": {
      "External id": 2481786,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618749.635, "dur": 163.759,
    "args": {
      "External id": 2481787,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618751.495, "dur": 2.293,
    "args": {
      "External id": 2481788,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618756.606, "dur": 156.223,
    "args": {
      "External id": 2481789,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 9492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618757.843, "dur": 0.430,
    "args": {
      "External id": 2481790,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618759.182, "dur": 20.733,
    "args": {
      "External id": 2481791,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618781.233, "dur": 4.320,
    "args": {
      "External id": 2481792,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 9495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618783.148, "dur": 2.111,
    "args": {
      "External id": 2481793,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 9496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618786.571, "dur": 20.126,
    "args": {
      "External id": 2481794,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618788.964, "dur": 1.504,
    "args": {
      "External id": 2481795,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618791.304, "dur": 15.074,
    "args": {
      "External id": 2481796,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 9499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618793.567, "dur": 2.145,
    "args": {
      "External id": 2481797,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618807.647, "dur": 17.917,
    "args": {
      "External id": 2481798,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618826.586, "dur": 12.057,
    "args": {
      "External id": 2481799,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618840.683, "dur": 12.563,
    "args": {
      "External id": 2481800,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 9503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618854.626, "dur": 10.985,
    "args": {
      "External id": 2481801,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618867.274, "dur": 20.847,
    "args": {
      "External id": 2481802,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 9505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618869.128, "dur": 1.253,
    "args": {
      "External id": 2481803,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618873.992, "dur": 1.955,
    "args": {
      "External id": 2481804,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 9507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618889.374, "dur": 10.424,
    "args": {
      "External id": 2481805,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618900.669, "dur": 11.172,
    "args": {
      "External id": 2481806,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618918.817, "dur": 1.677,
    "args": {
      "External id": 2481807,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618927.034, "dur": 3.456,
    "args": {
      "External id": 2481808,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656618929.220, "dur": 0.429,
    "args": {
      "External id": 2481809,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "57344"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619033.620, "dur": 54.385,
    "args": {
      "External id": 2481810,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 9513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619094.581, "dur": 6.354,
    "args": {
      "External id": 2481811,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619098.588, "dur": 0.939,
    "args": {
      "External id": 2481812,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "117440512"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619102.261, "dur": 24.479,
    "args": {
      "External id": 2481813,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 9516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619131.403, "dur": 5.290,
    "args": {
      "External id": 2481814,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 9517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619132.773, "dur": 3.341,
    "args": {
      "External id": 2481815,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 9518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619134.593, "dur": 1.296,
    "args": {
      "External id": 2481816,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 9519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619139.053, "dur": 44.491,
    "args": {
      "External id": 2481817,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619141.892, "dur": 41.178,
    "args": {
      "External id": 2481818,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619187.179, "dur": 14.018,
    "args": {
      "External id": 2481819,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 9522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619205.560, "dur": 21.865,
    "args": {
      "External id": 2481820,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 9523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619207.792, "dur": 19.266,
    "args": {
      "External id": 2481821,"Record function id": 0, "Concrete Inputs": ["", "[]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619212.749, "dur": 0.713,
    "args": {
      "External id": 2481822,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619232.394, "dur": 31.591,
    "args": {
      "External id": 2481823,"Record function id": 0, "Concrete Inputs": ["", "", "15", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[32000, 2048], [], [], [], [], []], "Ev Idx": 9526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619233.971, "dur": 29.795,
    "args": {
      "External id": 2481824,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], [], []], "Input Dims": [[32000, 2048], [], [], [], [], [], []], "Ev Idx": 9527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619238.385, "dur": 3.815,
    "args": {
      "External id": 2481825,"Record function id": 0, "Concrete Inputs": ["[32000, 2048]", "[2048, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619243.176, "dur": 20.141,
    "args": {
      "External id": 2481826,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "float", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 9529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619276.427, "dur": 6.176,
    "args": {
      "External id": 2481827,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 9530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619279.939, "dur": 2.391,
    "args": {
      "External id": 2481828,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 9531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619283.792, "dur": 1.280,
    "args": {
      "External id": 2481829,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 9532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619284.479, "dur": 0.500,
    "args": {
      "External id": 2481830,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 9533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619323.277, "dur": 20.572,
    "args": {
      "External id": 2481831,"Sequence number": 24740632, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1], [], []], "Ev Idx": 9534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619345.709, "dur": 12.032,
    "args": {
      "External id": 2481832,"Sequence number": 24740633, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1], [], []], "Ev Idx": 9535
    }
  },
  {
    "ph": "s", "id": 16, "pid": 1336759, "tid": 1336759, "ts": 1514656619345.709,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619363.594, "dur": 5.302,
    "args": {
      "External id": 2481833,"Sequence number": 24740634, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], [], []], "Input Dims": [[16, 4, 4096], [], [], [], []], "Ev Idx": 9536
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619366.483, "dur": 1.147,
    "args": {
      "External id": 2481834,"Record function id": 0, "Concrete Inputs": ["", "[16, 4, 4096]", "[8192, 1, 1]", "1"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[16, 4, 4096], [], [], []], "Ev Idx": 9537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619373.171, "dur": 5.970,
    "args": {
      "External id": 2481835,"Sequence number": 24740634, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "2"], "Input type": ["long int", "Scalar", "Scalar"], "Input Strides": [[8192, 1, 1], [], []], "Input Dims": [[16, 4, 4096], [], []], "Ev Idx": 9538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619376.955, "dur": 0.962,
    "args": {
      "External id": 2481836,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096]", "[8192, 1]", "3"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[16, 4, 4096], [], [], []], "Ev Idx": 9539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619380.485, "dur": 2.645,
    "args": {
      "External id": 2481837,"Sequence number": 24740634, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], [], [], []], "Input Dims": [[16, 4096], [], [], [], []], "Ev Idx": 9540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619381.990, "dur": 0.544,
    "args": {
      "External id": 2481838,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096]", "[8192, 1]", "3"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1], [], [], []], "Input Dims": [[16, 4096], [], [], []], "Ev Idx": 9541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619387.080, "dur": 6.986,
    "args": {
      "External id": 2481839,"Sequence number": 24740634, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], [], []], "Ev Idx": 9542
    }
  },
  {
    "ph": "s", "id": 15, "pid": 1336759, "tid": 1336759, "ts": 1514656619387.080,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619391.913, "dur": 0.896,
    "args": {
      "External id": 2481840,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 2048]", "[33554432, 8192, 2048, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], []], "Ev Idx": 9543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619394.998, "dur": 4.586,
    "args": {
      "External id": 2481841,"Sequence number": 24740635, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], [], []], "Ev Idx": 9544
    }
  },
  {
    "ph": "s", "id": 14, "pid": 1336759, "tid": 1336759, "ts": 1514656619394.998,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619398.161, "dur": 0.686,
    "args": {
      "External id": 2481842,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 2048]", "[33554432, 8192, 2048, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], []], "Ev Idx": 9545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619400.617, "dur": 6.309,
    "args": {
      "External id": 2481843,"Sequence number": 24740636, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2", "2"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], []], "Input Dims": [[16, 4096, 4, 2048], [], []], "Ev Idx": 9546
    }
  },
  {
    "ph": "s", "id": 13, "pid": 1336759, "tid": 1336759, "ts": 1514656619400.617,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619405.064, "dur": 1.015,
    "args": {
      "External id": 2481844,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]", "[33554432, 8192, 1]", "4096"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], []], "Ev Idx": 9547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619407.947, "dur": 4.592,
    "args": {
      "External id": 2481845,"Sequence number": 24740637, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 1], [], [], [], []], "Input Dims": [[16, 4096, 2048], [], [], [], []], "Ev Idx": 9548
    }
  },
  {
    "ph": "s", "id": 12, "pid": 1336759, "tid": 1336759, "ts": 1514656619407.947,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619410.915, "dur": 0.856,
    "args": {
      "External id": 2481846,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]", "[33554432, 8192, 1]", "4096"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 1], [], [], []], "Input Dims": [[16, 4096, 2048], [], [], []], "Ev Idx": 9549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619416.127, "dur": 27.811,
    "args": {
      "External id": 2481847,"Sequence number": 24740638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["long int", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[16, 4096], []], "Ev Idx": 9550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619417.599, "dur": 26.092,
    "args": {
      "External id": 2481848,"Sequence number": 24740638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["long int", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[16, 4096], []], "Ev Idx": 9551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619419.939, "dur": 5.151,
    "args": {
      "External id": 2481849,"Record function id": 0, "Concrete Inputs": ["", "4", "0", "", "", "0"], "Input type": ["long int", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[8192, 1], [], [], [], [], []], "Input Dims": [[16, 4096], [], [], [], [], []], "Ev Idx": 9552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619421.927, "dur": 2.640,
    "args": {
      "External id": 2481850,"Record function id": 0, "Concrete Inputs": ["[16, 4096]", "4", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619425.808, "dur": 17.450,
    "args": {
      "External id": 2481851,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[4096, 1], [8192, 1], []], "Input Dims": [[16, 4096], [16, 4096], []], "Ev Idx": 9554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619469.383, "dur": 5.872,
    "args": {
      "External id": 2481852,"Sequence number": 24740638, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[33554432, 8192, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 9555
    }
  },
  {
    "ph": "s", "id": 11, "pid": 1336759, "tid": 1336759, "ts": 1514656619469.383,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619477.737, "dur": 1.053,
    "args": {
      "External id": 2481853,"Sequence number": 24740639, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["long int", "ScalarList"], "Input Strides": [[4096, 1], []], "Input Dims": [[16, 4096], []], "Ev Idx": 9556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FusedLinearCrossEntropyFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619511.561, "dur": 44233.643,
    "args": {
      "External id": 2481854,"Sequence number": 24740639, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "-100", "0.", "1.", "8"], "Input type": ["c10::BFloat16", "long int", "c10::BFloat16", "", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [1], [2048, 1], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536], [32000, 2048], [], [], [], [], []], "Ev Idx": 9557
    }
  },
  {
    "ph": "s", "id": 10, "pid": 1336759, "tid": 1336759, "ts": 1514656619511.561,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619525.794, "dur": 44.980,
    "args": {
      "External id": 2481855,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 9558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619526.472, "dur": 44.113,
    "args": {
      "External id": 2481856,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 9559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619545.199, "dur": 6.898,
    "args": {
      "External id": 2481857,"Record function id": 0, "Concrete Inputs": ["", "15", "0", "", "", "0"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[8192, 1], [], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], [], []], "Ev Idx": 9560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619547.046, "dur": 4.427,
    "args": {
      "External id": 2481858,"Record function id": 0, "Concrete Inputs": ["[65536, 2048]", "15", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619552.773, "dur": 17.417,
    "args": {
      "External id": 2481859,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [8192, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 9562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619588.262, "dur": 22.266,
    "args": {
      "External id": 2481860,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], [], []], "Ev Idx": 9563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619589.357, "dur": 4.929,
    "args": {
      "External id": 2481861,"Record function id": 0, "Concrete Inputs": ["", "15", "0", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], [], []], "Ev Idx": 9564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619590.826, "dur": 3.166,
    "args": {
      "External id": 2481862,"Record function id": 0, "Concrete Inputs": ["[65536, 2048]", "[2048, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619595.508, "dur": 14.770,
    "args": {
      "External id": 2481863,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 9566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619597.141, "dur": 12.498,
    "args": {
      "External id": 2481864,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 9567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619613.916, "dur": 18.393,
    "args": {
      "External id": 2481865,"Record function id": 0, "Concrete Inputs": ["", "6", "", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[32000, 2048], [], [], [], [], []], "Ev Idx": 9568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619614.472, "dur": 5.431,
    "args": {
      "External id": 2481866,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[32000, 2048], [], [], [], [], []], "Ev Idx": 9569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619617.183, "dur": 2.489,
    "args": {
      "External id": 2481867,"Record function id": 0, "Concrete Inputs": ["[32000, 2048]", "[2048, 1]", "6", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619620.408, "dur": 11.678,
    "args": {
      "External id": 2481868,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 9571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619620.919, "dur": 10.729,
    "args": {
      "External id": 2481869,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[2048, 1], []], "Input Dims": [[32000, 2048], []], "Ev Idx": 9572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619637.842, "dur": 17.227,
    "args": {
      "External id": 2481870,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 9573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619641.741, "dur": 2.622,
    "args": {
      "External id": 2481871,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619644.858, "dur": 9.918,
    "args": {
      "External id": 2481872,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[65536]], "Ev Idx": 9575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619645.380, "dur": 9.103,
    "args": {
      "External id": 2481873,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 9576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::ne", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619661.876, "dur": 21.488,
    "args": {
      "External id": 2481874,"Record function id": 0, "Concrete Inputs": ["", "-100"], "Input type": ["long int", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 9577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619685.859, "dur": 44.330,
    "args": {
      "External id": 2481875,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["bool", ""], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 9578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619687.796, "dur": 41.974,
    "args": {
      "External id": 2481876,"Record function id": 0, "Concrete Inputs": ["", "[]", "False", ""], "Input type": ["bool", "ScalarList", "Scalar", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619692.437, "dur": 1.132,
    "args": {
      "External id": 2481877,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["long int", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619695.108, "dur": 20.534,
    "args": {
      "External id": 2481878,"Record function id": 0, "Concrete Inputs": ["", "4", "False", "False", ""], "Input type": ["bool", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619696.751, "dur": 18.687,
    "args": {
      "External id": 2481879,"Record function id": 0, "Concrete Inputs": ["", "4", "", "", "", "False", ""], "Input type": ["bool", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[65536], [], [], [], [], [], []], "Ev Idx": 9582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619700.538, "dur": 2.680,
    "args": {
      "External id": 2481880,"Record function id": 0, "Concrete Inputs": ["[65536]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619703.851, "dur": 11.277,
    "args": {
      "External id": 2481881,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[65536], [65536], []], "Ev Idx": 9584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619733.637, "dur": 38611.698,
    "args": {
      "External id": 2481882,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1336759, "tid": 1336759,
    "ts": 1514656619735.136, "dur": 38608.337,
    "args": {
      "External id": 2481883,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 9586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658358.254, "dur": 8.548,
    "args": {
      "External id": 2481884,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658363.835, "dur": 0.996,
    "args": {
      "External id": 2481885,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658371.735, "dur": 99.588,
    "args": {
      "External id": 2481886,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 9589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658373.200, "dur": 5.817,
    "args": {
      "External id": 2481887,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 9590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658375.200, "dur": 2.906,
    "args": {
      "External id": 2481888,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 9591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658377.004, "dur": 0.855,
    "args": {
      "External id": 2481889,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 9592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658380.378, "dur": 90.274,
    "args": {
      "External id": 2481890,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658383.648, "dur": 86.138,
    "args": {
      "External id": 2481891,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658474.949, "dur": 4.119,
    "args": {
      "External id": 2481892,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658476.922, "dur": 0.656,
    "args": {
      "External id": 2481893,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "0"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658486.225, "dur": 2.582,
    "args": {
      "External id": 2481894,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 9597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658496.953, "dur": 6.483,
    "args": {
      "External id": 2481895,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 9598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658499.083, "dur": 3.993,
    "args": {
      "External id": 2481896,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658647.411, "dur": 187.499,
    "args": {
      "External id": 2481897,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658650.553, "dur": 3.191,
    "args": {
      "External id": 2481898,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658657.046, "dur": 177.415,
    "args": {
      "External id": 2481899,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 9602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658658.548, "dur": 0.422,
    "args": {
      "External id": 2481900,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658660.447, "dur": 24.773,
    "args": {
      "External id": 2481901,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658687.227, "dur": 5.368,
    "args": {
      "External id": 2481902,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 9605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658689.915, "dur": 2.305,
    "args": {
      "External id": 2481903,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 9606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658693.622, "dur": 24.096,
    "args": {
      "External id": 2481904,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658694.673, "dur": 1.379,
    "args": {
      "External id": 2481905,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658699.390, "dur": 18.014,
    "args": {
      "External id": 2481906,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 9609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658702.947, "dur": 3.225,
    "args": {
      "External id": 2481907,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658719.434, "dur": 22.465,
    "args": {
      "External id": 2481908,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658743.485, "dur": 13.225,
    "args": {
      "External id": 2481909,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658759.582, "dur": 14.721,
    "args": {
      "External id": 2481910,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 9613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658775.676, "dur": 11.950,
    "args": {
      "External id": 2481911,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658789.315, "dur": 20.116,
    "args": {
      "External id": 2481912,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 9615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658791.395, "dur": 1.353,
    "args": {
      "External id": 2481913,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658796.187, "dur": 0.678,
    "args": {
      "External id": 2481914,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 9617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658810.804, "dur": 11.145,
    "args": {
      "External id": 2481915,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658823.143, "dur": 10.285,
    "args": {
      "External id": 2481916,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658842.570, "dur": 2.249,
    "args": {
      "External id": 2481917,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658851.391, "dur": 3.621,
    "args": {
      "External id": 2481918,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658853.462, "dur": 0.643,
    "args": {
      "External id": 2481919,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656658926.596, "dur": 101.144,
    "args": {
      "External id": 2481920,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 9623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659035.562, "dur": 10.503,
    "args": {
      "External id": 2481921,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659040.786, "dur": 2.623,
    "args": {
      "External id": 2481922,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659047.661, "dur": 26.812,
    "args": {
      "External id": 2481923,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 9626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659080.155, "dur": 5.836,
    "args": {
      "External id": 2481924,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 9627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659081.949, "dur": 3.360,
    "args": {
      "External id": 2481925,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 9628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659083.988, "dur": 1.052,
    "args": {
      "External id": 2481926,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 9629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659089.111, "dur": 49.326,
    "args": {
      "External id": 2481927,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659092.160, "dur": 45.631,
    "args": {
      "External id": 2481928,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659142.804, "dur": 16.000,
    "args": {
      "External id": 2481929,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 9632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659165.162, "dur": 4.008,
    "args": {
      "External id": 2481930,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659167.564, "dur": 0.722,
    "args": {
      "External id": 2481931,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "16777216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659173.339, "dur": 48.548,
    "args": {
      "External id": 2481932,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 9635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659174.089, "dur": 5.370,
    "args": {
      "External id": 2481933,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 9636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659174.796, "dur": 4.056,
    "args": {
      "External id": 2481934,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 9637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659178.018, "dur": 0.705,
    "args": {
      "External id": 2481935,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 9638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659180.021, "dur": 41.435,
    "args": {
      "External id": 2481936,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659180.649, "dur": 40.373,
    "args": {
      "External id": 2481937,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659226.146, "dur": 3.784,
    "args": {
      "External id": 2481938,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659228.113, "dur": 0.774,
    "args": {
      "External id": 2481939,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "8192"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659236.350, "dur": 1.636,
    "args": {
      "External id": 2481940,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 9643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659247.720, "dur": 6.583,
    "args": {
      "External id": 2481941,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 9644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659249.720, "dur": 4.322,
    "args": {
      "External id": 2481942,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659348.572, "dur": 170.587,
    "args": {
      "External id": 2481943,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659350.590, "dur": 3.884,
    "args": {
      "External id": 2481944,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659355.914, "dur": 162.743,
    "args": {
      "External id": 2481945,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 9648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659357.188, "dur": 0.356,
    "args": {
      "External id": 2481946,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659360.280, "dur": 19.922,
    "args": {
      "External id": 2481947,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659381.826, "dur": 4.986,
    "args": {
      "External id": 2481948,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 9651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659385.664, "dur": 0.921,
    "args": {
      "External id": 2481949,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 9652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659387.691, "dur": 19.288,
    "args": {
      "External id": 2481950,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659388.693, "dur": 1.611,
    "args": {
      "External id": 2481951,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659391.414, "dur": 15.324,
    "args": {
      "External id": 2481952,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 9655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659394.104, "dur": 2.560,
    "args": {
      "External id": 2481953,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659408.402, "dur": 19.957,
    "args": {
      "External id": 2481954,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659429.719, "dur": 12.803,
    "args": {
      "External id": 2481955,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659445.107, "dur": 12.112,
    "args": {
      "External id": 2481956,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 9659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659460.249, "dur": 12.089,
    "args": {
      "External id": 2481957,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659474.026, "dur": 19.233,
    "args": {
      "External id": 2481958,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 9661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659475.746, "dur": 1.042,
    "args": {
      "External id": 2481959,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659478.990, "dur": 2.131,
    "args": {
      "External id": 2481960,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 9663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659494.711, "dur": 11.756,
    "args": {
      "External id": 2481961,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659507.300, "dur": 10.394,
    "args": {
      "External id": 2481962,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659526.026, "dur": 18.452,
    "args": {
      "External id": 2481963,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659556.824, "dur": 4.033,
    "args": {
      "External id": 2481964,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659559.627, "dur": 0.478,
    "args": {
      "External id": 2481965,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "8192"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659633.756, "dur": 53.255,
    "args": {
      "External id": 2481966,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 9669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659692.478, "dur": 4.409,
    "args": {
      "External id": 2481967,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659694.913, "dur": 0.786,
    "args": {
      "External id": 2481968,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "16777216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659698.206, "dur": 23.054,
    "args": {
      "External id": 2481969,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 9672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659725.784, "dur": 6.883,
    "args": {
      "External id": 2481970,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 9673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659727.079, "dur": 4.960,
    "args": {
      "External id": 2481971,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 9674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659730.437, "dur": 1.407,
    "args": {
      "External id": 2481972,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 9675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659735.054, "dur": 39.526,
    "args": {
      "External id": 2481973,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659736.216, "dur": 37.885,
    "args": {
      "External id": 2481974,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659778.226, "dur": 14.783,
    "args": {
      "External id": 2481975,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 9678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659798.244, "dur": 3.237,
    "args": {
      "External id": 2481976,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659800.073, "dur": 0.672,
    "args": {
      "External id": 2481977,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "33554432"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659806.715, "dur": 44.938,
    "args": {
      "External id": 2481978,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 9681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659807.378, "dur": 3.340,
    "args": {
      "External id": 2481979,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 9682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659807.931, "dur": 2.264,
    "args": {
      "External id": 2481980,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 9683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659809.292, "dur": 0.773,
    "args": {
      "External id": 2481981,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 9684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659811.267, "dur": 40.005,
    "args": {
      "External id": 2481982,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659811.779, "dur": 38.957,
    "args": {
      "External id": 2481983,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659868.875, "dur": 3.760,
    "args": {
      "External id": 2481984,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659870.847, "dur": 0.693,
    "args": {
      "External id": 2481985,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "16384"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659880.283, "dur": 1.542,
    "args": {
      "External id": 2481986,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 9689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659889.056, "dur": 7.503,
    "args": {
      "External id": 2481987,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 9690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659890.748, "dur": 5.528,
    "args": {
      "External id": 2481988,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659976.215, "dur": 211.558,
    "args": {
      "External id": 2481989,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656659978.317, "dur": 2.172,
    "args": {
      "External id": 2481990,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660016.494, "dur": 170.727,
    "args": {
      "External id": 2481991,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 9694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660020.252, "dur": 0.519,
    "args": {
      "External id": 2481992,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660021.880, "dur": 25.045,
    "args": {
      "External id": 2481993,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660048.485, "dur": 3.394,
    "args": {
      "External id": 2481994,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 9697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660050.798, "dur": 0.777,
    "args": {
      "External id": 2481995,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 9698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660052.866, "dur": 22.329,
    "args": {
      "External id": 2481996,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660053.902, "dur": 1.698,
    "args": {
      "External id": 2481997,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660056.711, "dur": 18.150,
    "args": {
      "External id": 2481998,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 9701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660061.186, "dur": 2.801,
    "args": {
      "External id": 2481999,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660076.472, "dur": 21.960,
    "args": {
      "External id": 2482000,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660099.582, "dur": 12.052,
    "args": {
      "External id": 2482001,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660114.283, "dur": 12.580,
    "args": {
      "External id": 2482002,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 9705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660129.970, "dur": 11.359,
    "args": {
      "External id": 2482003,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660143.196, "dur": 18.217,
    "args": {
      "External id": 2482004,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 9707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660144.839, "dur": 1.407,
    "args": {
      "External id": 2482005,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660148.354, "dur": 0.656,
    "args": {
      "External id": 2482006,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 9709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660162.596, "dur": 10.838,
    "args": {
      "External id": 2482007,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660174.423, "dur": 10.350,
    "args": {
      "External id": 2482008,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660195.312, "dur": 2.252,
    "args": {
      "External id": 2482009,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660206.509, "dur": 3.553,
    "args": {
      "External id": 2482010,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660208.734, "dur": 0.576,
    "args": {
      "External id": 2482011,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "16384"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660276.760, "dur": 52.916,
    "args": {
      "External id": 2482012,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 9715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660334.780, "dur": 3.725,
    "args": {
      "External id": 2482013,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660336.940, "dur": 0.645,
    "args": {
      "External id": 2482014,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "33554432"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660340.033, "dur": 24.149,
    "args": {
      "External id": 2482015,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 9718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660368.767, "dur": 7.192,
    "args": {
      "External id": 2482016,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 9719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660371.836, "dur": 3.368,
    "args": {
      "External id": 2482017,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 9720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660373.831, "dur": 1.179,
    "args": {
      "External id": 2482018,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 9721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660381.710, "dur": 68.017,
    "args": {
      "External id": 2482019,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660382.636, "dur": 66.501,
    "args": {
      "External id": 2482020,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660453.349, "dur": 39.039,
    "args": {
      "External id": 2482021,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 9724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660498.290, "dur": 4.656,
    "args": {
      "External id": 2482022,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660501.635, "dur": 0.597,
    "args": {
      "External id": 2482023,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "50331648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660506.347, "dur": 78.721,
    "args": {
      "External id": 2482024,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 9727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660507.029, "dur": 3.283,
    "args": {
      "External id": 2482025,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 9728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660507.600, "dur": 2.191,
    "args": {
      "External id": 2482026,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 9729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660508.956, "dur": 0.691,
    "args": {
      "External id": 2482027,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 9730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660510.983, "dur": 73.450,
    "args": {
      "External id": 2482028,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660511.527, "dur": 72.044,
    "args": {
      "External id": 2482029,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660592.399, "dur": 6.058,
    "args": {
      "External id": 2482030,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660594.779, "dur": 2.246,
    "args": {
      "External id": 2482031,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "24576"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660605.018, "dur": 1.453,
    "args": {
      "External id": 2482032,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 9735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660614.160, "dur": 6.393,
    "args": {
      "External id": 2482033,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 9736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660616.044, "dur": 4.236,
    "args": {
      "External id": 2482034,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660703.512, "dur": 166.158,
    "args": {
      "External id": 2482035,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660705.626, "dur": 2.241,
    "args": {
      "External id": 2482036,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660709.261, "dur": 159.901,
    "args": {
      "External id": 2482037,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 9740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660710.663, "dur": 0.361,
    "args": {
      "External id": 2482038,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660712.072, "dur": 21.748,
    "args": {
      "External id": 2482039,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660735.544, "dur": 3.429,
    "args": {
      "External id": 2482040,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 9743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660737.769, "dur": 0.910,
    "args": {
      "External id": 2482041,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 9744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660739.813, "dur": 22.799,
    "args": {
      "External id": 2482042,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660740.701, "dur": 1.556,
    "args": {
      "External id": 2482043,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660744.953, "dur": 17.352,
    "args": {
      "External id": 2482044,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 9747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660749.452, "dur": 2.182,
    "args": {
      "External id": 2482045,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660763.887, "dur": 19.407,
    "args": {
      "External id": 2482046,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660784.429, "dur": 12.043,
    "args": {
      "External id": 2482047,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660799.339, "dur": 12.408,
    "args": {
      "External id": 2482048,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 9751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660812.889, "dur": 11.215,
    "args": {
      "External id": 2482049,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660825.948, "dur": 17.246,
    "args": {
      "External id": 2482050,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 9753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660827.406, "dur": 1.182,
    "args": {
      "External id": 2482051,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660830.569, "dur": 0.781,
    "args": {
      "External id": 2482052,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 9755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660845.841, "dur": 10.947,
    "args": {
      "External id": 2482053,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660857.553, "dur": 10.755,
    "args": {
      "External id": 2482054,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660875.994, "dur": 1.620,
    "args": {
      "External id": 2482055,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660884.567, "dur": 3.480,
    "args": {
      "External id": 2482056,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660886.638, "dur": 0.585,
    "args": {
      "External id": 2482057,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "24576"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656660952.262, "dur": 90.487,
    "args": {
      "External id": 2482058,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 9761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661049.876, "dur": 7.842,
    "args": {
      "External id": 2482059,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661054.841, "dur": 1.419,
    "args": {
      "External id": 2482060,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "50331648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661059.117, "dur": 24.280,
    "args": {
      "External id": 2482061,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 9764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661088.643, "dur": 5.146,
    "args": {
      "External id": 2482062,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 9765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661090.169, "dur": 2.926,
    "args": {
      "External id": 2482063,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 9766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661092.209, "dur": 0.689,
    "args": {
      "External id": 2482064,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 9767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661096.809, "dur": 43.033,
    "args": {
      "External id": 2482065,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661098.033, "dur": 41.073,
    "args": {
      "External id": 2482066,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661146.024, "dur": 14.125,
    "args": {
      "External id": 2482067,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 9770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661166.388, "dur": 3.370,
    "args": {
      "External id": 2482068,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661168.282, "dur": 0.770,
    "args": {
      "External id": 2482069,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "67108864"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661173.547, "dur": 49.541,
    "args": {
      "External id": 2482070,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 9773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661174.207, "dur": 6.817,
    "args": {
      "External id": 2482071,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 9774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661174.806, "dur": 5.598,
    "args": {
      "External id": 2482072,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 9775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661179.331, "dur": 0.945,
    "args": {
      "External id": 2482073,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 9776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661181.731, "dur": 40.971,
    "args": {
      "External id": 2482074,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661182.207, "dur": 39.939,
    "args": {
      "External id": 2482075,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661234.121, "dur": 3.434,
    "args": {
      "External id": 2482076,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661235.948, "dur": 0.602,
    "args": {
      "External id": 2482077,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "32768"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661243.799, "dur": 1.458,
    "args": {
      "External id": 2482078,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 9781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661252.493, "dur": 7.883,
    "args": {
      "External id": 2482079,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 9782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661256.127, "dur": 3.993,
    "args": {
      "External id": 2482080,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661345.361, "dur": 165.576,
    "args": {
      "External id": 2482081,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661348.926, "dur": 2.149,
    "args": {
      "External id": 2482082,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661352.483, "dur": 157.972,
    "args": {
      "External id": 2482083,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 9786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661353.632, "dur": 0.280,
    "args": {
      "External id": 2482084,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661354.882, "dur": 19.949,
    "args": {
      "External id": 2482085,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661376.321, "dur": 4.533,
    "args": {
      "External id": 2482086,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 9789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661379.835, "dur": 0.719,
    "args": {
      "External id": 2482087,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 9790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661381.692, "dur": 21.487,
    "args": {
      "External id": 2482088,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661384.328, "dur": 1.788,
    "args": {
      "External id": 2482089,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661387.185, "dur": 15.652,
    "args": {
      "External id": 2482090,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 9793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661389.767, "dur": 2.372,
    "args": {
      "External id": 2482091,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661404.482, "dur": 18.488,
    "args": {
      "External id": 2482092,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661424.127, "dur": 11.989,
    "args": {
      "External id": 2482093,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661438.546, "dur": 12.574,
    "args": {
      "External id": 2482094,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 9797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661452.309, "dur": 10.546,
    "args": {
      "External id": 2482095,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661464.575, "dur": 21.887,
    "args": {
      "External id": 2482096,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 9799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661468.105, "dur": 0.995,
    "args": {
      "External id": 2482097,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661473.332, "dur": 0.778,
    "args": {
      "External id": 2482098,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 9801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661487.729, "dur": 10.318,
    "args": {
      "External id": 2482099,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661498.958, "dur": 10.564,
    "args": {
      "External id": 2482100,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661517.118, "dur": 1.484,
    "args": {
      "External id": 2482101,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661525.846, "dur": 20.787,
    "args": {
      "External id": 2482102,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661544.489, "dur": 0.777,
    "args": {
      "External id": 2482103,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "32768"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661611.402, "dur": 51.291,
    "args": {
      "External id": 2482104,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 9807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661667.421, "dur": 5.906,
    "args": {
      "External id": 2482105,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661671.600, "dur": 0.739,
    "args": {
      "External id": 2482106,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "67108864"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661674.545, "dur": 23.579,
    "args": {
      "External id": 2482107,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 9810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661702.627, "dur": 4.838,
    "args": {
      "External id": 2482108,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 9811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661703.810, "dur": 3.049,
    "args": {
      "External id": 2482109,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 9812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661705.563, "dur": 1.132,
    "args": {
      "External id": 2482110,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 9813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661709.669, "dur": 39.808,
    "args": {
      "External id": 2482111,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661712.480, "dur": 36.482,
    "args": {
      "External id": 2482112,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661753.309, "dur": 14.113,
    "args": {
      "External id": 2482113,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 9816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661772.482, "dur": 3.411,
    "args": {
      "External id": 2482114,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661774.340, "dur": 0.698,
    "args": {
      "External id": 2482115,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "83886080"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661779.307, "dur": 52.915,
    "args": {
      "External id": 2482116,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 9819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661787.271, "dur": 5.617,
    "args": {
      "External id": 2482117,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 9820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661789.962, "dur": 2.391,
    "args": {
      "External id": 2482118,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 9821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661791.517, "dur": 0.696,
    "args": {
      "External id": 2482119,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 9822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661793.412, "dur": 38.450,
    "args": {
      "External id": 2482120,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661793.914, "dur": 37.472,
    "args": {
      "External id": 2482121,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661836.159, "dur": 3.680,
    "args": {
      "External id": 2482122,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661837.911, "dur": 1.008,
    "args": {
      "External id": 2482123,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "40960"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661847.645, "dur": 1.557,
    "args": {
      "External id": 2482124,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 9827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661855.804, "dur": 7.222,
    "args": {
      "External id": 2482125,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 9828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661857.904, "dur": 4.846,
    "args": {
      "External id": 2482126,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661937.688, "dur": 212.085,
    "args": {
      "External id": 2482127,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661939.874, "dur": 2.425,
    "args": {
      "External id": 2482128,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661943.419, "dur": 205.817,
    "args": {
      "External id": 2482129,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 9832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661946.354, "dur": 0.447,
    "args": {
      "External id": 2482130,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661947.679, "dur": 20.549,
    "args": {
      "External id": 2482131,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661969.634, "dur": 3.118,
    "args": {
      "External id": 2482132,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 9835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661971.600, "dur": 0.932,
    "args": {
      "External id": 2482133,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 9836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661973.577, "dur": 58.490,
    "args": {
      "External id": 2482134,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661974.372, "dur": 1.786,
    "args": {
      "External id": 2482135,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656661977.139, "dur": 54.557,
    "args": {
      "External id": 2482136,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 9839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662014.545, "dur": 3.755,
    "args": {
      "External id": 2482137,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662033.945, "dur": 20.829,
    "args": {
      "External id": 2482138,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662055.844, "dur": 12.309,
    "args": {
      "External id": 2482139,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662070.667, "dur": 12.234,
    "args": {
      "External id": 2482140,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 9843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662085.893, "dur": 11.575,
    "args": {
      "External id": 2482141,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662099.347, "dur": 21.656,
    "args": {
      "External id": 2482142,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 9845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662101.151, "dur": 1.554,
    "args": {
      "External id": 2482143,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662104.832, "dur": 0.855,
    "args": {
      "External id": 2482144,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 9847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662122.266, "dur": 11.958,
    "args": {
      "External id": 2482145,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662135.120, "dur": 11.610,
    "args": {
      "External id": 2482146,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662157.342, "dur": 2.284,
    "args": {
      "External id": 2482147,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662167.870, "dur": 3.348,
    "args": {
      "External id": 2482148,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662169.952, "dur": 0.572,
    "args": {
      "External id": 2482149,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "40960"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662234.873, "dur": 52.299,
    "args": {
      "External id": 2482150,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 9853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662291.843, "dur": 4.240,
    "args": {
      "External id": 2482151,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662294.047, "dur": 1.145,
    "args": {
      "External id": 2482152,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "83886080"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662297.325, "dur": 21.142,
    "args": {
      "External id": 2482153,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 9856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662322.799, "dur": 6.499,
    "args": {
      "External id": 2482154,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 9857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662325.616, "dur": 3.016,
    "args": {
      "External id": 2482155,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 9858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662327.364, "dur": 1.075,
    "args": {
      "External id": 2482156,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 9859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662331.520, "dur": 38.659,
    "args": {
      "External id": 2482157,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662332.705, "dur": 36.961,
    "args": {
      "External id": 2482158,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662373.429, "dur": 13.525,
    "args": {
      "External id": 2482159,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 9862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662392.740, "dur": 5.319,
    "args": {
      "External id": 2482160,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662396.562, "dur": 0.733,
    "args": {
      "External id": 2482161,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "100663296"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662401.499, "dur": 46.306,
    "args": {
      "External id": 2482162,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 9865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662402.214, "dur": 4.820,
    "args": {
      "External id": 2482163,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 9866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662402.995, "dur": 3.499,
    "args": {
      "External id": 2482164,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 9867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662404.154, "dur": 2.173,
    "args": {
      "External id": 2482165,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 9868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662407.644, "dur": 39.828,
    "args": {
      "External id": 2482166,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662408.126, "dur": 38.660,
    "args": {
      "External id": 2482167,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662452.195, "dur": 3.268,
    "args": {
      "External id": 2482168,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662453.853, "dur": 0.614,
    "args": {
      "External id": 2482169,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "49152"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662462.508, "dur": 1.299,
    "args": {
      "External id": 2482170,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 9873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662470.141, "dur": 5.389,
    "args": {
      "External id": 2482171,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 9874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662471.944, "dur": 3.299,
    "args": {
      "External id": 2482172,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662565.690, "dur": 195.665,
    "args": {
      "External id": 2482173,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662568.567, "dur": 3.042,
    "args": {
      "External id": 2482174,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662574.743, "dur": 186.157,
    "args": {
      "External id": 2482175,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 9878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662576.488, "dur": 0.447,
    "args": {
      "External id": 2482176,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662604.909, "dur": 22.294,
    "args": {
      "External id": 2482177,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662628.871, "dur": 3.584,
    "args": {
      "External id": 2482178,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 9881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662630.797, "dur": 1.316,
    "args": {
      "External id": 2482179,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 9882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662633.196, "dur": 20.188,
    "args": {
      "External id": 2482180,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662633.951, "dur": 2.897,
    "args": {
      "External id": 2482181,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662638.194, "dur": 14.857,
    "args": {
      "External id": 2482182,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 9885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662640.641, "dur": 2.214,
    "args": {
      "External id": 2482183,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662656.385, "dur": 20.621,
    "args": {
      "External id": 2482184,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662678.142, "dur": 11.515,
    "args": {
      "External id": 2482185,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662692.092, "dur": 12.556,
    "args": {
      "External id": 2482186,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 9889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662705.693, "dur": 11.337,
    "args": {
      "External id": 2482187,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662718.476, "dur": 16.840,
    "args": {
      "External id": 2482188,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 9891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662720.042, "dur": 1.007,
    "args": {
      "External id": 2482189,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662722.764, "dur": 0.692,
    "args": {
      "External id": 2482190,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 9893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662736.561, "dur": 10.479,
    "args": {
      "External id": 2482191,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662749.967, "dur": 10.075,
    "args": {
      "External id": 2482192,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662768.353, "dur": 2.005,
    "args": {
      "External id": 2482193,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662778.025, "dur": 3.284,
    "args": {
      "External id": 2482194,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662780.152, "dur": 0.388,
    "args": {
      "External id": 2482195,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "49152"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662842.639, "dur": 51.489,
    "args": {
      "External id": 2482196,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 9899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662898.842, "dur": 4.391,
    "args": {
      "External id": 2482197,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662901.281, "dur": 0.719,
    "args": {
      "External id": 2482198,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "100663296"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662904.702, "dur": 21.245,
    "args": {
      "External id": 2482199,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 9902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662932.389, "dur": 4.462,
    "args": {
      "External id": 2482200,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 9903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662933.656, "dur": 2.656,
    "args": {
      "External id": 2482201,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 9904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662935.263, "dur": 0.883,
    "args": {
      "External id": 2482202,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 9905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662938.945, "dur": 37.351,
    "args": {
      "External id": 2482203,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662939.908, "dur": 35.720,
    "args": {
      "External id": 2482204,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656662979.908, "dur": 51.834,
    "args": {
      "External id": 2482205,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 9908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663040.162, "dur": 5.860,
    "args": {
      "External id": 2482206,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663044.220, "dur": 0.876,
    "args": {
      "External id": 2482207,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "117440512"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663050.067, "dur": 52.769,
    "args": {
      "External id": 2482208,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 9911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663050.734, "dur": 4.965,
    "args": {
      "External id": 2482209,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 9912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663051.540, "dur": 3.603,
    "args": {
      "External id": 2482210,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 9913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663054.445, "dur": 0.563,
    "args": {
      "External id": 2482211,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 9914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663056.534, "dur": 45.868,
    "args": {
      "External id": 2482212,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663058.745, "dur": 42.998,
    "args": {
      "External id": 2482213,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 9916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663107.088, "dur": 3.546,
    "args": {
      "External id": 2482214,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663108.876, "dur": 0.728,
    "args": {
      "External id": 2482215,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "57344"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663116.643, "dur": 1.364,
    "args": {
      "External id": 2482216,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 9919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663125.203, "dur": 7.100,
    "args": {
      "External id": 2482217,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 9920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663128.549, "dur": 3.477,
    "args": {
      "External id": 2482218,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663214.631, "dur": 170.501,
    "args": {
      "External id": 2482219,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663216.744, "dur": 2.354,
    "args": {
      "External id": 2482220,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663220.375, "dur": 164.311,
    "args": {
      "External id": 2482221,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 9924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663223.191, "dur": 0.309,
    "args": {
      "External id": 2482222,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663224.350, "dur": 19.966,
    "args": {
      "External id": 2482223,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 9926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663245.811, "dur": 3.119,
    "args": {
      "External id": 2482224,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 9927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663247.762, "dur": 0.950,
    "args": {
      "External id": 2482225,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 9928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663249.803, "dur": 27.501,
    "args": {
      "External id": 2482226,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663258.261, "dur": 1.516,
    "args": {
      "External id": 2482227,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663260.808, "dur": 16.242,
    "args": {
      "External id": 2482228,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 9931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663263.424, "dur": 2.195,
    "args": {
      "External id": 2482229,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663278.797, "dur": 18.556,
    "args": {
      "External id": 2482230,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663298.679, "dur": 13.172,
    "args": {
      "External id": 2482231,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663314.675, "dur": 12.492,
    "args": {
      "External id": 2482232,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 9935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663328.365, "dur": 11.003,
    "args": {
      "External id": 2482233,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 9936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663340.852, "dur": 18.786,
    "args": {
      "External id": 2482234,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 9937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663342.518, "dur": 1.045,
    "args": {
      "External id": 2482235,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 9938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663347.286, "dur": 0.611,
    "args": {
      "External id": 2482236,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 9939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663360.671, "dur": 11.497,
    "args": {
      "External id": 2482237,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 9940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663373.078, "dur": 10.795,
    "args": {
      "External id": 2482238,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 9941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663390.967, "dur": 1.652,
    "args": {
      "External id": 2482239,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 9942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663400.286, "dur": 3.352,
    "args": {
      "External id": 2482240,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 9943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663402.472, "dur": 0.407,
    "args": {
      "External id": 2482241,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "57344"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663463.232, "dur": 46.736,
    "args": {
      "External id": 2482242,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 9945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663514.689, "dur": 7.507,
    "args": {
      "External id": 2482243,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 9946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663518.661, "dur": 2.448,
    "args": {
      "External id": 2482244,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "117440512"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 9947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663523.583, "dur": 42.887,
    "args": {
      "External id": 2482245,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 9948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663572.758, "dur": 5.467,
    "args": {
      "External id": 2482246,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 9949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663574.146, "dur": 3.001,
    "args": {
      "External id": 2482247,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 9950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663575.894, "dur": 1.039,
    "args": {
      "External id": 2482248,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 9951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663580.968, "dur": 44.800,
    "args": {
      "External id": 2482249,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663583.859, "dur": 41.288,
    "args": {
      "External id": 2482250,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 9953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663629.788, "dur": 14.422,
    "args": {
      "External id": 2482251,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 9954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663648.783, "dur": 21.543,
    "args": {
      "External id": 2482252,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 9955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663650.868, "dur": 19.032,
    "args": {
      "External id": 2482253,"Record function id": 0, "Concrete Inputs": ["", "[]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 9956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663655.570, "dur": 1.168,
    "args": {
      "External id": 2482254,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 9957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663675.283, "dur": 25.851,
    "args": {
      "External id": 2482255,"Record function id": 0, "Concrete Inputs": ["", "", "15", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[32000, 2048], [], [], [], [], []], "Ev Idx": 9958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663676.949, "dur": 23.988,
    "args": {
      "External id": 2482256,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], [], []], "Input Dims": [[32000, 2048], [], [], [], [], [], []], "Ev Idx": 9959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663681.020, "dur": 3.883,
    "args": {
      "External id": 2482257,"Record function id": 0, "Concrete Inputs": ["[32000, 2048]", "[2048, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663688.070, "dur": 12.472,
    "args": {
      "External id": 2482258,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "float", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 9961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663712.481, "dur": 4.426,
    "args": {
      "External id": 2482259,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 9962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663714.106, "dur": 2.570,
    "args": {
      "External id": 2482260,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 9963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663717.908, "dur": 1.262,
    "args": {
      "External id": 2482261,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 9964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663718.319, "dur": 0.662,
    "args": {
      "External id": 2482262,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 9965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663760.905, "dur": 22.461,
    "args": {
      "External id": 2482263,"Sequence number": 24740640, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1], [], []], "Ev Idx": 9966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663785.163, "dur": 14.939,
    "args": {
      "External id": 2482264,"Sequence number": 24740641, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1], [], []], "Ev Idx": 9967
    }
  },
  {
    "ph": "s", "id": 9, "pid": 1336759, "tid": 1336759, "ts": 1514656663785.163,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663806.054, "dur": 5.244,
    "args": {
      "External id": 2482265,"Sequence number": 24740642, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], [], []], "Input Dims": [[16, 4, 4096], [], [], [], []], "Ev Idx": 9968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663809.118, "dur": 0.963,
    "args": {
      "External id": 2482266,"Record function id": 0, "Concrete Inputs": ["", "[16, 4, 4096]", "[8192, 1, 1]", "1"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[16, 4, 4096], [], [], []], "Ev Idx": 9969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663817.518, "dur": 6.353,
    "args": {
      "External id": 2482267,"Sequence number": 24740642, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "3"], "Input type": ["long int", "Scalar", "Scalar"], "Input Strides": [[8192, 1, 1], [], []], "Input Dims": [[16, 4, 4096], [], []], "Ev Idx": 9970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663821.962, "dur": 0.636,
    "args": {
      "External id": 2482268,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096]", "[8192, 1]", "4"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1, 1], [], [], []], "Input Dims": [[16, 4, 4096], [], [], []], "Ev Idx": 9971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663825.162, "dur": 2.903,
    "args": {
      "External id": 2482269,"Sequence number": 24740642, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [], [], [], []], "Input Dims": [[16, 4096], [], [], [], []], "Ev Idx": 9972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663826.501, "dur": 0.984,
    "args": {
      "External id": 2482270,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096]", "[8192, 1]", "4"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[8192, 1], [], [], []], "Input Dims": [[16, 4096], [], [], []], "Ev Idx": 9973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663833.951, "dur": 6.850,
    "args": {
      "External id": 2482271,"Sequence number": 24740642, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], [], []], "Ev Idx": 9974
    }
  },
  {
    "ph": "s", "id": 8, "pid": 1336759, "tid": 1336759, "ts": 1514656663833.951,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663838.349, "dur": 1.102,
    "args": {
      "External id": 2482272,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 2048]", "[33554432, 8192, 2048, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], []], "Ev Idx": 9975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663841.740, "dur": 3.736,
    "args": {
      "External id": 2482273,"Sequence number": 24740643, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "1", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], [], []], "Ev Idx": 9976
    }
  },
  {
    "ph": "s", "id": 7, "pid": 1336759, "tid": 1336759, "ts": 1514656663841.740,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663844.145, "dur": 0.620,
    "args": {
      "External id": 2482274,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 4, 2048]", "[33554432, 8192, 2048, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], []], "Ev Idx": 9977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::select", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663846.413, "dur": 5.958,
    "args": {
      "External id": 2482275,"Sequence number": 24740644, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2", "3"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], []], "Input Dims": [[16, 4096, 4, 2048], [], []], "Ev Idx": 9978
    }
  },
  {
    "ph": "s", "id": 6, "pid": 1336759, "tid": 1336759, "ts": 1514656663846.413,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663850.787, "dur": 0.686,
    "args": {
      "External id": 2482276,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]", "[33554432, 8192, 1]", "6144"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 2048, 1], [], [], []], "Input Dims": [[16, 4096, 4, 2048], [], [], []], "Ev Idx": 9979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663853.313, "dur": 5.810,
    "args": {
      "External id": 2482277,"Sequence number": 24740645, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "2", "0", "9223372036854775807", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[33554432, 8192, 1], [], [], [], []], "Input Dims": [[16, 4096, 2048], [], [], [], []], "Ev Idx": 9980
    }
  },
  {
    "ph": "s", "id": 5, "pid": 1336759, "tid": 1336759, "ts": 1514656663853.313,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663855.822, "dur": 2.529,
    "args": {
      "External id": 2482278,"Record function id": 0, "Concrete Inputs": ["", "[16, 4096, 2048]", "[33554432, 8192, 1]", "6144"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[33554432, 8192, 1], [], [], []], "Input Dims": [[16, 4096, 2048], [], [], []], "Ev Idx": 9981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663862.740, "dur": 25.340,
    "args": {
      "External id": 2482279,"Sequence number": 24740646, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["long int", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[16, 4096], []], "Ev Idx": 9982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663864.126, "dur": 23.756,
    "args": {
      "External id": 2482280,"Sequence number": 24740646, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["long int", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[16, 4096], []], "Ev Idx": 9983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663866.366, "dur": 5.541,
    "args": {
      "External id": 2482281,"Record function id": 0, "Concrete Inputs": ["", "4", "0", "", "", "0"], "Input type": ["long int", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[8192, 1], [], [], [], [], []], "Input Dims": [[16, 4096], [], [], [], [], []], "Ev Idx": 9984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663868.413, "dur": 2.975,
    "args": {
      "External id": 2482282,"Record function id": 0, "Concrete Inputs": ["[16, 4096]", "4", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663872.727, "dur": 14.736,
    "args": {
      "External id": 2482283,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "long int", "Scalar"], "Input Strides": [[4096, 1], [8192, 1], []], "Input Dims": [[16, 4096], [16, 4096], []], "Ev Idx": 9986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663912.227, "dur": 5.063,
    "args": {
      "External id": 2482284,"Sequence number": 24740646, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1, 2048]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[33554432, 8192, 1], []], "Input Dims": [[16, 4096, 2048], []], "Ev Idx": 9987
    }
  },
  {
    "ph": "s", "id": 4, "pid": 1336759, "tid": 1336759, "ts": 1514656663912.227,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663919.483, "dur": 1.100,
    "args": {
      "External id": 2482285,"Sequence number": 24740647, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["long int", "ScalarList"], "Input Strides": [[4096, 1], []], "Input Dims": [[16, 4096], []], "Ev Idx": 9988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "FusedLinearCrossEntropyFunction", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663951.408, "dur": 44574.566,
    "args": {
      "External id": 2482286,"Sequence number": 24740647, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "", "", "-100", "0.", "1.", "8"], "Input type": ["c10::BFloat16", "long int", "c10::BFloat16", "", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[8192, 1], [1], [2048, 1], [], [], [], [], []], "Input Dims": [[65536, 2048], [65536], [32000, 2048], [], [], [], [], []], "Ev Idx": 9989
    }
  },
  {
    "ph": "s", "id": 3, "pid": 1336759, "tid": 1336759, "ts": 1514656663951.408,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::contiguous", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663964.584, "dur": 63.394,
    "args": {
      "External id": 2482287,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 9990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663965.293, "dur": 62.429,
    "args": {
      "External id": 2482288,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[8192, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 9991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663966.604, "dur": 5.936,
    "args": {
      "External id": 2482289,"Record function id": 0, "Concrete Inputs": ["", "15", "0", "", "", "0"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[8192, 1], [], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], [], []], "Ev Idx": 9992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663967.770, "dur": 4.305,
    "args": {
      "External id": 2482290,"Record function id": 0, "Concrete Inputs": ["[65536, 2048]", "15", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656663973.176, "dur": 53.841,
    "args": {
      "External id": 2482291,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [8192, 1], []], "Input Dims": [[65536, 2048], [65536, 2048], []], "Ev Idx": 9994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656664046.800, "dur": 30.654,
    "args": {
      "External id": 2482292,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", ""], "Input type": ["c10::BFloat16", "", "", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], [], []], "Ev Idx": 9995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656664047.905, "dur": 9.425,
    "args": {
      "External id": 2482293,"Record function id": 0, "Concrete Inputs": ["", "15", "0", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], [], []], "Ev Idx": 9996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656664051.936, "dur": 5.082,
    "args": {
      "External id": 2482294,"Record function id": 0, "Concrete Inputs": ["[65536, 2048]", "[2048, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 9997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656664058.563, "dur": 18.676,
    "args": {
      "External id": 2482295,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 9998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656664060.356, "dur": 16.454,
    "args": {
      "External id": 2482296,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], []], "Input Dims": [[65536, 2048], []], "Ev Idx": 9999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656664080.781, "dur": 17.757,
    "args": {
      "External id": 2482297,"Record function id": 0, "Concrete Inputs": ["", "6", "", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[32000, 2048], [], [], [], [], []], "Ev Idx": 10000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656664081.676, "dur": 3.860,
    "args": {
      "External id": 2482298,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "False", ""], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[32000, 2048], [], [], [], [], []], "Ev Idx": 10001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656664082.735, "dur": 2.571,
    "args": {
      "External id": 2482299,"Record function id": 0, "Concrete Inputs": ["[32000, 2048]", "[2048, 1]", "6", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656664086.198, "dur": 12.140,
    "args": {
      "External id": 2482300,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 10003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656664086.582, "dur": 11.466,
    "args": {
      "External id": 2482301,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[2048, 1], []], "Input Dims": [[32000, 2048], []], "Ev Idx": 10004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 1336759, "tid": 1336759,
    "ts": 1514656664104.664, "dur": 17.641,
    "args": {
      "External id": 2482302,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False"], "Input type": ["ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 10005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656664108.079, "dur": 3.017,
    "args": {
      "External id": 2482303,"Record function id": 0, "Concrete Inputs": ["[65536]", "6", "", "", "False", ""], "Input type": ["ScalarList", "Scalar", "", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656664111.927, "dur": 10.116,
    "args": {
      "External id": 2482304,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[65536]], "Ev Idx": 10007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656664112.442, "dur": 9.311,
    "args": {
      "External id": 2482305,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 10008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::ne", "pid": 1336759, "tid": 1336759,
    "ts": 1514656664127.280, "dur": 19.616,
    "args": {
      "External id": 2482306,"Record function id": 0, "Concrete Inputs": ["", "-100"], "Input type": ["long int", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 10009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656664149.312, "dur": 44.583,
    "args": {
      "External id": 2482307,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["bool", ""], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 10010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656664151.284, "dur": 42.166,
    "args": {
      "External id": 2482308,"Record function id": 0, "Concrete Inputs": ["", "[]", "False", ""], "Input type": ["bool", "ScalarList", "Scalar", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 10011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656664157.382, "dur": 1.211,
    "args": {
      "External id": 2482309,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["long int", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656664159.830, "dur": 20.255,
    "args": {
      "External id": 2482310,"Record function id": 0, "Concrete Inputs": ["", "4", "False", "False", ""], "Input type": ["bool", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 10013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656664161.436, "dur": 18.393,
    "args": {
      "External id": 2482311,"Record function id": 0, "Concrete Inputs": ["", "4", "", "", "", "False", ""], "Input type": ["bool", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[1], [], [], [], [], [], []], "Input Dims": [[65536], [], [], [], [], [], []], "Ev Idx": 10014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656664163.540, "dur": 2.469,
    "args": {
      "External id": 2482312,"Record function id": 0, "Concrete Inputs": ["[65536]", "[1]", "4", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656664166.813, "dur": 12.520,
    "args": {
      "External id": 2482313,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["long int", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[65536], [65536], []], "Ev Idx": 10016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1336759, "tid": 1336759,
    "ts": 1514656664199.460, "dur": 39137.999,
    "args": {
      "External id": 2482314,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 10017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1336759, "tid": 1336759,
    "ts": 1514656664200.923, "dur": 39135.564,
    "args": {
      "External id": 2482315,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["long int"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 10018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703347.646, "dur": 6.192,
    "args": {
      "External id": 2482316,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 10019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703351.086, "dur": 1.097,
    "args": {
      "External id": 2482317,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 10020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703358.784, "dur": 92.864,
    "args": {
      "External id": 2482318,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 10021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703360.252, "dur": 7.525,
    "args": {
      "External id": 2482319,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 10022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703362.503, "dur": 4.484,
    "args": {
      "External id": 2482320,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 10023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703365.881, "dur": 0.845,
    "args": {
      "External id": 2482321,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 10024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703368.822, "dur": 82.170,
    "args": {
      "External id": 2482322,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 10025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703370.299, "dur": 79.833,
    "args": {
      "External id": 2482323,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 10026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703454.924, "dur": 3.987,
    "args": {
      "External id": 2482324,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 10027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703456.846, "dur": 0.800,
    "args": {
      "External id": 2482325,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "0"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 10028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703465.238, "dur": 2.158,
    "args": {
      "External id": 2482326,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 10029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703476.791, "dur": 6.145,
    "args": {
      "External id": 2482327,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 10030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703478.431, "dur": 4.205,
    "args": {
      "External id": 2482328,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703609.510, "dur": 185.400,
    "args": {
      "External id": 2482329,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 10032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703613.503, "dur": 2.944,
    "args": {
      "External id": 2482330,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703617.910, "dur": 176.504,
    "args": {
      "External id": 2482331,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 10034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703619.578, "dur": 0.399,
    "args": {
      "External id": 2482332,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 10035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703621.243, "dur": 23.323,
    "args": {
      "External id": 2482333,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 10036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703647.996, "dur": 4.736,
    "args": {
      "External id": 2482334,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 10037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703651.340, "dur": 1.076,
    "args": {
      "External id": 2482335,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 10038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703653.748, "dur": 21.857,
    "args": {
      "External id": 2482336,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 10039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703654.888, "dur": 1.247,
    "args": {
      "External id": 2482337,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703657.302, "dur": 18.068,
    "args": {
      "External id": 2482338,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 10041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703660.635, "dur": 3.234,
    "args": {
      "External id": 2482339,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 10042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703677.285, "dur": 20.742,
    "args": {
      "External id": 2482340,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 10043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703699.525, "dur": 15.980,
    "args": {
      "External id": 2482341,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 10044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703718.289, "dur": 14.153,
    "args": {
      "External id": 2482342,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 10045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703733.790, "dur": 11.754,
    "args": {
      "External id": 2482343,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 10046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703749.294, "dur": 19.609,
    "args": {
      "External id": 2482344,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 10047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703751.357, "dur": 1.281,
    "args": {
      "External id": 2482345,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 10048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703754.987, "dur": 0.916,
    "args": {
      "External id": 2482346,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 10049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703770.479, "dur": 11.085,
    "args": {
      "External id": 2482347,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 10050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703782.843, "dur": 10.392,
    "args": {
      "External id": 2482348,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 10051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703801.764, "dur": 2.336,
    "args": {
      "External id": 2482349,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 10052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703812.374, "dur": 4.178,
    "args": {
      "External id": 2482350,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 10053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703814.494, "dur": 1.027,
    "args": {
      "External id": 2482351,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "0"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 10054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703882.734, "dur": 61.227,
    "args": {
      "External id": 2482352,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 10055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703949.156, "dur": 4.990,
    "args": {
      "External id": 2482353,"Record function id": 0, "Concrete Inputs": ["", "0", "0", "8192", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 10056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703951.521, "dur": 0.713,
    "args": {
      "External id": 2482354,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "0"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 10057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656703955.724, "dur": 24.995,
    "args": {
      "External id": 2482355,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 10058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704022.606, "dur": 9.110,
    "args": {
      "External id": 2482356,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 10059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704024.330, "dur": 6.524,
    "args": {
      "External id": 2482357,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 10060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704028.703, "dur": 1.768,
    "args": {
      "External id": 2482358,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 10061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704035.023, "dur": 49.629,
    "args": {
      "External id": 2482359,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 10062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704036.148, "dur": 47.854,
    "args": {
      "External id": 2482360,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 10063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704089.282, "dur": 16.440,
    "args": {
      "External id": 2482361,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 10064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704112.117, "dur": 3.932,
    "args": {
      "External id": 2482362,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 10065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704114.386, "dur": 0.706,
    "args": {
      "External id": 2482363,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "16777216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 10066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704122.299, "dur": 47.032,
    "args": {
      "External id": 2482364,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 10067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704123.296, "dur": 3.606,
    "args": {
      "External id": 2482365,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 10068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704124.098, "dur": 2.241,
    "args": {
      "External id": 2482366,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 10069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704125.387, "dur": 0.838,
    "args": {
      "External id": 2482367,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 10070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704127.522, "dur": 41.453,
    "args": {
      "External id": 2482368,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 10071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704128.235, "dur": 40.020,
    "args": {
      "External id": 2482369,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 10072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704174.662, "dur": 3.518,
    "args": {
      "External id": 2482370,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 10073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704176.289, "dur": 0.742,
    "args": {
      "External id": 2482371,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "8192"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 10074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704184.895, "dur": 1.574,
    "args": {
      "External id": 2482372,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 10075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704193.684, "dur": 8.165,
    "args": {
      "External id": 2482373,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 10076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704195.688, "dur": 5.829,
    "args": {
      "External id": 2482374,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704297.001, "dur": 171.845,
    "args": {
      "External id": 2482375,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 10078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704299.266, "dur": 2.353,
    "args": {
      "External id": 2482376,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704302.755, "dur": 165.515,
    "args": {
      "External id": 2482377,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 10080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704304.491, "dur": 0.448,
    "args": {
      "External id": 2482378,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 10081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704306.016, "dur": 22.342,
    "args": {
      "External id": 2482379,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 10082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704329.828, "dur": 3.345,
    "args": {
      "External id": 2482380,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 10083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704331.937, "dur": 0.897,
    "args": {
      "External id": 2482381,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 10084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704334.086, "dur": 23.367,
    "args": {
      "External id": 2482382,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 10085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704335.216, "dur": 1.732,
    "args": {
      "External id": 2482383,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10086
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704338.032, "dur": 19.079,
    "args": {
      "External id": 2482384,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 10087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704343.828, "dur": 2.415,
    "args": {
      "External id": 2482385,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 10088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704358.840, "dur": 21.851,
    "args": {
      "External id": 2482386,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 10089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704381.944, "dur": 11.977,
    "args": {
      "External id": 2482387,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 10090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704396.374, "dur": 13.614,
    "args": {
      "External id": 2482388,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 10091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704411.212, "dur": 11.285,
    "args": {
      "External id": 2482389,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 10092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704424.196, "dur": 18.409,
    "args": {
      "External id": 2482390,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 10093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704426.123, "dur": 1.477,
    "args": {
      "External id": 2482391,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 10094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704429.650, "dur": 0.801,
    "args": {
      "External id": 2482392,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 10095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704445.285, "dur": 10.751,
    "args": {
      "External id": 2482393,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 10096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704457.167, "dur": 10.355,
    "args": {
      "External id": 2482394,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 10097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704475.807, "dur": 1.507,
    "args": {
      "External id": 2482395,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 10098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704485.431, "dur": 3.388,
    "args": {
      "External id": 2482396,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 10099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704487.466, "dur": 0.568,
    "args": {
      "External id": 2482397,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "8192"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 10100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704573.006, "dur": 53.404,
    "args": {
      "External id": 2482398,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 10101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704631.668, "dur": 4.962,
    "args": {
      "External id": 2482399,"Record function id": 0, "Concrete Inputs": ["", "0", "8192", "16384", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 10102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704634.273, "dur": 0.877,
    "args": {
      "External id": 2482400,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "16777216"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 10103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704639.761, "dur": 23.979,
    "args": {
      "External id": 2482401,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 10104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704668.472, "dur": 5.009,
    "args": {
      "External id": 2482402,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 10105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704669.657, "dur": 3.295,
    "args": {
      "External id": 2482403,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 10106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704671.425, "dur": 1.327,
    "args": {
      "External id": 2482404,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 10107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704676.098, "dur": 38.136,
    "args": {
      "External id": 2482405,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 10108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704677.306, "dur": 36.357,
    "args": {
      "External id": 2482406,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 10109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704717.903, "dur": 14.166,
    "args": {
      "External id": 2482407,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 10110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704739.189, "dur": 3.480,
    "args": {
      "External id": 2482408,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 10111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704741.224, "dur": 0.673,
    "args": {
      "External id": 2482409,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "33554432"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 10112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704746.241, "dur": 46.005,
    "args": {
      "External id": 2482410,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 10113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704747.361, "dur": 3.281,
    "args": {
      "External id": 2482411,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 10114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704748.300, "dur": 1.809,
    "args": {
      "External id": 2482412,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 10115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704749.411, "dur": 0.558,
    "args": {
      "External id": 2482413,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 10116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704752.961, "dur": 38.948,
    "args": {
      "External id": 2482414,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 10117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704753.397, "dur": 37.955,
    "args": {
      "External id": 2482415,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 10118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704796.248, "dur": 4.975,
    "args": {
      "External id": 2482416,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 10119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704798.090, "dur": 2.063,
    "args": {
      "External id": 2482417,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "16384"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 10120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704806.589, "dur": 1.297,
    "args": {
      "External id": 2482418,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 10121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704814.836, "dur": 5.510,
    "args": {
      "External id": 2482419,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 10122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704816.618, "dur": 3.478,
    "args": {
      "External id": 2482420,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704899.978, "dur": 209.989,
    "args": {
      "External id": 2482421,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 10124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704901.725, "dur": 1.966,
    "args": {
      "External id": 2482422,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704905.254, "dur": 204.104,
    "args": {
      "External id": 2482423,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 10126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704906.442, "dur": 0.402,
    "args": {
      "External id": 2482424,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 10127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704907.802, "dur": 20.104,
    "args": {
      "External id": 2482425,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 10128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704929.673, "dur": 2.936,
    "args": {
      "External id": 2482426,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 10129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704931.500, "dur": 0.832,
    "args": {
      "External id": 2482427,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 10130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704933.538, "dur": 22.010,
    "args": {
      "External id": 2482428,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 10131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704934.764, "dur": 1.425,
    "args": {
      "External id": 2482429,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704937.276, "dur": 18.001,
    "args": {
      "External id": 2482430,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 10133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704942.630, "dur": 2.278,
    "args": {
      "External id": 2482431,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 10134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704956.749, "dur": 18.033,
    "args": {
      "External id": 2482432,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 10135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656704976.155, "dur": 50.813,
    "args": {
      "External id": 2482433,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 10136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705031.041, "dur": 13.597,
    "args": {
      "External id": 2482434,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 10137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705046.223, "dur": 13.070,
    "args": {
      "External id": 2482435,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 10138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705060.978, "dur": 18.900,
    "args": {
      "External id": 2482436,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 10139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705062.897, "dur": 1.724,
    "args": {
      "External id": 2482437,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 10140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705066.632, "dur": 0.812,
    "args": {
      "External id": 2482438,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 10141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705083.158, "dur": 14.011,
    "args": {
      "External id": 2482439,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 10142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705098.256, "dur": 10.096,
    "args": {
      "External id": 2482440,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 10143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705117.986, "dur": 2.104,
    "args": {
      "External id": 2482441,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 10144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705128.504, "dur": 3.442,
    "args": {
      "External id": 2482442,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 10145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705130.776, "dur": 0.418,
    "args": {
      "External id": 2482443,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "16384"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 10146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705196.507, "dur": 52.302,
    "args": {
      "External id": 2482444,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 10147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705253.502, "dur": 6.392,
    "args": {
      "External id": 2482445,"Record function id": 0, "Concrete Inputs": ["", "0", "16384", "24576", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 10148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705257.545, "dur": 1.338,
    "args": {
      "External id": 2482446,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "33554432"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 10149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705261.304, "dur": 22.193,
    "args": {
      "External id": 2482447,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 10150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705287.756, "dur": 4.732,
    "args": {
      "External id": 2482448,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 10151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705289.048, "dur": 2.773,
    "args": {
      "External id": 2482449,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 10152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705290.546, "dur": 1.090,
    "args": {
      "External id": 2482450,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 10153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705294.848, "dur": 38.029,
    "args": {
      "External id": 2482451,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 10154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705295.916, "dur": 36.483,
    "args": {
      "External id": 2482452,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 10155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705338.109, "dur": 16.074,
    "args": {
      "External id": 2482453,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 10156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705359.189, "dur": 3.096,
    "args": {
      "External id": 2482454,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 10157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705360.964, "dur": 0.540,
    "args": {
      "External id": 2482455,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "50331648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 10158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705365.712, "dur": 79.115,
    "args": {
      "External id": 2482456,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 10159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705366.350, "dur": 6.966,
    "args": {
      "External id": 2482457,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 10160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705367.182, "dur": 5.570,
    "args": {
      "External id": 2482458,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 10161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705372.148, "dur": 0.469,
    "args": {
      "External id": 2482459,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 10162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705373.878, "dur": 70.603,
    "args": {
      "External id": 2482460,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 10163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705374.468, "dur": 69.413,
    "args": {
      "External id": 2482461,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 10164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705448.434, "dur": 14.405,
    "args": {
      "External id": 2482462,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 10165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705460.945, "dur": 0.704,
    "args": {
      "External id": 2482463,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "24576"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 10166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705468.098, "dur": 1.327,
    "args": {
      "External id": 2482464,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 10167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705477.970, "dur": 5.375,
    "args": {
      "External id": 2482465,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 10168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705479.590, "dur": 3.489,
    "args": {
      "External id": 2482466,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705579.749, "dur": 172.836,
    "args": {
      "External id": 2482467,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 10170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705583.659, "dur": 3.066,
    "args": {
      "External id": 2482468,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705588.412, "dur": 163.678,
    "args": {
      "External id": 2482469,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 10172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705589.835, "dur": 0.523,
    "args": {
      "External id": 2482470,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 10173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705591.318, "dur": 20.997,
    "args": {
      "External id": 2482471,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 10174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705615.789, "dur": 4.949,
    "args": {
      "External id": 2482472,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 10175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705619.160, "dur": 1.200,
    "args": {
      "External id": 2482473,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 10176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705621.751, "dur": 21.390,
    "args": {
      "External id": 2482474,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 10177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705622.746, "dur": 1.362,
    "args": {
      "External id": 2482475,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705625.114, "dur": 17.712,
    "args": {
      "External id": 2482476,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 10179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705627.453, "dur": 2.222,
    "args": {
      "External id": 2482477,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 10180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705644.553, "dur": 20.598,
    "args": {
      "External id": 2482478,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 10181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705666.200, "dur": 12.068,
    "args": {
      "External id": 2482479,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 10182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705680.643, "dur": 12.970,
    "args": {
      "External id": 2482480,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 10183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705694.749, "dur": 10.741,
    "args": {
      "External id": 2482481,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 10184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705708.495, "dur": 18.785,
    "args": {
      "External id": 2482482,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 10185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705711.841, "dur": 1.098,
    "args": {
      "External id": 2482483,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 10186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705715.039, "dur": 0.574,
    "args": {
      "External id": 2482484,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 10187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705728.419, "dur": 10.744,
    "args": {
      "External id": 2482485,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 10188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705740.147, "dur": 11.152,
    "args": {
      "External id": 2482486,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 10189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705758.925, "dur": 1.840,
    "args": {
      "External id": 2482487,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 10190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705767.838, "dur": 3.653,
    "args": {
      "External id": 2482488,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 10191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705770.051, "dur": 0.414,
    "args": {
      "External id": 2482489,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "24576"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 10192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705831.476, "dur": 50.229,
    "args": {
      "External id": 2482490,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 10193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705886.480, "dur": 4.546,
    "args": {
      "External id": 2482491,"Record function id": 0, "Concrete Inputs": ["", "0", "24576", "32768", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 10194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705888.950, "dur": 0.796,
    "args": {
      "External id": 2482492,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "50331648"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 10195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705892.441, "dur": 22.665,
    "args": {
      "External id": 2482493,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 10196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705919.123, "dur": 6.573,
    "args": {
      "External id": 2482494,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 10197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705920.304, "dur": 4.835,
    "args": {
      "External id": 2482495,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 10198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705923.544, "dur": 1.425,
    "args": {
      "External id": 2482496,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 10199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705927.850, "dur": 38.118,
    "args": {
      "External id": 2482497,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 10200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705928.749, "dur": 36.442,
    "args": {
      "External id": 2482498,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 10201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656705969.713, "dur": 49.726,
    "args": {
      "External id": 2482499,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 10202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706027.845, "dur": 4.335,
    "args": {
      "External id": 2482500,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 10203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706030.310, "dur": 0.861,
    "args": {
      "External id": 2482501,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "67108864"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 10204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706036.188, "dur": 51.488,
    "args": {
      "External id": 2482502,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 10205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706038.805, "dur": 4.247,
    "args": {
      "External id": 2482503,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 10206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706039.912, "dur": 2.587,
    "args": {
      "External id": 2482504,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 10207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706041.495, "dur": 0.846,
    "args": {
      "External id": 2482505,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 10208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706043.652, "dur": 43.563,
    "args": {
      "External id": 2482506,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 10209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706044.271, "dur": 42.336,
    "args": {
      "External id": 2482507,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 10210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706091.790, "dur": 3.601,
    "args": {
      "External id": 2482508,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 10211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706093.595, "dur": 0.673,
    "args": {
      "External id": 2482509,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "32768"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 10212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706101.160, "dur": 1.492,
    "args": {
      "External id": 2482510,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 10213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706110.920, "dur": 5.719,
    "args": {
      "External id": 2482511,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 10214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706112.594, "dur": 3.795,
    "args": {
      "External id": 2482512,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706192.049, "dur": 170.352,
    "args": {
      "External id": 2482513,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 10216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706196.147, "dur": 2.133,
    "args": {
      "External id": 2482514,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706199.229, "dur": 162.673,
    "args": {
      "External id": 2482515,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 10218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706200.610, "dur": 0.236,
    "args": {
      "External id": 2482516,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 10219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706203.640, "dur": 20.258,
    "args": {
      "External id": 2482517,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 10220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706230.000, "dur": 5.926,
    "args": {
      "External id": 2482518,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 10221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706234.555, "dur": 1.142,
    "args": {
      "External id": 2482519,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 10222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706236.802, "dur": 19.211,
    "args": {
      "External id": 2482520,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 10223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706237.617, "dur": 1.609,
    "args": {
      "External id": 2482521,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706240.288, "dur": 15.379,
    "args": {
      "External id": 2482522,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 10225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706242.857, "dur": 2.491,
    "args": {
      "External id": 2482523,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 10226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706257.227, "dur": 19.741,
    "args": {
      "External id": 2482524,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 10227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706278.129, "dur": 12.111,
    "args": {
      "External id": 2482525,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 10228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706293.769, "dur": 12.424,
    "args": {
      "External id": 2482526,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 10229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706307.262, "dur": 11.666,
    "args": {
      "External id": 2482527,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 10230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706320.595, "dur": 16.866,
    "args": {
      "External id": 2482528,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 10231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706322.291, "dur": 1.076,
    "args": {
      "External id": 2482529,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 10232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706325.177, "dur": 0.709,
    "args": {
      "External id": 2482530,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 10233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706338.577, "dur": 11.290,
    "args": {
      "External id": 2482531,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 10234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706351.115, "dur": 9.988,
    "args": {
      "External id": 2482532,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 10235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706370.145, "dur": 1.463,
    "args": {
      "External id": 2482533,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 10236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706378.813, "dur": 3.578,
    "args": {
      "External id": 2482534,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 10237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706381.118, "dur": 0.457,
    "args": {
      "External id": 2482535,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "32768"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 10238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706436.545, "dur": 45.076,
    "args": {
      "External id": 2482536,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 10239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706486.047, "dur": 4.185,
    "args": {
      "External id": 2482537,"Record function id": 0, "Concrete Inputs": ["", "0", "32768", "40960", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 10240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706488.369, "dur": 0.849,
    "args": {
      "External id": 2482538,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "67108864"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 10241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706491.680, "dur": 20.581,
    "args": {
      "External id": 2482539,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 10242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706516.323, "dur": 6.959,
    "args": {
      "External id": 2482540,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 10243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706519.401, "dur": 3.353,
    "args": {
      "External id": 2482541,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 10244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706521.270, "dur": 1.284,
    "args": {
      "External id": 2482542,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 10245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706525.451, "dur": 58.351,
    "args": {
      "External id": 2482543,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 10246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706526.402, "dur": 56.370,
    "args": {
      "External id": 2482544,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 10247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706588.922, "dur": 14.384,
    "args": {
      "External id": 2482545,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 10248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706609.223, "dur": 5.672,
    "args": {
      "External id": 2482546,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 10249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706613.082, "dur": 0.792,
    "args": {
      "External id": 2482547,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "83886080"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 10250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706618.449, "dur": 44.227,
    "args": {
      "External id": 2482548,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 10251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706619.196, "dur": 3.750,
    "args": {
      "External id": 2482549,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 10252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706619.965, "dur": 2.441,
    "args": {
      "External id": 2482550,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 10253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706621.532, "dur": 0.734,
    "args": {
      "External id": 2482551,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 10254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706623.517, "dur": 38.827,
    "args": {
      "External id": 2482552,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 10255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706624.269, "dur": 37.452,
    "args": {
      "External id": 2482553,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 10256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706667.996, "dur": 4.604,
    "args": {
      "External id": 2482554,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 10257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706669.681, "dur": 1.903,
    "args": {
      "External id": 2482555,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "40960"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 10258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706678.145, "dur": 1.335,
    "args": {
      "External id": 2482556,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 10259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706686.244, "dur": 5.900,
    "args": {
      "External id": 2482557,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 10260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706687.968, "dur": 3.901,
    "args": {
      "External id": 2482558,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706772.842, "dur": 162.032,
    "args": {
      "External id": 2482559,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 10262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706774.911, "dur": 2.141,
    "args": {
      "External id": 2482560,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706778.390, "dur": 156.017,
    "args": {
      "External id": 2482561,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 10264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706779.690, "dur": 0.313,
    "args": {
      "External id": 2482562,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 10265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706780.947, "dur": 20.544,
    "args": {
      "External id": 2482563,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 10266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706802.832, "dur": 3.621,
    "args": {
      "External id": 2482564,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 10267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706804.896, "dur": 1.310,
    "args": {
      "External id": 2482565,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 10268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706807.255, "dur": 23.817,
    "args": {
      "External id": 2482566,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 10269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706808.096, "dur": 1.631,
    "args": {
      "External id": 2482567,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706810.886, "dur": 19.891,
    "args": {
      "External id": 2482568,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 10271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706817.283, "dur": 2.623,
    "args": {
      "External id": 2482569,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 10272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706832.324, "dur": 17.724,
    "args": {
      "External id": 2482570,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 10273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706851.415, "dur": 12.013,
    "args": {
      "External id": 2482571,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 10274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706865.873, "dur": 12.447,
    "args": {
      "External id": 2482572,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 10275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706879.533, "dur": 10.708,
    "args": {
      "External id": 2482573,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 10276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706891.460, "dur": 17.095,
    "args": {
      "External id": 2482574,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 10277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706892.987, "dur": 1.305,
    "args": {
      "External id": 2482575,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 10278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706896.022, "dur": 0.552,
    "args": {
      "External id": 2482576,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 10279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706911.375, "dur": 10.579,
    "args": {
      "External id": 2482577,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 10280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706923.159, "dur": 10.269,
    "args": {
      "External id": 2482578,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 10281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706941.353, "dur": 1.427,
    "args": {
      "External id": 2482579,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 10282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706950.093, "dur": 3.162,
    "args": {
      "External id": 2482580,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 10283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656706952.109, "dur": 0.456,
    "args": {
      "External id": 2482581,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "40960"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 10284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707055.035, "dur": 53.188,
    "args": {
      "External id": 2482582,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 10285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707113.421, "dur": 6.843,
    "args": {
      "External id": 2482583,"Record function id": 0, "Concrete Inputs": ["", "0", "40960", "49152", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 10286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707117.739, "dur": 0.918,
    "args": {
      "External id": 2482584,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "83886080"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 10287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707121.646, "dur": 27.198,
    "args": {
      "External id": 2482585,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 10288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707153.427, "dur": 4.634,
    "args": {
      "External id": 2482586,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 10289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707154.448, "dur": 3.026,
    "args": {
      "External id": 2482587,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 10290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707156.240, "dur": 1.047,
    "args": {
      "External id": 2482588,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 10291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707160.633, "dur": 39.723,
    "args": {
      "External id": 2482589,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 10292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707161.583, "dur": 38.119,
    "args": {
      "External id": 2482590,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 10293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707206.065, "dur": 14.363,
    "args": {
      "External id": 2482591,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 10294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707225.693, "dur": 4.013,
    "args": {
      "External id": 2482592,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 10295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707228.133, "dur": 0.818,
    "args": {
      "External id": 2482593,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "100663296"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 10296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707233.082, "dur": 48.949,
    "args": {
      "External id": 2482594,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 10297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707233.901, "dur": 7.460,
    "args": {
      "External id": 2482595,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 10298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707234.631, "dur": 6.145,
    "args": {
      "External id": 2482596,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 10299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707237.805, "dur": 2.740,
    "args": {
      "External id": 2482597,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 10300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707241.933, "dur": 39.586,
    "args": {
      "External id": 2482598,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 10301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707242.735, "dur": 38.286,
    "args": {
      "External id": 2482599,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 10302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707286.229, "dur": 3.726,
    "args": {
      "External id": 2482600,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 10303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707288.103, "dur": 0.853,
    "args": {
      "External id": 2482601,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "49152"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 10304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707295.549, "dur": 1.446,
    "args": {
      "External id": 2482602,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 10305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707303.734, "dur": 8.260,
    "args": {
      "External id": 2482603,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 10306
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707307.263, "dur": 4.460,
    "args": {
      "External id": 2482604,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707387.291, "dur": 180.622,
    "args": {
      "External id": 2482605,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 10308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707389.273, "dur": 2.165,
    "args": {
      "External id": 2482606,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707392.825, "dur": 174.671,
    "args": {
      "External id": 2482607,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 10310
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707394.004, "dur": 0.294,
    "args": {
      "External id": 2482608,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 10311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707397.224, "dur": 19.746,
    "args": {
      "External id": 2482609,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 10312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707418.372, "dur": 2.971,
    "args": {
      "External id": 2482610,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 10313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707420.361, "dur": 0.701,
    "args": {
      "External id": 2482611,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 10314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707422.312, "dur": 23.371,
    "args": {
      "External id": 2482612,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 10315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707425.510, "dur": 3.337,
    "args": {
      "External id": 2482613,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707430.155, "dur": 15.265,
    "args": {
      "External id": 2482614,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 10317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707432.686, "dur": 2.417,
    "args": {
      "External id": 2482615,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 10318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707446.876, "dur": 18.843,
    "args": {
      "External id": 2482616,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 10319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707467.123, "dur": 11.729,
    "args": {
      "External id": 2482617,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 10320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707481.114, "dur": 12.192,
    "args": {
      "External id": 2482618,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 10321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707494.379, "dur": 10.821,
    "args": {
      "External id": 2482619,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 10322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707506.756, "dur": 18.704,
    "args": {
      "External id": 2482620,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 10323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707508.665, "dur": 1.273,
    "args": {
      "External id": 2482621,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 10324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707512.996, "dur": 0.705,
    "args": {
      "External id": 2482622,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 10325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707540.707, "dur": 13.893,
    "args": {
      "External id": 2482623,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 10326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707556.336, "dur": 9.831,
    "args": {
      "External id": 2482624,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 10327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707575.080, "dur": 2.065,
    "args": {
      "External id": 2482625,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 10328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707585.122, "dur": 3.256,
    "args": {
      "External id": 2482626,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 10329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707587.072, "dur": 0.602,
    "args": {
      "External id": 2482627,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "49152"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 10330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707649.083, "dur": 48.804,
    "args": {
      "External id": 2482628,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 10331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707702.587, "dur": 6.126,
    "args": {
      "External id": 2482629,"Record function id": 0, "Concrete Inputs": ["", "0", "49152", "57344", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 10332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707706.881, "dur": 0.787,
    "args": {
      "External id": 2482630,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "100663296"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 10333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707709.885, "dur": 22.372,
    "args": {
      "External id": 2482631,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 10334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707736.342, "dur": 4.733,
    "args": {
      "External id": 2482632,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 10335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707737.661, "dur": 2.815,
    "args": {
      "External id": 2482633,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 10336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707739.325, "dur": 0.994,
    "args": {
      "External id": 2482634,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 10337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707743.457, "dur": 38.573,
    "args": {
      "External id": 2482635,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 10338
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707746.315, "dur": 35.189,
    "args": {
      "External id": 2482636,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 10339
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707785.732, "dur": 13.200,
    "args": {
      "External id": 2482637,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 10340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707803.952, "dur": 3.502,
    "args": {
      "External id": 2482638,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 10341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707805.967, "dur": 0.724,
    "args": {
      "External id": 2482639,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "117440512"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 10342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linear", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707810.572, "dur": 47.082,
    "args": {
      "External id": 2482640,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["c10::BFloat16", "c10::BFloat16", ""], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [32000, 2048], []], "Ev Idx": 10343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707811.281, "dur": 6.717,
    "args": {
      "External id": 2482641,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 10344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707811.990, "dur": 5.428,
    "args": {
      "External id": 2482642,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], []], "Input Dims": [[32000, 2048], [], []], "Ev Idx": 10345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707816.764, "dur": 0.509,
    "args": {
      "External id": 2482643,"Record function id": 0, "Concrete Inputs": ["", "[2048, 32000]", "[1, 2048]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[32000, 2048], [], [], []], "Ev Idx": 10346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707818.557, "dur": 38.771,
    "args": {
      "External id": 2482644,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 10347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707819.315, "dur": 37.368,
    "args": {
      "External id": 2482645,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[2048, 1], [1, 2048]], "Input Dims": [[8192, 2048], [2048, 32000]], "Ev Idx": 10348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707861.549, "dur": 3.552,
    "args": {
      "External id": 2482646,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["long int", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 10349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707863.252, "dur": 0.744,
    "args": {
      "External id": 2482647,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "57344"], "Input type": ["long int", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 10350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707870.135, "dur": 1.251,
    "args": {
      "External id": 2482648,"Record function id": 0, "Concrete Inputs": ["", "[-1, 32000]"], "Input type": ["c10::BFloat16", "ScalarList"], "Input Strides": [[32000, 1], []], "Input Dims": [[8192, 32000], []], "Ev Idx": 10351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::new_empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707878.640, "dur": 5.361,
    "args": {
      "External id": 2482649,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "6", "", "", "False"], "Input type": ["c10::BFloat16", "ScalarList", "Scalar", "", "", "Scalar"], "Input Strides": [[32000, 1], [], [], [], [], []], "Input Dims": [[8192, 32000], [], [], [], [], []], "Ev Idx": 10352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707880.056, "dur": 3.681,
    "args": {
      "External id": 2482650,"Record function id": 0, "Concrete Inputs": ["[8192, 1]", "6", "0", "", "False", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "Scalar", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707953.567, "dur": 203.125,
    "args": {
      "External id": 2482651,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 10354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707955.348, "dur": 1.928,
    "args": {
      "External id": 2482652,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::logsumexp", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707958.497, "dur": 197.611,
    "args": {
      "External id": 2482653,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", "float"], "Input Strides": [[1, 1], [], [], [1]], "Input Dims": [[8192, 1], [], [], [0]], "Ev Idx": 10356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::real", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707961.511, "dur": 0.275,
    "args": {
      "External id": 2482654,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 10357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::amax", "pid": 1336759, "tid": 1336759,
    "ts": 1514656707964.555, "dur": 18.992,
    "args": {
      "External id": 2482655,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "True"], "Input type": ["float", "ScalarList", "Scalar"], "Input Strides": [[1, 1], [], []], "Input Dims": [[8192, 1], [], []], "Ev Idx": 10358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::squeeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708022.459, "dur": 3.674,
    "args": {
      "External id": 2482656,"Record function id": 0, "Concrete Inputs": ["", "[-1]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1, 1], []], "Input Dims": [[8192, 1], []], "Ev Idx": 10359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708024.408, "dur": 1.210,
    "args": {
      "External id": 2482657,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1, 1], [], [], []], "Input Dims": [[8192, 1], [], [], []], "Ev Idx": 10360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708026.995, "dur": 24.470,
    "args": {
      "External id": 2482658,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 10361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708029.777, "dur": 2.450,
    "args": {
      "External id": 2482659,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708033.242, "dur": 17.960,
    "args": {
      "External id": 2482660,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[1], [1]], "Input Dims": [[8192], [0]], "Ev Idx": 10363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708035.980, "dur": 2.666,
    "args": {
      "External id": 2482661,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 10364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708052.624, "dur": 18.435,
    "args": {
      "External id": 2482662,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 10365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::masked_fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708072.301, "dur": 12.315,
    "args": {
      "External id": 2482663,"Record function id": 0, "Concrete Inputs": ["", "", "0"], "Input type": ["float", "bool", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 10366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sub", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708086.857, "dur": 11.875,
    "args": {
      "External id": 2482664,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1, 1], [1, 1], []], "Input Dims": [[8192, 1], [8192, 1], []], "Ev Idx": 10367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::exp_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708101.665, "dur": 11.551,
    "args": {
      "External id": 2482665,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1, 1]], "Input Dims": [[8192, 1]], "Ev Idx": 10368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708114.676, "dur": 17.189,
    "args": {
      "External id": 2482666,"Record function id": 0, "Concrete Inputs": ["", "[-1]", "False", "", ""], "Input type": ["float", "ScalarList", "Scalar", "", "float"], "Input Strides": [[1, 1], [], [], [], [1]], "Input Dims": [[8192, 1], [], [], [], [0]], "Ev Idx": 10369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708116.286, "dur": 1.093,
    "args": {
      "External id": 2482667,"Record function id": 0, "Concrete Inputs": ["", "[8192]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 10370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708118.963, "dur": 0.447,
    "args": {
      "External id": 2482668,"Record function id": 0, "Concrete Inputs": ["", "[8192, 1]", "[1, 0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[8192], [], [], []], "Ev Idx": 10371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::log_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708133.190, "dur": 10.418,
    "args": {
      "External id": 2482669,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[8192]], "Ev Idx": 10372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708144.381, "dur": 10.983,
    "args": {
      "External id": 2482670,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [1], []], "Input Dims": [[8192], [8192], []], "Ev Idx": 10373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708163.491, "dur": 1.899,
    "args": {
      "External id": 2482671,"Record function id": 0, "Concrete Inputs": ["", "[8192]"], "Input type": ["float", "ScalarList"], "Input Strides": [[1], []], "Input Dims": [[8192], []], "Ev Idx": 10374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708174.981, "dur": 3.660,
    "args": {
      "External id": 2482672,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["float", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], []], "Input Dims": [[65536], [], [], [], []], "Ev Idx": 10375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708177.376, "dur": 0.430,
    "args": {
      "External id": 2482673,"Record function id": 0, "Concrete Inputs": ["", "[8192]", "[1]", "57344"], "Input type": ["float", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 10376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708240.107, "dur": 49.897,
    "args": {
      "External id": 2482674,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[32000, 1], [2048, 1]], "Input Dims": [[8192, 32000], [32000, 2048]], "Ev Idx": 10377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::slice", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708294.855, "dur": 6.369,
    "args": {
      "External id": 2482675,"Record function id": 0, "Concrete Inputs": ["", "0", "57344", "65536", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar", "Scalar", "Scalar"], "Input Strides": [[2048, 1], [], [], [], []], "Input Dims": [[65536, 2048], [], [], [], []], "Ev Idx": 10378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708297.328, "dur": 2.566,
    "args": {
      "External id": 2482676,"Record function id": 0, "Concrete Inputs": ["", "[8192, 2048]", "[2048, 1]", "117440512"], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", "Scalar"], "Input Strides": [[2048, 1], [], [], []], "Input Dims": [[65536, 2048], [], [], []], "Ev Idx": 10379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708302.540, "dur": 22.158,
    "args": {
      "External id": 2482677,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[8192, 2048], [8192, 2048], []], "Ev Idx": 10380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::t", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708329.019, "dur": 6.826,
    "args": {
      "External id": 2482678,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[32000, 1]], "Input Dims": [[8192, 32000]], "Ev Idx": 10381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::transpose", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708330.566, "dur": 4.630,
    "args": {
      "External id": 2482679,"Record function id": 0, "Concrete Inputs": ["", "0", "1"], "Input type": ["c10::BFloat16", "Scalar", "Scalar"], "Input Strides": [[32000, 1], [], []], "Input Dims": [[8192, 32000], [], []], "Ev Idx": 10382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708333.855, "dur": 1.132,
    "args": {
      "External id": 2482680,"Record function id": 0, "Concrete Inputs": ["", "[32000, 8192]", "[1, 32000]", ""], "Input type": ["c10::BFloat16", "ScalarList", "ScalarList", ""], "Input Strides": [[32000, 1], [], [], []], "Input Dims": [[8192, 32000], [], [], []], "Ev Idx": 10383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::matmul", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708337.903, "dur": 37.790,
    "args": {
      "External id": 2482681,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 10384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mm", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708338.744, "dur": 36.402,
    "args": {
      "External id": 2482682,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["c10::BFloat16", "c10::BFloat16"], "Input Strides": [[1, 32000], [2048, 1]], "Input Dims": [[32000, 8192], [8192, 2048]], "Ev Idx": 10385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708379.202, "dur": 12.896,
    "args": {
      "External id": 2482683,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "c10::BFloat16", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 10386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708396.427, "dur": 20.859,
    "args": {
      "External id": 2482684,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[1], []], "Input Dims": [[65536], []], "Ev Idx": 10387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::sum", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708398.512, "dur": 18.403,
    "args": {
      "External id": 2482685,"Record function id": 0, "Concrete Inputs": ["", "[]", "False", ""], "Input type": ["float", "ScalarList", "Scalar", ""], "Input Strides": [[1], [], [], []], "Input Dims": [[65536], [], [], []], "Ev Idx": 10388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708403.313, "dur": 0.750,
    "args": {
      "External id": 2482686,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708421.812, "dur": 62.137,
    "args": {
      "External id": 2482687,"Record function id": 0, "Concrete Inputs": ["", "", "15", "False", "False", ""], "Input type": ["float", "", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], []], "Input Dims": [[32000, 2048], [], [], [], [], []], "Ev Idx": 10390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_to_copy", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708423.485, "dur": 60.188,
    "args": {
      "External id": 2482688,"Record function id": 0, "Concrete Inputs": ["", "15", "", "", "", "False", ""], "Input type": ["float", "Scalar", "", "", "", "Scalar", ""], "Input Strides": [[2048, 1], [], [], [], [], [], []], "Input Dims": [[32000, 2048], [], [], [], [], [], []], "Ev Idx": 10391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708461.739, "dur": 7.260,
    "args": {
      "External id": 2482689,"Record function id": 0, "Concrete Inputs": ["[32000, 2048]", "[2048, 1]", "15", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708470.124, "dur": 13.017,
    "args": {
      "External id": 2482690,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["c10::BFloat16", "float", "Scalar"], "Input Strides": [[2048, 1], [2048, 1], []], "Input Dims": [[32000, 2048], [32000, 2048], []], "Ev Idx": 10393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708495.680, "dur": 4.284,
    "args": {
      "External id": 2482691,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 10394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708497.190, "dur": 2.487,
    "args": {
      "External id": 2482692,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[65536, 2048]], "Ev Idx": 10395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::detach", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708500.991, "dur": 0.842,
    "args": {
      "External id": 2482693,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 10396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "detach", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708501.260, "dur": 0.404,
    "args": {
      "External id": 2482694,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["c10::BFloat16"], "Input Strides": [[2048, 1]], "Input Dims": [[32000, 2048]], "Ev Idx": 10397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708560.370, "dur": 26.279,
    "args": {
      "External id": 2482695,"Sequence number": 24740648, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1], [], []], "Ev Idx": 10398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708588.574, "dur": 12.738,
    "args": {
      "External id": 2482696,"Sequence number": 24740649, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1], [], []], "Ev Idx": 10399
    }
  },
  {
    "ph": "s", "id": 2, "pid": 1336759, "tid": 1336759, "ts": 1514656708588.574,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "FSDP::post_forward", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708705.561, "dur": 35.001,
    "args": {
      "External id": 2482697,"Record function id": 0, "Ev Idx": 10400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::div", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708840.432, "dur": 33.493,
    "args": {
      "External id": 2482698,"Sequence number": 24740650, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "long int"], "Input Strides": [[1], []], "Input Dims": [[1], []], "Ev Idx": 10401
    }
  },
  {
    "ph": "s", "id": 1, "pid": 1336759, "tid": 1336759, "ts": 1514656708840.432,
    "cat": "fwdbwd", "name": "fwdbwd"
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::ones_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708927.495, "dur": 25.509,
    "args": {
      "External id": 2482699,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "1"], "Input type": ["float", "", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[1], [], [], [], [], []], "Ev Idx": 10402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708929.011, "dur": 8.103,
    "args": {
      "External id": 2482700,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "False", "1"], "Input type": ["float", "", "", "", "Scalar", "Scalar"], "Input Strides": [[1], [], [], [], [], []], "Input Dims": [[1], [], [], [], [], []], "Ev Idx": 10403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708932.485, "dur": 4.007,
    "args": {
      "External id": 2482701,"Record function id": 0, "Concrete Inputs": ["[1]", "[1]", "6", "0", "", "False"], "Input type": ["ScalarList", "ScalarList", "Scalar", "Scalar", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514656708938.514, "dur": 14.209,
    "args": {
      "External id": 2482702,"Record function id": 0, "Concrete Inputs": ["", "1."], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[1], []], "Ev Idx": 10405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::div", "pid": 1336759, "tid": 1336759,
    "ts": 1514658027096.903, "dur": 48.960,
    "args": {
      "External id": 2482703,"Sequence number": 24740651, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "long int"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::div", "pid": 1336759, "tid": 1336759,
    "ts": 1514658027154.560, "dur": 16.462,
    "args": {
      "External id": 2482704,"Sequence number": 24740652, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "long int"], "Input Strides": [[1], []], "Input Dims": [[1], []], "Ev Idx": 10407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 1336759, "tid": 1336759,
    "ts": 1514658027177.819, "dur": 102.320,
    "args": {
      "External id": 2482705,"Sequence number": 24740653, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "long int", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1], [], []], "Ev Idx": 10408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 1336759, "tid": 1336759,
    "ts": 1514658027709.764, "dur": 27.579,
    "args": {
      "External id": 2482706,"Sequence number": 24740654, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "long int", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 10409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 1336759, "tid": 1336759,
    "ts": 1514658028083.424, "dur": 30.115,
    "args": {
      "External id": 2482707,"Sequence number": 24740655, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "long int", "Scalar"], "Input Strides": [[1], [], []], "Input Dims": [[1], [], []], "Ev Idx": 10410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_norm", "pid": 1336759, "tid": 1336759,
    "ts": 1514658029373.634, "dur": 2876.463,
    "args": {
      "External id": 2482708,"Record function id": 0, "Concrete Inputs": ["", "2.", ""], "Input type": ["TensorList", "Scalar", ""], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 10411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_norm", "pid": 1336759, "tid": 1336759,
    "ts": 1514658029955.887, "dur": 827.613,
    "args": {
      "External id": 2482709,"Record function id": 0, "Concrete Inputs": ["", "2.", ""], "Input type": ["TensorList", "Scalar", ""], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 10412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zeros", "pid": 1336759, "tid": 1336759,
    "ts": 1514658029977.732, "dur": 102.965,
    "args": {
      "External id": 2482710,"Record function id": 0, "Concrete Inputs": ["[36375]", "6", "0", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 10413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514658029981.202, "dur": 43.367,
    "args": {
      "External id": 2482711,"Record function id": 0, "Concrete Inputs": ["[36375]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 10414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::zero_", "pid": 1336759, "tid": 1336759,
    "ts": 1514658030027.810, "dur": 52.546,
    "args": {
      "External id": 2482712,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[1]], "Input Dims": [[36375]], "Ev Idx": 10415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::fill_", "pid": 1336759, "tid": 1336759,
    "ts": 1514658030030.413, "dur": 49.398,
    "args": {
      "External id": 2482713,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[1], []], "Input Dims": [[36375], []], "Ev Idx": 10416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032284.557, "dur": 4.546,
    "args": {
      "External id": 2482714,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032291.532, "dur": 0.465,
    "args": {
      "External id": 2482715,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032293.278, "dur": 0.409,
    "args": {
      "External id": 2482716,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032294.887, "dur": 0.235,
    "args": {
      "External id": 2482717,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032296.242, "dur": 0.323,
    "args": {
      "External id": 2482718,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032297.941, "dur": 0.312,
    "args": {
      "External id": 2482719,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032299.252, "dur": 0.226,
    "args": {
      "External id": 2482720,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032302.069, "dur": 0.195,
    "args": {
      "External id": 2482721,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032303.187, "dur": 0.402,
    "args": {
      "External id": 2482722,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032304.566, "dur": 0.225,
    "args": {
      "External id": 2482723,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032305.605, "dur": 0.354,
    "args": {
      "External id": 2482724,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032306.801, "dur": 0.486,
    "args": {
      "External id": 2482725,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032308.300, "dur": 0.462,
    "args": {
      "External id": 2482726,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032309.618, "dur": 0.360,
    "args": {
      "External id": 2482727,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032310.727, "dur": 0.326,
    "args": {
      "External id": 2482728,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032313.927, "dur": 0.293,
    "args": {
      "External id": 2482729,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032315.119, "dur": 0.255,
    "args": {
      "External id": 2482730,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032316.315, "dur": 0.194,
    "args": {
      "External id": 2482731,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032317.357, "dur": 0.192,
    "args": {
      "External id": 2482732,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032318.318, "dur": 0.194,
    "args": {
      "External id": 2482733,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032319.425, "dur": 0.232,
    "args": {
      "External id": 2482734,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032320.429, "dur": 0.197,
    "args": {
      "External id": 2482735,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032321.451, "dur": 0.197,
    "args": {
      "External id": 2482736,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032324.270, "dur": 0.233,
    "args": {
      "External id": 2482737,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032325.371, "dur": 0.194,
    "args": {
      "External id": 2482738,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032326.641, "dur": 0.196,
    "args": {
      "External id": 2482739,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032327.720, "dur": 0.194,
    "args": {
      "External id": 2482740,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032328.976, "dur": 0.205,
    "args": {
      "External id": 2482741,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032330.178, "dur": 0.199,
    "args": {
      "External id": 2482742,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032331.370, "dur": 0.195,
    "args": {
      "External id": 2482743,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032332.421, "dur": 0.232,
    "args": {
      "External id": 2482744,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032335.114, "dur": 0.278,
    "args": {
      "External id": 2482745,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032336.424, "dur": 0.368,
    "args": {
      "External id": 2482746,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032337.787, "dur": 0.200,
    "args": {
      "External id": 2482747,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032338.807, "dur": 0.320,
    "args": {
      "External id": 2482748,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032339.925, "dur": 0.440,
    "args": {
      "External id": 2482749,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032341.274, "dur": 0.175,
    "args": {
      "External id": 2482750,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032342.221, "dur": 0.313,
    "args": {
      "External id": 2482751,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032343.342, "dur": 0.365,
    "args": {
      "External id": 2482752,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032345.850, "dur": 0.307,
    "args": {
      "External id": 2482753,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032346.980, "dur": 0.321,
    "args": {
      "External id": 2482754,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032348.418, "dur": 0.196,
    "args": {
      "External id": 2482755,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032349.430, "dur": 0.317,
    "args": {
      "External id": 2482756,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032350.567, "dur": 0.196,
    "args": {
      "External id": 2482757,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032351.535, "dur": 0.196,
    "args": {
      "External id": 2482758,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032353.752, "dur": 0.377,
    "args": {
      "External id": 2482759,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032354.935, "dur": 0.332,
    "args": {
      "External id": 2482760,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032358.326, "dur": 0.306,
    "args": {
      "External id": 2482761,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032359.421, "dur": 0.308,
    "args": {
      "External id": 2482762,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032360.515, "dur": 0.243,
    "args": {
      "External id": 2482763,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032361.522, "dur": 0.368,
    "args": {
      "External id": 2482764,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032362.649, "dur": 0.292,
    "args": {
      "External id": 2482765,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032363.699, "dur": 0.326,
    "args": {
      "External id": 2482766,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032365.076, "dur": 0.329,
    "args": {
      "External id": 2482767,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032366.185, "dur": 0.342,
    "args": {
      "External id": 2482768,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032369.555, "dur": 0.195,
    "args": {
      "External id": 2482769,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032370.592, "dur": 0.203,
    "args": {
      "External id": 2482770,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032371.667, "dur": 0.191,
    "args": {
      "External id": 2482771,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032372.628, "dur": 0.194,
    "args": {
      "External id": 2482772,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032373.629, "dur": 0.192,
    "args": {
      "External id": 2482773,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032374.841, "dur": 0.220,
    "args": {
      "External id": 2482774,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032375.974, "dur": 0.214,
    "args": {
      "External id": 2482775,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032377.013, "dur": 0.194,
    "args": {
      "External id": 2482776,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032379.606, "dur": 0.229,
    "args": {
      "External id": 2482777,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032380.684, "dur": 0.337,
    "args": {
      "External id": 2482778,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032382.149, "dur": 0.347,
    "args": {
      "External id": 2482779,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032383.373, "dur": 0.192,
    "args": {
      "External id": 2482780,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032384.344, "dur": 0.223,
    "args": {
      "External id": 2482781,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032385.392, "dur": 0.222,
    "args": {
      "External id": 2482782,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032386.496, "dur": 0.192,
    "args": {
      "External id": 2482783,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032387.683, "dur": 0.191,
    "args": {
      "External id": 2482784,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032390.274, "dur": 0.310,
    "args": {
      "External id": 2482785,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032391.453, "dur": 0.327,
    "args": {
      "External id": 2482786,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032392.957, "dur": 0.197,
    "args": {
      "External id": 2482787,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032394.078, "dur": 0.226,
    "args": {
      "External id": 2482788,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032395.305, "dur": 0.197,
    "args": {
      "External id": 2482789,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032396.302, "dur": 0.195,
    "args": {
      "External id": 2482790,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032398.080, "dur": 0.194,
    "args": {
      "External id": 2482791,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032399.030, "dur": 0.305,
    "args": {
      "External id": 2482792,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032401.914, "dur": 0.306,
    "args": {
      "External id": 2482793,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032403.014, "dur": 0.468,
    "args": {
      "External id": 2482794,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032404.242, "dur": 0.191,
    "args": {
      "External id": 2482795,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032405.219, "dur": 0.295,
    "args": {
      "External id": 2482796,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032406.267, "dur": 0.289,
    "args": {
      "External id": 2482797,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032407.317, "dur": 0.319,
    "args": {
      "External id": 2482798,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032408.411, "dur": 0.313,
    "args": {
      "External id": 2482799,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032409.498, "dur": 0.313,
    "args": {
      "External id": 2482800,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032412.359, "dur": 0.326,
    "args": {
      "External id": 2482801,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032413.574, "dur": 0.204,
    "args": {
      "External id": 2482802,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032414.530, "dur": 0.192,
    "args": {
      "External id": 2482803,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032415.471, "dur": 0.191,
    "args": {
      "External id": 2482804,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032416.440, "dur": 0.217,
    "args": {
      "External id": 2482805,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032417.415, "dur": 0.211,
    "args": {
      "External id": 2482806,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032419.236, "dur": 0.225,
    "args": {
      "External id": 2482807,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032420.356, "dur": 0.197,
    "args": {
      "External id": 2482808,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032422.871, "dur": 0.195,
    "args": {
      "External id": 2482809,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032423.932, "dur": 0.224,
    "args": {
      "External id": 2482810,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032424.997, "dur": 0.200,
    "args": {
      "External id": 2482811,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032426.072, "dur": 0.195,
    "args": {
      "External id": 2482812,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032427.037, "dur": 0.195,
    "args": {
      "External id": 2482813,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032428.013, "dur": 0.194,
    "args": {
      "External id": 2482814,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032429.287, "dur": 0.205,
    "args": {
      "External id": 2482815,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032430.453, "dur": 0.196,
    "args": {
      "External id": 2482816,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032433.109, "dur": 0.223,
    "args": {
      "External id": 2482817,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032434.108, "dur": 0.191,
    "args": {
      "External id": 2482818,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032435.101, "dur": 0.195,
    "args": {
      "External id": 2482819,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032436.067, "dur": 0.196,
    "args": {
      "External id": 2482820,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032437.077, "dur": 0.334,
    "args": {
      "External id": 2482821,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032438.266, "dur": 0.316,
    "args": {
      "External id": 2482822,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032439.880, "dur": 0.197,
    "args": {
      "External id": 2482823,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032440.925, "dur": 0.320,
    "args": {
      "External id": 2482824,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032443.765, "dur": 0.322,
    "args": {
      "External id": 2482825,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032444.989, "dur": 0.514,
    "args": {
      "External id": 2482826,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032446.483, "dur": 0.316,
    "args": {
      "External id": 2482827,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032447.651, "dur": 0.312,
    "args": {
      "External id": 2482828,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032448.970, "dur": 0.363,
    "args": {
      "External id": 2482829,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032450.207, "dur": 0.399,
    "args": {
      "External id": 2482830,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032451.451, "dur": 0.193,
    "args": {
      "External id": 2482831,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032452.404, "dur": 0.356,
    "args": {
      "External id": 2482832,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032455.211, "dur": 0.302,
    "args": {
      "External id": 2482833,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10536
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032456.290, "dur": 0.334,
    "args": {
      "External id": 2482834,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032457.472, "dur": 0.230,
    "args": {
      "External id": 2482835,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032458.516, "dur": 0.255,
    "args": {
      "External id": 2482836,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032459.616, "dur": 0.258,
    "args": {
      "External id": 2482837,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032460.757, "dur": 0.227,
    "args": {
      "External id": 2482838,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032461.771, "dur": 0.339,
    "args": {
      "External id": 2482839,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032462.922, "dur": 0.343,
    "args": {
      "External id": 2482840,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032465.864, "dur": 0.194,
    "args": {
      "External id": 2482841,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032466.938, "dur": 0.195,
    "args": {
      "External id": 2482842,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032468.273, "dur": 0.198,
    "args": {
      "External id": 2482843,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032469.224, "dur": 0.195,
    "args": {
      "External id": 2482844,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032470.245, "dur": 0.200,
    "args": {
      "External id": 2482845,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032471.353, "dur": 0.198,
    "args": {
      "External id": 2482846,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032472.331, "dur": 0.197,
    "args": {
      "External id": 2482847,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032473.376, "dur": 0.201,
    "args": {
      "External id": 2482848,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032475.916, "dur": 0.190,
    "args": {
      "External id": 2482849,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032476.964, "dur": 0.205,
    "args": {
      "External id": 2482850,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032478.047, "dur": 0.389,
    "args": {
      "External id": 2482851,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032479.232, "dur": 0.353,
    "args": {
      "External id": 2482852,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032480.342, "dur": 0.333,
    "args": {
      "External id": 2482853,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032481.468, "dur": 0.354,
    "args": {
      "External id": 2482854,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032482.596, "dur": 0.195,
    "args": {
      "External id": 2482855,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032483.581, "dur": 0.196,
    "args": {
      "External id": 2482856,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032486.032, "dur": 0.195,
    "args": {
      "External id": 2482857,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032487.036, "dur": 0.190,
    "args": {
      "External id": 2482858,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032488.000, "dur": 0.192,
    "args": {
      "External id": 2482859,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032488.934, "dur": 0.194,
    "args": {
      "External id": 2482860,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032489.907, "dur": 0.198,
    "args": {
      "External id": 2482861,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032490.859, "dur": 0.192,
    "args": {
      "External id": 2482862,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032492.121, "dur": 0.190,
    "args": {
      "External id": 2482863,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032493.082, "dur": 0.192,
    "args": {
      "External id": 2482864,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032495.737, "dur": 0.233,
    "args": {
      "External id": 2482865,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032496.744, "dur": 0.198,
    "args": {
      "External id": 2482866,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032497.939, "dur": 0.217,
    "args": {
      "External id": 2482867,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032499.054, "dur": 0.190,
    "args": {
      "External id": 2482868,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032500.280, "dur": 0.235,
    "args": {
      "External id": 2482869,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032501.333, "dur": 0.199,
    "args": {
      "External id": 2482870,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032502.429, "dur": 0.193,
    "args": {
      "External id": 2482871,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032503.461, "dur": 0.193,
    "args": {
      "External id": 2482872,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032507.102, "dur": 0.240,
    "args": {
      "External id": 2482873,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032508.250, "dur": 0.193,
    "args": {
      "External id": 2482874,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032509.219, "dur": 0.192,
    "args": {
      "External id": 2482875,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032510.195, "dur": 0.297,
    "args": {
      "External id": 2482876,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032511.282, "dur": 0.326,
    "args": {
      "External id": 2482877,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032512.391, "dur": 0.333,
    "args": {
      "External id": 2482878,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032513.473, "dur": 0.383,
    "args": {
      "External id": 2482879,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032514.608, "dur": 0.325,
    "args": {
      "External id": 2482880,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032517.302, "dur": 0.318,
    "args": {
      "External id": 2482881,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032518.388, "dur": 0.338,
    "args": {
      "External id": 2482882,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032519.542, "dur": 0.198,
    "args": {
      "External id": 2482883,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032520.499, "dur": 0.316,
    "args": {
      "External id": 2482884,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032521.582, "dur": 0.315,
    "args": {
      "External id": 2482885,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032522.666, "dur": 0.317,
    "args": {
      "External id": 2482886,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032524.098, "dur": 0.197,
    "args": {
      "External id": 2482887,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032525.154, "dur": 0.197,
    "args": {
      "External id": 2482888,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032527.660, "dur": 0.196,
    "args": {
      "External id": 2482889,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032542.440, "dur": 0.618,
    "args": {
      "External id": 2482890,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032543.991, "dur": 0.211,
    "args": {
      "External id": 2482891,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032544.962, "dur": 0.208,
    "args": {
      "External id": 2482892,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032545.986, "dur": 0.355,
    "args": {
      "External id": 2482893,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032547.095, "dur": 0.349,
    "args": {
      "External id": 2482894,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032548.322, "dur": 0.215,
    "args": {
      "External id": 2482895,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032549.295, "dur": 0.205,
    "args": {
      "External id": 2482896,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032551.863, "dur": 0.223,
    "args": {
      "External id": 2482897,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032552.931, "dur": 0.201,
    "args": {
      "External id": 2482898,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032554.037, "dur": 0.350,
    "args": {
      "External id": 2482899,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032555.171, "dur": 0.346,
    "args": {
      "External id": 2482900,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032556.354, "dur": 0.332,
    "args": {
      "External id": 2482901,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032557.431, "dur": 0.444,
    "args": {
      "External id": 2482902,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032558.612, "dur": 0.319,
    "args": {
      "External id": 2482903,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032559.734, "dur": 0.434,
    "args": {
      "External id": 2482904,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032562.734, "dur": 0.375,
    "args": {
      "External id": 2482905,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032564.017, "dur": 0.325,
    "args": {
      "External id": 2482906,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032565.315, "dur": 0.209,
    "args": {
      "External id": 2482907,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032566.301, "dur": 0.331,
    "args": {
      "External id": 2482908,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032567.500, "dur": 0.312,
    "args": {
      "External id": 2482909,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032568.635, "dur": 0.309,
    "args": {
      "External id": 2482910,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032569.695, "dur": 0.357,
    "args": {
      "External id": 2482911,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032570.891, "dur": 0.328,
    "args": {
      "External id": 2482912,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032573.612, "dur": 0.327,
    "args": {
      "External id": 2482913,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032574.791, "dur": 0.309,
    "args": {
      "External id": 2482914,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032576.473, "dur": 0.208,
    "args": {
      "External id": 2482915,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032577.522, "dur": 0.207,
    "args": {
      "External id": 2482916,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032578.467, "dur": 0.211,
    "args": {
      "External id": 2482917,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032579.419, "dur": 0.209,
    "args": {
      "External id": 2482918,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032580.387, "dur": 0.217,
    "args": {
      "External id": 2482919,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032581.351, "dur": 0.207,
    "args": {
      "External id": 2482920,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032584.362, "dur": 0.217,
    "args": {
      "External id": 2482921,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032585.326, "dur": 0.196,
    "args": {
      "External id": 2482922,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032586.277, "dur": 0.205,
    "args": {
      "External id": 2482923,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032587.282, "dur": 0.202,
    "args": {
      "External id": 2482924,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032588.218, "dur": 0.206,
    "args": {
      "External id": 2482925,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032589.164, "dur": 0.203,
    "args": {
      "External id": 2482926,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032590.105, "dur": 0.211,
    "args": {
      "External id": 2482927,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032591.074, "dur": 0.201,
    "args": {
      "External id": 2482928,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032593.436, "dur": 0.345,
    "args": {
      "External id": 2482929,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032594.532, "dur": 0.305,
    "args": {
      "External id": 2482930,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032595.610, "dur": 0.351,
    "args": {
      "External id": 2482931,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10634
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032596.704, "dur": 0.344,
    "args": {
      "External id": 2482932,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10635
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032597.790, "dur": 0.361,
    "args": {
      "External id": 2482933,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10636
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032598.913, "dur": 0.207,
    "args": {
      "External id": 2482934,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10637
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032599.860, "dur": 0.210,
    "args": {
      "External id": 2482935,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10638
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032600.901, "dur": 0.225,
    "args": {
      "External id": 2482936,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10639
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032603.295, "dur": 0.208,
    "args": {
      "External id": 2482937,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10640
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032604.287, "dur": 0.198,
    "args": {
      "External id": 2482938,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10641
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032605.244, "dur": 0.210,
    "args": {
      "External id": 2482939,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10642
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032606.237, "dur": 0.200,
    "args": {
      "External id": 2482940,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10643
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032607.252, "dur": 0.209,
    "args": {
      "External id": 2482941,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10644
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032608.284, "dur": 0.206,
    "args": {
      "External id": 2482942,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10645
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032609.337, "dur": 0.210,
    "args": {
      "External id": 2482943,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10646
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032610.430, "dur": 0.202,
    "args": {
      "External id": 2482944,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10647
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032612.702, "dur": 0.225,
    "args": {
      "External id": 2482945,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10648
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032613.841, "dur": 0.211,
    "args": {
      "External id": 2482946,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10649
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032615.217, "dur": 0.206,
    "args": {
      "External id": 2482947,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10650
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032616.255, "dur": 0.205,
    "args": {
      "External id": 2482948,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10651
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032617.296, "dur": 0.214,
    "args": {
      "External id": 2482949,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10652
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032618.412, "dur": 0.205,
    "args": {
      "External id": 2482950,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10653
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032619.378, "dur": 0.216,
    "args": {
      "External id": 2482951,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10654
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032620.363, "dur": 0.208,
    "args": {
      "External id": 2482952,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10655
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032622.449, "dur": 0.215,
    "args": {
      "External id": 2482953,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10656
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032623.584, "dur": 0.197,
    "args": {
      "External id": 2482954,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10657
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032624.623, "dur": 0.207,
    "args": {
      "External id": 2482955,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10658
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032625.675, "dur": 0.202,
    "args": {
      "External id": 2482956,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10659
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032626.864, "dur": 0.206,
    "args": {
      "External id": 2482957,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10660
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032627.874, "dur": 0.222,
    "args": {
      "External id": 2482958,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10661
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032628.923, "dur": 0.210,
    "args": {
      "External id": 2482959,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10662
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032629.995, "dur": 0.217,
    "args": {
      "External id": 2482960,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10663
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032632.442, "dur": 0.348,
    "args": {
      "External id": 2482961,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10664
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032633.525, "dur": 0.318,
    "args": {
      "External id": 2482962,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10665
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032634.596, "dur": 0.205,
    "args": {
      "External id": 2482963,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10666
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032635.559, "dur": 0.311,
    "args": {
      "External id": 2482964,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10667
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032636.648, "dur": 0.316,
    "args": {
      "External id": 2482965,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10668
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032637.705, "dur": 0.267,
    "args": {
      "External id": 2482966,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10669
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032638.696, "dur": 0.330,
    "args": {
      "External id": 2482967,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10670
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032639.752, "dur": 0.318,
    "args": {
      "External id": 2482968,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10671
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032642.425, "dur": 0.347,
    "args": {
      "External id": 2482969,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10672
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032646.533, "dur": 0.346,
    "args": {
      "External id": 2482970,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10673
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032647.700, "dur": 0.243,
    "args": {
      "External id": 2482971,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10674
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032648.708, "dur": 0.240,
    "args": {
      "External id": 2482972,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10675
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032649.722, "dur": 0.240,
    "args": {
      "External id": 2482973,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10676
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032650.805, "dur": 0.203,
    "args": {
      "External id": 2482974,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10677
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032651.786, "dur": 0.205,
    "args": {
      "External id": 2482975,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10678
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032652.735, "dur": 0.219,
    "args": {
      "External id": 2482976,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10679
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032655.443, "dur": 0.206,
    "args": {
      "External id": 2482977,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10680
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032656.391, "dur": 0.299,
    "args": {
      "External id": 2482978,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10681
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032657.443, "dur": 0.205,
    "args": {
      "External id": 2482979,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10682
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032658.428, "dur": 0.321,
    "args": {
      "External id": 2482980,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10683
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032659.489, "dur": 0.348,
    "args": {
      "External id": 2482981,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10684
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032660.574, "dur": 0.305,
    "args": {
      "External id": 2482982,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10685
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032662.016, "dur": 0.357,
    "args": {
      "External id": 2482983,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10686
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032663.133, "dur": 0.328,
    "args": {
      "External id": 2482984,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10687
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032665.540, "dur": 0.427,
    "args": {
      "External id": 2482985,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10688
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032666.714, "dur": 0.322,
    "args": {
      "External id": 2482986,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10689
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032668.901, "dur": 0.388,
    "args": {
      "External id": 2482987,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10690
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032670.066, "dur": 0.321,
    "args": {
      "External id": 2482988,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10691
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032671.122, "dur": 0.372,
    "args": {
      "External id": 2482989,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10692
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032672.234, "dur": 0.338,
    "args": {
      "External id": 2482990,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10693
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032673.351, "dur": 0.205,
    "args": {
      "External id": 2482991,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10694
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032674.292, "dur": 0.200,
    "args": {
      "External id": 2482992,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10695
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032676.839, "dur": 0.209,
    "args": {
      "External id": 2482993,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10696
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032677.793, "dur": 0.194,
    "args": {
      "External id": 2482994,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10697
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032678.732, "dur": 0.203,
    "args": {
      "External id": 2482995,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10698
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032679.714, "dur": 0.253,
    "args": {
      "External id": 2482996,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10699
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032680.859, "dur": 0.205,
    "args": {
      "External id": 2482997,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10700
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032681.802, "dur": 0.201,
    "args": {
      "External id": 2482998,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10701
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032682.819, "dur": 0.207,
    "args": {
      "External id": 2482999,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10702
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032683.861, "dur": 0.206,
    "args": {
      "External id": 2483000,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10703
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032686.567, "dur": 0.357,
    "args": {
      "External id": 2483001,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10704
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032687.665, "dur": 0.195,
    "args": {
      "External id": 2483002,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10705
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032688.604, "dur": 0.203,
    "args": {
      "External id": 2483003,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10706
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032689.578, "dur": 0.217,
    "args": {
      "External id": 2483004,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 10707
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::stack", "pid": 1336759, "tid": 1336759,
    "ts": 1514658032745.909, "dur": 1490.051,
    "args": {
      "External id": 2483005,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["TensorList", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10708
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::stack", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033194.180, "dur": 972.729,
    "args": {
      "External id": 2483006,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["TensorList", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10709
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033202.697, "dur": 7.268,
    "args": {
      "External id": 2483007,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10710
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033206.541, "dur": 2.861,
    "args": {
      "External id": 2483008,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10711
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033210.685, "dur": 3.555,
    "args": {
      "External id": 2483009,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10712
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033211.763, "dur": 2.370,
    "args": {
      "External id": 2483010,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10713
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033214.747, "dur": 2.987,
    "args": {
      "External id": 2483011,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10714
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033216.840, "dur": 0.600,
    "args": {
      "External id": 2483012,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10715
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033218.115, "dur": 1.370,
    "args": {
      "External id": 2483013,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10716
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033218.684, "dur": 0.732,
    "args": {
      "External id": 2483014,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10717
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033221.337, "dur": 3.540,
    "args": {
      "External id": 2483015,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10718
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033224.387, "dur": 0.425,
    "args": {
      "External id": 2483016,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10719
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033225.478, "dur": 1.275,
    "args": {
      "External id": 2483017,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10720
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033226.137, "dur": 0.547,
    "args": {
      "External id": 2483018,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10721
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033227.248, "dur": 1.473,
    "args": {
      "External id": 2483019,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10722
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033227.960, "dur": 0.676,
    "args": {
      "External id": 2483020,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10723
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033229.322, "dur": 3.508,
    "args": {
      "External id": 2483021,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10724
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033232.035, "dur": 0.724,
    "args": {
      "External id": 2483022,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10725
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033233.184, "dur": 1.396,
    "args": {
      "External id": 2483023,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10726
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033233.794, "dur": 0.717,
    "args": {
      "External id": 2483024,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10727
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033235.081, "dur": 2.456,
    "args": {
      "External id": 2483025,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10728
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033235.687, "dur": 1.765,
    "args": {
      "External id": 2483026,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10729
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033237.770, "dur": 2.838,
    "args": {
      "External id": 2483027,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10730
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033239.842, "dur": 0.697,
    "args": {
      "External id": 2483028,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10731
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033240.839, "dur": 1.472,
    "args": {
      "External id": 2483029,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10732
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033241.406, "dur": 0.835,
    "args": {
      "External id": 2483030,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10733
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033243.999, "dur": 4.024,
    "args": {
      "External id": 2483031,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10734
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033247.505, "dur": 0.452,
    "args": {
      "External id": 2483032,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10735
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033248.268, "dur": 1.220,
    "args": {
      "External id": 2483033,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10736
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033248.780, "dur": 0.642,
    "args": {
      "External id": 2483034,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10737
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033249.739, "dur": 1.309,
    "args": {
      "External id": 2483035,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10738
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033250.331, "dur": 0.649,
    "args": {
      "External id": 2483036,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10739
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033251.277, "dur": 3.057,
    "args": {
      "External id": 2483037,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10740
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033253.772, "dur": 0.496,
    "args": {
      "External id": 2483038,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10741
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033254.705, "dur": 0.965,
    "args": {
      "External id": 2483039,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10742
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033255.177, "dur": 0.427,
    "args": {
      "External id": 2483040,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10743
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033256.020, "dur": 2.685,
    "args": {
      "External id": 2483041,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10744
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033256.655, "dur": 1.977,
    "args": {
      "External id": 2483042,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10745
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033259.057, "dur": 2.341,
    "args": {
      "External id": 2483043,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10746
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033260.900, "dur": 0.432,
    "args": {
      "External id": 2483044,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10747
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033261.628, "dur": 1.356,
    "args": {
      "External id": 2483045,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10748
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033262.015, "dur": 0.899,
    "args": {
      "External id": 2483046,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10749
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033264.908, "dur": 4.107,
    "args": {
      "External id": 2483047,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10750
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033268.275, "dur": 0.671,
    "args": {
      "External id": 2483048,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10751
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033269.273, "dur": 1.105,
    "args": {
      "External id": 2483049,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10752
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033269.763, "dur": 0.549,
    "args": {
      "External id": 2483050,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10753
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033270.639, "dur": 1.233,
    "args": {
      "External id": 2483051,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10754
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033271.168, "dur": 0.639,
    "args": {
      "External id": 2483052,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10755
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033272.101, "dur": 3.189,
    "args": {
      "External id": 2483053,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10756
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033274.617, "dur": 0.606,
    "args": {
      "External id": 2483054,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10757
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033275.815, "dur": 1.113,
    "args": {
      "External id": 2483055,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10758
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033276.324, "dur": 0.532,
    "args": {
      "External id": 2483056,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10759
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033277.432, "dur": 2.538,
    "args": {
      "External id": 2483057,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10760
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033278.105, "dur": 1.790,
    "args": {
      "External id": 2483058,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10761
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033280.233, "dur": 2.654,
    "args": {
      "External id": 2483059,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10762
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033282.204, "dur": 0.616,
    "args": {
      "External id": 2483060,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10763
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033283.130, "dur": 1.438,
    "args": {
      "External id": 2483061,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10764
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033283.674, "dur": 0.823,
    "args": {
      "External id": 2483062,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10765
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033285.999, "dur": 3.865,
    "args": {
      "External id": 2483063,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10766
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033289.029, "dur": 0.763,
    "args": {
      "External id": 2483064,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10767
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033290.108, "dur": 1.275,
    "args": {
      "External id": 2483065,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10768
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033290.597, "dur": 0.719,
    "args": {
      "External id": 2483066,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10769
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033291.630, "dur": 1.073,
    "args": {
      "External id": 2483067,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10770
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033292.126, "dur": 0.507,
    "args": {
      "External id": 2483068,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10771
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033292.931, "dur": 3.237,
    "args": {
      "External id": 2483069,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10772
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033295.337, "dur": 0.766,
    "args": {
      "External id": 2483070,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10773
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033296.520, "dur": 1.067,
    "args": {
      "External id": 2483071,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10774
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033297.075, "dur": 0.447,
    "args": {
      "External id": 2483072,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10775
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033297.928, "dur": 2.782,
    "args": {
      "External id": 2483073,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10776
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033298.763, "dur": 1.876,
    "args": {
      "External id": 2483074,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10777
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033301.041, "dur": 3.816,
    "args": {
      "External id": 2483075,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10778
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033304.332, "dur": 0.463,
    "args": {
      "External id": 2483076,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10779
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033305.091, "dur": 2.598,
    "args": {
      "External id": 2483077,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10780
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033305.621, "dur": 1.798,
    "args": {
      "External id": 2483078,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10781
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033309.769, "dur": 2.209,
    "args": {
      "External id": 2483079,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10782
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033311.466, "dur": 0.446,
    "args": {
      "External id": 2483080,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10783
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033312.306, "dur": 1.277,
    "args": {
      "External id": 2483081,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10784
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033312.928, "dur": 0.587,
    "args": {
      "External id": 2483082,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10785
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033313.949, "dur": 2.830,
    "args": {
      "External id": 2483083,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10786
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033316.004, "dur": 0.709,
    "args": {
      "External id": 2483084,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10787
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033317.174, "dur": 2.560,
    "args": {
      "External id": 2483085,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10788
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033318.971, "dur": 0.698,
    "args": {
      "External id": 2483086,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10789
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033320.089, "dur": 1.192,
    "args": {
      "External id": 2483087,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10790
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033320.590, "dur": 0.622,
    "args": {
      "External id": 2483088,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10791
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033321.511, "dur": 2.830,
    "args": {
      "External id": 2483089,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10792
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033323.572, "dur": 0.703,
    "args": {
      "External id": 2483090,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10793
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033324.593, "dur": 2.763,
    "args": {
      "External id": 2483091,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10794
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033326.309, "dur": 0.986,
    "args": {
      "External id": 2483092,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10795
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033327.586, "dur": 2.344,
    "args": {
      "External id": 2483093,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10796
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033328.114, "dur": 1.745,
    "args": {
      "External id": 2483094,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10797
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033331.799, "dur": 2.182,
    "args": {
      "External id": 2483095,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10798
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033333.436, "dur": 0.477,
    "args": {
      "External id": 2483096,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10799
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033334.546, "dur": 1.250,
    "args": {
      "External id": 2483097,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10800
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033335.008, "dur": 0.720,
    "args": {
      "External id": 2483098,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10801
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033336.150, "dur": 3.173,
    "args": {
      "External id": 2483099,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10802
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033338.614, "dur": 0.644,
    "args": {
      "External id": 2483100,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10803
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033339.688, "dur": 1.288,
    "args": {
      "External id": 2483101,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10804
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033340.325, "dur": 0.583,
    "args": {
      "External id": 2483102,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10805
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033341.378, "dur": 1.815,
    "args": {
      "External id": 2483103,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10806
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033342.566, "dur": 0.558,
    "args": {
      "External id": 2483104,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10807
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033343.455, "dur": 2.546,
    "args": {
      "External id": 2483105,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10808
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033345.421, "dur": 0.516,
    "args": {
      "External id": 2483106,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10809
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033346.296, "dur": 1.373,
    "args": {
      "External id": 2483107,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10810
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033347.199, "dur": 0.405,
    "args": {
      "External id": 2483108,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10811
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033347.910, "dur": 3.730,
    "args": {
      "External id": 2483109,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10812
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033349.104, "dur": 2.272,
    "args": {
      "External id": 2483110,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10813
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033353.240, "dur": 1.382,
    "args": {
      "External id": 2483111,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10814
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033353.999, "dur": 0.561,
    "args": {
      "External id": 2483112,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10815
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033354.870, "dur": 1.924,
    "args": {
      "External id": 2483113,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10816
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033355.985, "dur": 0.739,
    "args": {
      "External id": 2483114,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10817
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033357.159, "dur": 3.517,
    "args": {
      "External id": 2483115,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10818
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033360.011, "dur": 0.602,
    "args": {
      "External id": 2483116,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10819
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033361.035, "dur": 1.866,
    "args": {
      "External id": 2483117,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10820
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033362.047, "dur": 0.788,
    "args": {
      "External id": 2483118,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10821
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033363.357, "dur": 1.210,
    "args": {
      "External id": 2483119,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10822
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033363.975, "dur": 0.526,
    "args": {
      "External id": 2483120,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10823
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033364.802, "dur": 3.432,
    "args": {
      "External id": 2483121,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10824
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033367.712, "dur": 0.456,
    "args": {
      "External id": 2483122,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10825
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033368.484, "dur": 1.413,
    "args": {
      "External id": 2483123,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10826
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033369.359, "dur": 0.471,
    "args": {
      "External id": 2483124,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10827
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033370.232, "dur": 2.900,
    "args": {
      "External id": 2483125,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10828
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033371.469, "dur": 1.583,
    "args": {
      "External id": 2483126,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10829
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033375.046, "dur": 1.457,
    "args": {
      "External id": 2483127,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10830
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033375.836, "dur": 0.600,
    "args": {
      "External id": 2483128,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10831
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033376.836, "dur": 2.008,
    "args": {
      "External id": 2483129,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10832
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033377.971, "dur": 0.803,
    "args": {
      "External id": 2483130,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10833
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033379.099, "dur": 2.929,
    "args": {
      "External id": 2483131,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10834
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033381.231, "dur": 0.728,
    "args": {
      "External id": 2483132,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10835
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033382.272, "dur": 1.806,
    "args": {
      "External id": 2483133,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10836
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033383.303, "dur": 0.707,
    "args": {
      "External id": 2483134,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10837
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033384.337, "dur": 1.348,
    "args": {
      "External id": 2483135,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10838
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033385.098, "dur": 0.519,
    "args": {
      "External id": 2483136,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10839
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033386.014, "dur": 3.226,
    "args": {
      "External id": 2483137,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10840
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033388.701, "dur": 0.475,
    "args": {
      "External id": 2483138,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10841
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033389.492, "dur": 1.362,
    "args": {
      "External id": 2483139,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10842
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033390.141, "dur": 0.652,
    "args": {
      "External id": 2483140,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10843
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033391.194, "dur": 2.755,
    "args": {
      "External id": 2483141,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10844
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033392.047, "dur": 1.829,
    "args": {
      "External id": 2483142,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10845
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033395.740, "dur": 1.425,
    "args": {
      "External id": 2483143,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10846
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033396.468, "dur": 0.630,
    "args": {
      "External id": 2483144,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10847
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033397.587, "dur": 1.819,
    "args": {
      "External id": 2483145,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10848
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033398.595, "dur": 0.746,
    "args": {
      "External id": 2483146,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10849
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033399.660, "dur": 3.216,
    "args": {
      "External id": 2483147,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10850
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033402.202, "dur": 0.607,
    "args": {
      "External id": 2483148,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10851
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033403.125, "dur": 1.766,
    "args": {
      "External id": 2483149,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10852
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033404.243, "dur": 0.577,
    "args": {
      "External id": 2483150,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10853
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033405.325, "dur": 1.395,
    "args": {
      "External id": 2483151,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10854
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033406.032, "dur": 0.619,
    "args": {
      "External id": 2483152,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10855
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033406.951, "dur": 3.633,
    "args": {
      "External id": 2483153,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10856
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033409.802, "dur": 0.716,
    "args": {
      "External id": 2483154,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10857
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033410.963, "dur": 1.473,
    "args": {
      "External id": 2483155,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10858
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033411.893, "dur": 0.478,
    "args": {
      "External id": 2483156,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10859
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033412.864, "dur": 2.902,
    "args": {
      "External id": 2483157,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10860
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033413.734, "dur": 1.956,
    "args": {
      "External id": 2483158,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10861
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033417.577, "dur": 1.423,
    "args": {
      "External id": 2483159,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10862
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033418.394, "dur": 0.543,
    "args": {
      "External id": 2483160,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10863
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033419.251, "dur": 1.787,
    "args": {
      "External id": 2483161,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10864
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033420.303, "dur": 0.667,
    "args": {
      "External id": 2483162,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10865
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033421.291, "dur": 2.926,
    "args": {
      "External id": 2483163,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10866
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033423.590, "dur": 0.561,
    "args": {
      "External id": 2483164,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10867
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033424.461, "dur": 2.059,
    "args": {
      "External id": 2483165,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10868
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033425.694, "dur": 0.757,
    "args": {
      "External id": 2483166,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10869
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033426.772, "dur": 1.639,
    "args": {
      "External id": 2483167,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10870
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033427.610, "dur": 0.737,
    "args": {
      "External id": 2483168,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10871
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033428.763, "dur": 3.198,
    "args": {
      "External id": 2483169,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10872
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033431.205, "dur": 0.692,
    "args": {
      "External id": 2483170,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10873
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033432.336, "dur": 1.369,
    "args": {
      "External id": 2483171,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10874
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033433.158, "dur": 0.484,
    "args": {
      "External id": 2483172,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10875
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033434.042, "dur": 3.064,
    "args": {
      "External id": 2483173,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10876
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033434.970, "dur": 1.974,
    "args": {
      "External id": 2483174,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10877
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033438.696, "dur": 1.505,
    "args": {
      "External id": 2483175,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10878
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033439.594, "dur": 0.541,
    "args": {
      "External id": 2483176,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10879
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033440.531, "dur": 1.796,
    "args": {
      "External id": 2483177,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10880
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033441.494, "dur": 0.766,
    "args": {
      "External id": 2483178,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10881
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033442.580, "dur": 2.937,
    "args": {
      "External id": 2483179,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10882
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033444.749, "dur": 0.701,
    "args": {
      "External id": 2483180,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10883
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033445.794, "dur": 1.868,
    "args": {
      "External id": 2483181,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10884
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033447.025, "dur": 0.570,
    "args": {
      "External id": 2483182,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10885
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033447.914, "dur": 1.408,
    "args": {
      "External id": 2483183,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10886
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033448.476, "dur": 0.780,
    "args": {
      "External id": 2483184,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10887
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033449.552, "dur": 3.267,
    "args": {
      "External id": 2483185,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10888
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033452.135, "dur": 0.615,
    "args": {
      "External id": 2483186,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10889
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033453.191, "dur": 1.216,
    "args": {
      "External id": 2483187,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10890
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033453.867, "dur": 0.474,
    "args": {
      "External id": 2483188,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10891
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033454.729, "dur": 2.807,
    "args": {
      "External id": 2483189,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10892
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033455.486, "dur": 1.781,
    "args": {
      "External id": 2483190,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10893
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033459.145, "dur": 1.259,
    "args": {
      "External id": 2483191,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10894
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033459.835, "dur": 0.503,
    "args": {
      "External id": 2483192,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10895
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033460.722, "dur": 1.520,
    "args": {
      "External id": 2483193,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10896
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033461.595, "dur": 0.578,
    "args": {
      "External id": 2483194,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10897
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033462.598, "dur": 3.111,
    "args": {
      "External id": 2483195,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10898
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033464.995, "dur": 0.649,
    "args": {
      "External id": 2483196,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10899
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033465.957, "dur": 2.254,
    "args": {
      "External id": 2483197,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10900
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033467.008, "dur": 1.137,
    "args": {
      "External id": 2483198,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10901
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033468.488, "dur": 1.308,
    "args": {
      "External id": 2483199,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10902
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033469.150, "dur": 0.579,
    "args": {
      "External id": 2483200,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10903
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033494.145, "dur": 2.895,
    "args": {
      "External id": 2489345,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10904
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033496.593, "dur": 0.378,
    "args": {
      "External id": 2489346,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10905
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033497.346, "dur": 1.098,
    "args": {
      "External id": 2489347,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10906
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033497.913, "dur": 0.462,
    "args": {
      "External id": 2489348,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10907
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033498.698, "dur": 2.758,
    "args": {
      "External id": 2489349,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10908
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033499.732, "dur": 1.655,
    "args": {
      "External id": 2489350,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10909
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033503.440, "dur": 1.159,
    "args": {
      "External id": 2489351,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10910
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033504.064, "dur": 0.464,
    "args": {
      "External id": 2489352,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10911
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033504.851, "dur": 1.550,
    "args": {
      "External id": 2489353,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10912
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033505.624, "dur": 0.711,
    "args": {
      "External id": 2489354,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10913
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033506.686, "dur": 2.991,
    "args": {
      "External id": 2489355,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10914
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033509.145, "dur": 0.466,
    "args": {
      "External id": 2489356,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10915
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033509.940, "dur": 1.435,
    "args": {
      "External id": 2489357,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10916
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033510.784, "dur": 0.522,
    "args": {
      "External id": 2489358,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10917
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033511.637, "dur": 1.240,
    "args": {
      "External id": 2489359,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10918
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033512.293, "dur": 0.514,
    "args": {
      "External id": 2489360,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10919
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033513.108, "dur": 2.987,
    "args": {
      "External id": 2489361,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10920
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033515.551, "dur": 0.480,
    "args": {
      "External id": 2489362,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10921
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033516.352, "dur": 1.287,
    "args": {
      "External id": 2489363,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10922
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033516.929, "dur": 0.644,
    "args": {
      "External id": 2489364,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10923
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033517.872, "dur": 3.210,
    "args": {
      "External id": 2489365,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10924
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033518.801, "dur": 2.210,
    "args": {
      "External id": 2489366,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10925
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033522.832, "dur": 1.073,
    "args": {
      "External id": 2489367,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10926
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033523.428, "dur": 0.410,
    "args": {
      "External id": 2489368,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10927
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033524.137, "dur": 1.434,
    "args": {
      "External id": 2489369,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10928
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033525.108, "dur": 0.394,
    "args": {
      "External id": 2489370,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10929
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033526.019, "dur": 14.416,
    "args": {
      "External id": 2489371,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10930
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033538.902, "dur": 0.910,
    "args": {
      "External id": 2489372,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10931
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033541.418, "dur": 2.151,
    "args": {
      "External id": 2489373,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10932
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033542.649, "dur": 0.855,
    "args": {
      "External id": 2489374,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10933
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033543.875, "dur": 1.566,
    "args": {
      "External id": 2489375,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10934
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033544.490, "dur": 0.877,
    "args": {
      "External id": 2489376,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10935
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033545.701, "dur": 4.025,
    "args": {
      "External id": 2489377,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10936
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033549.086, "dur": 0.577,
    "args": {
      "External id": 2489378,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10937
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033550.179, "dur": 1.461,
    "args": {
      "External id": 2489379,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10938
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033550.962, "dur": 0.614,
    "args": {
      "External id": 2489380,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10939
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033552.126, "dur": 3.120,
    "args": {
      "External id": 2489381,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10940
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033553.212, "dur": 1.960,
    "args": {
      "External id": 2489382,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10941
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033557.135, "dur": 1.365,
    "args": {
      "External id": 2489383,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10942
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033557.809, "dur": 0.624,
    "args": {
      "External id": 2489384,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10943
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033558.949, "dur": 1.787,
    "args": {
      "External id": 2489385,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10944
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033560.089, "dur": 0.581,
    "args": {
      "External id": 2489386,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10945
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033561.039, "dur": 2.951,
    "args": {
      "External id": 2489387,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10946
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033563.295, "dur": 0.631,
    "args": {
      "External id": 2489388,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10947
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033564.246, "dur": 1.377,
    "args": {
      "External id": 2489389,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10948
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033565.121, "dur": 0.434,
    "args": {
      "External id": 2489390,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10949
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033565.961, "dur": 1.114,
    "args": {
      "External id": 2489391,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10950
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033566.622, "dur": 0.382,
    "args": {
      "External id": 2489392,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10951
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033567.313, "dur": 3.258,
    "args": {
      "External id": 2489393,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10952
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033569.962, "dur": 0.547,
    "args": {
      "External id": 2489394,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10953
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033570.968, "dur": 1.281,
    "args": {
      "External id": 2489395,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10954
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033571.701, "dur": 0.484,
    "args": {
      "External id": 2489396,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10955
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033572.505, "dur": 2.584,
    "args": {
      "External id": 2489397,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10956
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033573.504, "dur": 1.508,
    "args": {
      "External id": 2489398,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10957
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033577.209, "dur": 2.195,
    "args": {
      "External id": 2489399,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10958
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033578.843, "dur": 0.494,
    "args": {
      "External id": 2489400,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10959
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033579.639, "dur": 1.687,
    "args": {
      "External id": 2489401,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10960
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033580.667, "dur": 0.589,
    "args": {
      "External id": 2489402,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10961
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033581.589, "dur": 2.668,
    "args": {
      "External id": 2489403,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10962
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033583.662, "dur": 0.531,
    "args": {
      "External id": 2489404,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10963
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033584.581, "dur": 1.872,
    "args": {
      "External id": 2489405,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10964
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033585.618, "dur": 0.775,
    "args": {
      "External id": 2489406,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10965
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033586.716, "dur": 1.204,
    "args": {
      "External id": 2489407,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10966
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033587.351, "dur": 0.504,
    "args": {
      "External id": 2489408,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10967
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033588.172, "dur": 2.920,
    "args": {
      "External id": 2489409,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10968
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033590.607, "dur": 0.423,
    "args": {
      "External id": 2489410,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10969
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033591.349, "dur": 1.384,
    "args": {
      "External id": 2489411,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10970
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033592.162, "dur": 0.508,
    "args": {
      "External id": 2489412,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10971
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033592.975, "dur": 2.894,
    "args": {
      "External id": 2489413,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10972
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033593.944, "dur": 1.851,
    "args": {
      "External id": 2489414,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10973
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033597.560, "dur": 1.142,
    "args": {
      "External id": 2489415,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10974
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033598.168, "dur": 0.474,
    "args": {
      "External id": 2489416,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10975
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033599.054, "dur": 1.675,
    "args": {
      "External id": 2489417,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10976
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033600.130, "dur": 0.531,
    "args": {
      "External id": 2489418,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10977
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033600.985, "dur": 2.394,
    "args": {
      "External id": 2489419,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10978
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033602.907, "dur": 0.408,
    "args": {
      "External id": 2489420,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10979
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033603.633, "dur": 1.528,
    "args": {
      "External id": 2489421,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10980
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033604.610, "dur": 0.489,
    "args": {
      "External id": 2489422,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10981
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033605.416, "dur": 1.370,
    "args": {
      "External id": 2489423,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10982
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033606.188, "dur": 0.531,
    "args": {
      "External id": 2489424,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10983
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033607.024, "dur": 3.072,
    "args": {
      "External id": 2489425,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10984
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033609.488, "dur": 0.537,
    "args": {
      "External id": 2489426,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10985
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033610.394, "dur": 1.230,
    "args": {
      "External id": 2489427,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10986
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033611.077, "dur": 0.474,
    "args": {
      "External id": 2489428,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10987
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033611.882, "dur": 3.008,
    "args": {
      "External id": 2489429,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10988
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033612.854, "dur": 1.965,
    "args": {
      "External id": 2489430,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10989
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033616.428, "dur": 1.424,
    "args": {
      "External id": 2489431,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10990
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033617.186, "dur": 0.596,
    "args": {
      "External id": 2489432,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10991
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033618.084, "dur": 1.557,
    "args": {
      "External id": 2489433,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10992
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033618.886, "dur": 0.688,
    "args": {
      "External id": 2489434,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10993
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033620.005, "dur": 2.537,
    "args": {
      "External id": 2489435,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10994
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033621.964, "dur": 0.512,
    "args": {
      "External id": 2489436,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10995
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033622.788, "dur": 1.569,
    "args": {
      "External id": 2489437,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10996
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033623.741, "dur": 0.547,
    "args": {
      "External id": 2489438,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10997
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033624.614, "dur": 1.266,
    "args": {
      "External id": 2489439,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 10998
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033625.268, "dur": 0.543,
    "args": {
      "External id": 2489440,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 10999
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033626.250, "dur": 2.956,
    "args": {
      "External id": 2489441,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11000
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033628.657, "dur": 0.485,
    "args": {
      "External id": 2489442,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11001
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033629.479, "dur": 1.044,
    "args": {
      "External id": 2489443,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11002
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033630.045, "dur": 0.414,
    "args": {
      "External id": 2489444,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11003
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033630.753, "dur": 2.660,
    "args": {
      "External id": 2489445,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11004
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033631.614, "dur": 1.727,
    "args": {
      "External id": 2489446,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11005
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033635.153, "dur": 1.095,
    "args": {
      "External id": 2489447,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11006
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033635.752, "dur": 0.429,
    "args": {
      "External id": 2489448,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11007
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033636.481, "dur": 1.502,
    "args": {
      "External id": 2489449,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11008
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033637.383, "dur": 0.531,
    "args": {
      "External id": 2489450,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11009
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033638.347, "dur": 2.567,
    "args": {
      "External id": 2489451,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11010
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033640.438, "dur": 0.410,
    "args": {
      "External id": 2489452,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11011
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033641.193, "dur": 1.332,
    "args": {
      "External id": 2489453,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11012
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033642.052, "dur": 0.405,
    "args": {
      "External id": 2489454,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11013
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033642.779, "dur": 1.373,
    "args": {
      "External id": 2489455,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11014
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033643.383, "dur": 0.697,
    "args": {
      "External id": 2489456,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11015
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033644.395, "dur": 3.186,
    "args": {
      "External id": 2489457,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11016
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033646.911, "dur": 0.601,
    "args": {
      "External id": 2489458,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11017
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033647.836, "dur": 1.264,
    "args": {
      "External id": 2489459,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11018
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033648.571, "dur": 0.461,
    "args": {
      "External id": 2489460,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11019
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033649.334, "dur": 2.304,
    "args": {
      "External id": 2489461,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11020
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033650.123, "dur": 1.444,
    "args": {
      "External id": 2489462,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11021
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033653.332, "dur": 1.242,
    "args": {
      "External id": 2489463,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11022
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033653.944, "dur": 0.565,
    "args": {
      "External id": 2489464,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11023
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033654.805, "dur": 1.753,
    "args": {
      "External id": 2489465,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11024
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033655.840, "dur": 0.649,
    "args": {
      "External id": 2489466,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11025
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033656.810, "dur": 2.501,
    "args": {
      "External id": 2489467,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11026
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033658.768, "dur": 0.476,
    "args": {
      "External id": 2489468,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11027
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033659.582, "dur": 1.452,
    "args": {
      "External id": 2489469,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11028
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033660.541, "dur": 0.432,
    "args": {
      "External id": 2489470,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11029
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033661.290, "dur": 1.334,
    "args": {
      "External id": 2489471,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11030
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033662.033, "dur": 0.524,
    "args": {
      "External id": 2489472,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11031
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033662.856, "dur": 2.666,
    "args": {
      "External id": 2489473,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11032
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033664.937, "dur": 0.521,
    "args": {
      "External id": 2489474,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11033
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033665.778, "dur": 2.628,
    "args": {
      "External id": 2489475,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11034
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033667.925, "dur": 0.409,
    "args": {
      "External id": 2489476,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11035
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033668.665, "dur": 2.756,
    "args": {
      "External id": 2489477,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11036
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033669.594, "dur": 1.755,
    "args": {
      "External id": 2489478,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11037
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033672.935, "dur": 1.080,
    "args": {
      "External id": 2489479,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11038
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033673.532, "dur": 0.415,
    "args": {
      "External id": 2489480,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11039
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033674.242, "dur": 1.732,
    "args": {
      "External id": 2489481,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11040
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033675.341, "dur": 0.557,
    "args": {
      "External id": 2489482,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11041
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033676.227, "dur": 3.410,
    "args": {
      "External id": 2489483,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11042
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033678.931, "dur": 0.637,
    "args": {
      "External id": 2489484,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11043
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033679.882, "dur": 1.767,
    "args": {
      "External id": 2489485,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11044
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033680.840, "dur": 0.744,
    "args": {
      "External id": 2489486,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11045
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033681.903, "dur": 1.214,
    "args": {
      "External id": 2489487,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11046
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033682.471, "dur": 0.577,
    "args": {
      "External id": 2489488,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11047
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033683.346, "dur": 3.005,
    "args": {
      "External id": 2489489,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11048
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033685.857, "dur": 0.427,
    "args": {
      "External id": 2489490,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11049
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033686.741, "dur": 1.189,
    "args": {
      "External id": 2489491,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11050
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033687.401, "dur": 0.465,
    "args": {
      "External id": 2489492,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11051
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033688.157, "dur": 3.227,
    "args": {
      "External id": 2489493,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11052
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033689.392, "dur": 1.739,
    "args": {
      "External id": 2489494,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11053
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033693.444, "dur": 1.218,
    "args": {
      "External id": 2489495,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11054
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033694.144, "dur": 0.458,
    "args": {
      "External id": 2489496,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11055
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033695.021, "dur": 1.589,
    "args": {
      "External id": 2489497,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11056
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033696.017, "dur": 0.524,
    "args": {
      "External id": 2489498,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11057
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033697.055, "dur": 2.988,
    "args": {
      "External id": 2489499,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11058
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033699.485, "dur": 0.489,
    "args": {
      "External id": 2489500,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11059
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033700.303, "dur": 1.474,
    "args": {
      "External id": 2489501,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11060
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033701.163, "dur": 0.541,
    "args": {
      "External id": 2489502,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11061
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033702.204, "dur": 1.191,
    "args": {
      "External id": 2489503,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11062
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033702.791, "dur": 0.534,
    "args": {
      "External id": 2489504,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11063
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033703.768, "dur": 3.208,
    "args": {
      "External id": 2489505,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11064
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033706.352, "dur": 0.558,
    "args": {
      "External id": 2489506,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11065
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033707.236, "dur": 1.252,
    "args": {
      "External id": 2489507,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11066
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033707.872, "dur": 0.552,
    "args": {
      "External id": 2489508,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11067
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033708.717, "dur": 3.009,
    "args": {
      "External id": 2489509,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11068
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033709.836, "dur": 1.820,
    "args": {
      "External id": 2489510,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11069
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033713.829, "dur": 1.818,
    "args": {
      "External id": 2489511,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11070
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033714.865, "dur": 0.723,
    "args": {
      "External id": 2489512,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11071
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033715.883, "dur": 1.483,
    "args": {
      "External id": 2489513,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11072
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033716.764, "dur": 0.536,
    "args": {
      "External id": 2489514,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11073
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033717.620, "dur": 2.648,
    "args": {
      "External id": 2489515,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11074
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033719.806, "dur": 0.396,
    "args": {
      "External id": 2489516,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11075
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033720.518, "dur": 1.443,
    "args": {
      "External id": 2489517,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11076
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033721.440, "dur": 0.454,
    "args": {
      "External id": 2489518,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11077
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033722.392, "dur": 1.184,
    "args": {
      "External id": 2489519,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11078
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033722.958, "dur": 0.554,
    "args": {
      "External id": 2489520,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11079
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033723.808, "dur": 2.881,
    "args": {
      "External id": 2489521,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11080
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033726.215, "dur": 0.408,
    "args": {
      "External id": 2489522,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11081
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033726.958, "dur": 1.258,
    "args": {
      "External id": 2489523,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11082
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033727.675, "dur": 0.471,
    "args": {
      "External id": 2489524,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11083
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033728.445, "dur": 2.790,
    "args": {
      "External id": 2489525,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11084
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033729.470, "dur": 1.692,
    "args": {
      "External id": 2489526,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11085
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033732.873, "dur": 1.220,
    "args": {
      "External id": 2489527,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11086
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033733.541, "dur": 0.489,
    "args": {
      "External id": 2489528,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11087
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033734.330, "dur": 1.629,
    "args": {
      "External id": 2489529,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11088
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033735.326, "dur": 0.565,
    "args": {
      "External id": 2489530,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11089
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033736.213, "dur": 2.400,
    "args": {
      "External id": 2489531,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11090
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033738.049, "dur": 0.486,
    "args": {
      "External id": 2489532,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11091
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033738.895, "dur": 1.509,
    "args": {
      "External id": 2489533,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11092
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033739.846, "dur": 0.492,
    "args": {
      "External id": 2489534,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11093
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033740.664, "dur": 1.417,
    "args": {
      "External id": 2489535,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11094
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033741.388, "dur": 0.621,
    "args": {
      "External id": 2489536,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11095
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033742.330, "dur": 3.164,
    "args": {
      "External id": 2489537,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11096
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033744.941, "dur": 0.487,
    "args": {
      "External id": 2489538,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11097
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033745.754, "dur": 1.203,
    "args": {
      "External id": 2489539,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11098
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033746.428, "dur": 0.461,
    "args": {
      "External id": 2489540,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11099
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033747.212, "dur": 3.069,
    "args": {
      "External id": 2489541,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11100
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033748.155, "dur": 2.049,
    "args": {
      "External id": 2489542,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11101
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033752.175, "dur": 1.077,
    "args": {
      "External id": 2489543,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11102
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033752.751, "dur": 0.433,
    "args": {
      "External id": 2489544,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11103
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033753.482, "dur": 1.941,
    "args": {
      "External id": 2489545,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11104
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033754.654, "dur": 0.703,
    "args": {
      "External id": 2489546,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11105
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033755.677, "dur": 2.828,
    "args": {
      "External id": 2489547,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11106
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033758.025, "dur": 0.410,
    "args": {
      "External id": 2489548,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11107
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033758.786, "dur": 1.528,
    "args": {
      "External id": 2489549,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11108
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033759.740, "dur": 0.505,
    "args": {
      "External id": 2489550,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11109
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033760.833, "dur": 1.525,
    "args": {
      "External id": 2489551,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11110
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033761.603, "dur": 0.685,
    "args": {
      "External id": 2489552,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11111
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033762.700, "dur": 2.896,
    "args": {
      "External id": 2489553,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11112
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033764.933, "dur": 0.600,
    "args": {
      "External id": 2489554,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11113
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033765.852, "dur": 1.193,
    "args": {
      "External id": 2489555,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11114
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033766.573, "dur": 0.409,
    "args": {
      "External id": 2489556,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11115
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033767.359, "dur": 2.919,
    "args": {
      "External id": 2489557,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11116
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033768.346, "dur": 1.851,
    "args": {
      "External id": 2489558,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11117
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033771.877, "dur": 1.142,
    "args": {
      "External id": 2489559,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11118
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033772.525, "dur": 0.418,
    "args": {
      "External id": 2489560,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11119
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033773.273, "dur": 1.545,
    "args": {
      "External id": 2489561,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11120
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033774.187, "dur": 0.563,
    "args": {
      "External id": 2489562,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11121
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033775.072, "dur": 2.516,
    "args": {
      "External id": 2489563,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11122
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033776.959, "dur": 0.561,
    "args": {
      "External id": 2489564,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11123
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033777.858, "dur": 1.523,
    "args": {
      "External id": 2489565,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11124
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033778.835, "dur": 0.478,
    "args": {
      "External id": 2489566,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11125
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033779.689, "dur": 1.357,
    "args": {
      "External id": 2489567,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11126
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033780.283, "dur": 0.696,
    "args": {
      "External id": 2489568,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11127
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033781.276, "dur": 3.032,
    "args": {
      "External id": 2489569,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11128
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033783.672, "dur": 0.569,
    "args": {
      "External id": 2489570,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11129
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033784.565, "dur": 1.408,
    "args": {
      "External id": 2489571,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11130
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033785.450, "dur": 0.459,
    "args": {
      "External id": 2489572,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11131
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033786.216, "dur": 2.504,
    "args": {
      "External id": 2489573,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11132
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033787.126, "dur": 1.524,
    "args": {
      "External id": 2489574,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11133
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033790.774, "dur": 1.056,
    "args": {
      "External id": 2489575,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11134
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033791.339, "dur": 0.424,
    "args": {
      "External id": 2489576,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11135
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033792.063, "dur": 1.772,
    "args": {
      "External id": 2489577,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11136
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033793.155, "dur": 0.611,
    "args": {
      "External id": 2489578,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11137
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033794.091, "dur": 2.696,
    "args": {
      "External id": 2489579,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11138
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033796.329, "dur": 0.394,
    "args": {
      "External id": 2489580,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11139
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033797.033, "dur": 1.417,
    "args": {
      "External id": 2489581,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11140
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033797.900, "dur": 0.479,
    "args": {
      "External id": 2489582,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11141
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033798.703, "dur": 1.249,
    "args": {
      "External id": 2489583,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11142
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033799.392, "dur": 0.493,
    "args": {
      "External id": 2489584,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11143
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033800.367, "dur": 3.096,
    "args": {
      "External id": 2489585,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11144
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033802.712, "dur": 0.689,
    "args": {
      "External id": 2489586,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11145
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033803.717, "dur": 1.361,
    "args": {
      "External id": 2489587,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11146
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033804.491, "dur": 0.516,
    "args": {
      "External id": 2489588,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11147
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033805.305, "dur": 3.145,
    "args": {
      "External id": 2489589,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11148
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033806.177, "dur": 2.203,
    "args": {
      "External id": 2489590,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11149
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033810.055, "dur": 1.308,
    "args": {
      "External id": 2489591,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11150
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033810.674, "dur": 0.624,
    "args": {
      "External id": 2489592,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11151
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033811.593, "dur": 1.983,
    "args": {
      "External id": 2489593,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11152
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033812.633, "dur": 0.877,
    "args": {
      "External id": 2489594,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11153
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033813.853, "dur": 3.150,
    "args": {
      "External id": 2489595,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11154
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033816.588, "dur": 0.348,
    "args": {
      "External id": 2489596,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11155
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033817.248, "dur": 1.319,
    "args": {
      "External id": 2489597,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11156
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033818.083, "dur": 0.421,
    "args": {
      "External id": 2489598,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11157
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033819.264, "dur": 1.342,
    "args": {
      "External id": 2489599,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11158
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033820.063, "dur": 0.470,
    "args": {
      "External id": 2489600,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11159
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033820.841, "dur": 2.752,
    "args": {
      "External id": 2489601,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11160
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033822.928, "dur": 0.602,
    "args": {
      "External id": 2489602,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11161
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033824.063, "dur": 1.367,
    "args": {
      "External id": 2489603,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11162
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033824.891, "dur": 0.470,
    "args": {
      "External id": 2489604,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11163
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033825.661, "dur": 2.809,
    "args": {
      "External id": 2489605,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11164
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033826.649, "dur": 1.747,
    "args": {
      "External id": 2489606,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11165
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033830.220, "dur": 1.300,
    "args": {
      "External id": 2489607,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11166
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033830.956, "dur": 0.501,
    "args": {
      "External id": 2489608,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11167
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033831.918, "dur": 1.588,
    "args": {
      "External id": 2489609,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11168
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033832.713, "dur": 0.724,
    "args": {
      "External id": 2489610,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11169
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033833.763, "dur": 2.620,
    "args": {
      "External id": 2489611,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11170
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033835.750, "dur": 0.570,
    "args": {
      "External id": 2489612,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11171
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033836.628, "dur": 1.383,
    "args": {
      "External id": 2489613,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11172
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033837.394, "dur": 0.557,
    "args": {
      "External id": 2489614,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11173
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033838.279, "dur": 1.422,
    "args": {
      "External id": 2489615,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11174
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033838.898, "dur": 0.734,
    "args": {
      "External id": 2489616,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11175
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033840.012, "dur": 3.201,
    "args": {
      "External id": 2489617,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11176
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033842.633, "dur": 0.514,
    "args": {
      "External id": 2489618,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11177
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033843.469, "dur": 1.138,
    "args": {
      "External id": 2489619,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11178
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033844.124, "dur": 0.413,
    "args": {
      "External id": 2489620,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11179
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033844.836, "dur": 2.490,
    "args": {
      "External id": 2489621,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11180
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033845.877, "dur": 1.379,
    "args": {
      "External id": 2489622,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11181
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033848.833, "dur": 1.217,
    "args": {
      "External id": 2489623,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11182
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033849.494, "dur": 0.481,
    "args": {
      "External id": 2489624,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11183
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033850.281, "dur": 1.813,
    "args": {
      "External id": 2489625,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11184
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033851.357, "dur": 0.666,
    "args": {
      "External id": 2489626,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11185
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033852.358, "dur": 2.849,
    "args": {
      "External id": 2489627,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11186
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033854.278, "dur": 0.856,
    "args": {
      "External id": 2489628,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11187
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033855.464, "dur": 1.492,
    "args": {
      "External id": 2489629,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11188
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033856.480, "dur": 0.411,
    "args": {
      "External id": 2489630,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11189
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033857.211, "dur": 1.241,
    "args": {
      "External id": 2489631,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11190
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033857.850, "dur": 0.534,
    "args": {
      "External id": 2489632,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11191
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033858.717, "dur": 2.645,
    "args": {
      "External id": 2489633,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11192
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033860.733, "dur": 0.551,
    "args": {
      "External id": 2489634,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11193
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033861.769, "dur": 1.275,
    "args": {
      "External id": 2489635,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11194
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033862.548, "dur": 0.427,
    "args": {
      "External id": 2489636,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11195
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033863.558, "dur": 3.052,
    "args": {
      "External id": 2489637,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11196
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033864.575, "dur": 1.774,
    "args": {
      "External id": 2489638,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11197
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033868.395, "dur": 1.125,
    "args": {
      "External id": 2489639,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11198
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033869.030, "dur": 0.423,
    "args": {
      "External id": 2489640,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11199
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033869.750, "dur": 1.554,
    "args": {
      "External id": 2489641,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11200
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033870.606, "dur": 0.629,
    "args": {
      "External id": 2489642,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11201
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033871.655, "dur": 2.473,
    "args": {
      "External id": 2489643,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11202
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033873.674, "dur": 0.381,
    "args": {
      "External id": 2489644,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11203
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033874.401, "dur": 1.535,
    "args": {
      "External id": 2489645,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11204
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033875.448, "dur": 0.420,
    "args": {
      "External id": 2489646,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11205
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033876.284, "dur": 1.275,
    "args": {
      "External id": 2489647,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11206
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033876.995, "dur": 0.497,
    "args": {
      "External id": 2489648,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11207
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033877.794, "dur": 2.894,
    "args": {
      "External id": 2489649,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11208
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033880.182, "dur": 0.436,
    "args": {
      "External id": 2489650,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11209
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033880.942, "dur": 1.186,
    "args": {
      "External id": 2489651,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11210
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033881.591, "dur": 0.462,
    "args": {
      "External id": 2489652,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11211
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033882.446, "dur": 2.755,
    "args": {
      "External id": 2489653,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11212
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033883.494, "dur": 1.637,
    "args": {
      "External id": 2489654,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11213
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033886.801, "dur": 1.306,
    "args": {
      "External id": 2489655,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11214
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033887.518, "dur": 0.528,
    "args": {
      "External id": 2489656,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11215
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033888.337, "dur": 1.696,
    "args": {
      "External id": 2489657,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11216
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033889.445, "dur": 0.517,
    "args": {
      "External id": 2489658,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11217
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033890.287, "dur": 2.547,
    "args": {
      "External id": 2489659,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11218
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033892.324, "dur": 0.438,
    "args": {
      "External id": 2489660,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11219
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033893.155, "dur": 1.492,
    "args": {
      "External id": 2489661,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11220
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033894.180, "dur": 0.388,
    "args": {
      "External id": 2489662,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11221
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033894.905, "dur": 1.280,
    "args": {
      "External id": 2489663,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11222
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033895.515, "dur": 0.600,
    "args": {
      "External id": 2489664,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11223
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033896.414, "dur": 2.551,
    "args": {
      "External id": 2489665,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11224
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033898.494, "dur": 0.409,
    "args": {
      "External id": 2489666,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11225
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033899.251, "dur": 1.198,
    "args": {
      "External id": 2489667,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11226
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033899.927, "dur": 0.455,
    "args": {
      "External id": 2489668,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11227
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033900.679, "dur": 2.668,
    "args": {
      "External id": 2489669,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11228
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033901.741, "dur": 1.439,
    "args": {
      "External id": 2489670,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11229
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033904.995, "dur": 1.332,
    "args": {
      "External id": 2489671,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11230
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033905.616, "dur": 0.641,
    "args": {
      "External id": 2489672,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11231
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033906.555, "dur": 1.733,
    "args": {
      "External id": 2489673,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11232
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033907.526, "dur": 0.695,
    "args": {
      "External id": 2489674,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11233
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033908.567, "dur": 2.736,
    "args": {
      "External id": 2489675,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11234
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033910.720, "dur": 0.514,
    "args": {
      "External id": 2489676,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11235
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033911.548, "dur": 1.709,
    "args": {
      "External id": 2489677,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11236
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033912.664, "dur": 0.527,
    "args": {
      "External id": 2489678,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11237
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033913.529, "dur": 1.426,
    "args": {
      "External id": 2489679,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11238
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033914.258, "dur": 0.629,
    "args": {
      "External id": 2489680,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11239
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033915.186, "dur": 3.638,
    "args": {
      "External id": 2489681,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11240
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033917.718, "dur": 1.042,
    "args": {
      "External id": 2489682,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11241
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033919.078, "dur": 1.262,
    "args": {
      "External id": 2489683,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11242
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033919.836, "dur": 0.436,
    "args": {
      "External id": 2489684,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11243
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033920.574, "dur": 2.201,
    "args": {
      "External id": 2489685,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11244
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033921.517, "dur": 0.975,
    "args": {
      "External id": 2489686,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11245
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033924.658, "dur": 1.607,
    "args": {
      "External id": 2489687,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11246
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033925.454, "dur": 0.743,
    "args": {
      "External id": 2489688,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11247
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033926.735, "dur": 1.517,
    "args": {
      "External id": 2489689,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11248
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033927.539, "dur": 0.643,
    "args": {
      "External id": 2489690,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11249
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033928.524, "dur": 2.840,
    "args": {
      "External id": 2489691,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11250
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033930.748, "dur": 0.550,
    "args": {
      "External id": 2489692,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11251
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033931.652, "dur": 1.872,
    "args": {
      "External id": 2489693,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11252
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033932.610, "dur": 0.842,
    "args": {
      "External id": 2489694,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11253
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033933.797, "dur": 1.305,
    "args": {
      "External id": 2489695,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11254
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033934.475, "dur": 0.558,
    "args": {
      "External id": 2489696,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11255
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033935.331, "dur": 3.002,
    "args": {
      "External id": 2489697,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11256
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033937.792, "dur": 0.472,
    "args": {
      "External id": 2489698,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11257
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033938.587, "dur": 1.450,
    "args": {
      "External id": 2489699,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11258
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033939.462, "dur": 0.505,
    "args": {
      "External id": 2489700,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11259
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033940.266, "dur": 2.849,
    "args": {
      "External id": 2489701,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11260
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033941.452, "dur": 1.593,
    "args": {
      "External id": 2489702,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11261
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033944.978, "dur": 1.498,
    "args": {
      "External id": 2489703,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11262
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033945.533, "dur": 0.671,
    "args": {
      "External id": 2489704,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11263
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033946.866, "dur": 1.684,
    "args": {
      "External id": 2489705,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11264
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033947.763, "dur": 0.717,
    "args": {
      "External id": 2489706,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11265
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033948.808, "dur": 2.332,
    "args": {
      "External id": 2489707,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11266
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033950.639, "dur": 0.439,
    "args": {
      "External id": 2489708,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11267
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033951.389, "dur": 1.532,
    "args": {
      "External id": 2489709,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11268
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033952.322, "dur": 0.531,
    "args": {
      "External id": 2489710,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11269
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033953.182, "dur": 1.462,
    "args": {
      "External id": 2489711,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11270
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033953.895, "dur": 0.681,
    "args": {
      "External id": 2489712,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11271
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033954.875, "dur": 3.269,
    "args": {
      "External id": 2489713,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11272
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033957.489, "dur": 0.593,
    "args": {
      "External id": 2489714,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11273
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033958.471, "dur": 1.296,
    "args": {
      "External id": 2489715,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11274
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033959.263, "dur": 0.440,
    "args": {
      "External id": 2489716,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11275
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033959.992, "dur": 2.860,
    "args": {
      "External id": 2489717,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11276
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033960.973, "dur": 1.807,
    "args": {
      "External id": 2489718,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11277
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033964.441, "dur": 1.312,
    "args": {
      "External id": 2489719,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11278
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033965.124, "dur": 0.558,
    "args": {
      "External id": 2489720,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11279
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033966.190, "dur": 1.721,
    "args": {
      "External id": 2489721,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11280
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033967.274, "dur": 0.569,
    "args": {
      "External id": 2489722,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11281
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033968.190, "dur": 2.607,
    "args": {
      "External id": 2489723,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11282
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033970.322, "dur": 0.408,
    "args": {
      "External id": 2489724,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11283
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033971.062, "dur": 1.396,
    "args": {
      "External id": 2489725,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11284
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033971.843, "dur": 0.549,
    "args": {
      "External id": 2489726,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11285
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033972.735, "dur": 1.673,
    "args": {
      "External id": 2489727,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11286
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033973.405, "dur": 0.937,
    "args": {
      "External id": 2489728,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11287
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033974.639, "dur": 3.047,
    "args": {
      "External id": 2489729,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11288
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033977.198, "dur": 0.418,
    "args": {
      "External id": 2489730,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11289
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::unsqueeze", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033977.946, "dur": 2.190,
    "args": {
      "External id": 2489731,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11290
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658033979.730, "dur": 0.335,
    "args": {
      "External id": 2489732,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[1]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11291
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::cat", "pid": 1336759, "tid": 1336759,
    "ts": 1514658034025.281, "dur": 130.745,
    "args": {
      "External id": 2489733,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["TensorList", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11292
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linalg_vector_norm", "pid": 1336759, "tid": 1336759,
    "ts": 1514658034251.710, "dur": 127.665,
    "args": {
      "External id": 2489734,"Record function id": 0, "Concrete Inputs": ["", "2.", "", "False", ""], "Input type": ["float", "Scalar", "", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "Input Dims": [[291], [], [], [], []], "Ev Idx": 11293
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::linalg_vector_norm", "pid": 1336759, "tid": 1336759,
    "ts": 1514658034313.132, "dur": 47.560,
    "args": {
      "External id": 2489735,"Record function id": 0, "Concrete Inputs": ["", "2.", "", "False", ""], "Input type": ["float", "Scalar", "", "Scalar", ""], "Input Strides": [[1], [], [], [], []], "Input Dims": [[291], [], [], [], []], "Ev Idx": 11294
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::as_strided", "pid": 1336759, "tid": 1336759,
    "ts": 1514658034329.712, "dur": 1.116,
    "args": {
      "External id": 2489736,"Record function id": 0, "Concrete Inputs": ["", "[1]", "[0]", ""], "Input type": ["float", "ScalarList", "ScalarList", ""], "Input Strides": [[], [], [], []], "Input Dims": [[], [], [], []], "Ev Idx": 11295
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "Redistribute", "pid": 1336759, "tid": 1336759,
    "ts": 1514658034719.201, "dur": 1001.797,
    "args": {
      "External id": 2489737,"Sequence number": 24740656, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", "False"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11296
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 1336759, "tid": 1336759,
    "ts": 1514658034773.098, "dur": 103.933,
    "args": {
      "External id": 2489738,"Record function id": 0, "Concrete Inputs": ["", "2."], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11297
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658034780.018, "dur": 1.687,
    "args": {
      "External id": 2489739,"Record function id": 0, "Concrete Inputs": ["", "2."], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11298
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658034830.957, "dur": 0.942,
    "args": {
      "External id": 2489740,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 11299
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "_c10d_functional::all_reduce", "pid": 1336759, "tid": 1336759,
    "ts": 1514658034904.139, "dur": 475.719,
    "args": {
      "External id": 2489741,"Record function id": 0, "Concrete Inputs": ["", "", ""], "Input type": ["float", "", ""], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 11300
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clone", "pid": 1336759, "tid": 1336759,
    "ts": 1514658034910.268, "dur": 41.441,
    "args": {
      "External id": 2489742,"Record function id": 0, "Concrete Inputs": ["", "0"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11301
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty_like", "pid": 1336759, "tid": 1336759,
    "ts": 1514658034913.322, "dur": 8.580,
    "args": {
      "External id": 2489743,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "0"], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11302
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514658034917.200, "dur": 3.886,
    "args": {
      "External id": 2489744,"Record function id": 0, "Concrete Inputs": ["[]", "6", "0", "", "", "0"], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", "Scalar"], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11303
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::copy_", "pid": 1336759, "tid": 1336759,
    "ts": 1514658034923.142, "dur": 28.115,
    "args": {
      "External id": 2489745,"Record function id": 0, "Concrete Inputs": ["", "", "False"], "Input type": ["float", "float", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 11304
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "c10d::allreduce_", "pid": 1336759, "tid": 1336759,
    "ts": 1514658034961.653, "dur": 413.077,
    "args": {
      "External id": 2489746,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "-1"], "Input type": ["TensorList", "", "", "", "Scalar"], "Input Strides": [[[]], [], [], [], []], "Input Dims": [[[]], [], [], [], []], "Ev Idx": 11305
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514658035025.420, "dur": 342.337,
    "args": {
      "External id": 2489747,"Record function id": 0, "Collective name": "allreduce", "Process Group Description": "default_pg", "dtype": "Float", "Rank": 6, "Input Strides": [[[]], [], [], [], [], [], [], [], [], []], "Concrete Inputs": ["", "", "", "6", "", "[]", "[]", "0", "1", "8"], "Out msg nelems": 1, "Process Group Name": "0", "Input type": ["TensorList", "", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Global rank stride": 1, "Out split size": "[]", "In split size": "[]", "Process Group Ranks": "[0, 1, 2, 3, 4, 5, 6, 7]", "Group size": 8, "Input Dims": [[[]], [], [], [], [], [], [], [], [], []], "Global rank start": 0, "Ev Idx": 11306, "In msg nelems": 1
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "nccl:all_reduce", "pid": 1336759, "tid": 1336759,
    "ts": 1514658035042.199, "dur": 320.002,
    "args": {
      "External id": 2489748,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 11307
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 1336759, "tid": 1336759,
    "ts": 1514658035445.887, "dur": 235.179,
    "args": {
      "External id": 2489749,"Record function id": 0, "Concrete Inputs": ["", "0.5"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11308
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "_c10d_functional::wait_tensor", "pid": 1336759, "tid": 1336759,
    "ts": 1514658035553.606, "dur": 40.892,
    "args": {
      "External id": 2489750,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 11309
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "record_param_comms", "pid": 1336759, "tid": 1336759,
    "ts": 1514658035573.706, "dur": 10.557,
    "args": {
      "External id": 2489751,"Record function id": 0, "Collective name": "wait", "Process Group Description": "default_pg", "dtype": "Byte", "Input Dims": [[], [], [], [], [], [], [], [], []], "Input Strides": [[], [], [], [], [], [], [], [], []], "Input type": ["", "", "Scalar", "", "ScalarList", "ScalarList", "Scalar", "Scalar", "Scalar"], "Concrete Inputs": ["", "", "6", "", "[]", "[]", "-1", "-1", "1"], "Out msg nelems": 0, "Process Group Name": "0", "Process Group Ranks": "[]", "Group size": 1, "Ev Idx": 11310, "In msg nelems": 0, "Rank": 6, "In split size": "[]", "Out split size": "[]"
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::pow", "pid": 1336759, "tid": 1336759,
    "ts": 1514658035626.417, "dur": 48.780,
    "args": {
      "External id": 2489752,"Record function id": 0, "Concrete Inputs": ["", "0.5"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11311
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658035629.504, "dur": 1.318,
    "args": {
      "External id": 2489753,"Record function id": 0, "Concrete Inputs": ["", "0.5"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11312
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658035632.326, "dur": 0.410,
    "args": {
      "External id": 2489754,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 11313
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "_ToTorchTensor", "pid": 1336759, "tid": 1336759,
    "ts": 1514658035736.190, "dur": 19.333,
    "args": {
      "External id": 2489755,"Sequence number": 24740657, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", ""], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11314
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view_as", "pid": 1336759, "tid": 1336759,
    "ts": 1514658035745.124, "dur": 7.432,
    "args": {
      "External id": 2489756,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11315
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::view", "pid": 1336759, "tid": 1336759,
    "ts": 1514658035747.180, "dur": 5.217,
    "args": {
      "External id": 2489757,"Record function id": 0, "Concrete Inputs": ["", "[]"], "Input type": ["float", "ScalarList"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11316
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::add", "pid": 1336759, "tid": 1336759,
    "ts": 1514658036107.088, "dur": 41.271,
    "args": {
      "External id": 2489758,"Record function id": 0, "Concrete Inputs": ["", "", "1"], "Input type": ["float", "double", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 11317
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::reciprocal", "pid": 1336759, "tid": 1336759,
    "ts": 1514658036159.556, "dur": 22.665,
    "args": {
      "External id": 2489759,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 11318
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::mul", "pid": 1336759, "tid": 1336759,
    "ts": 1514658036190.274, "dur": 25.084,
    "args": {
      "External id": 2489760,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "double"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11319
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::clamp", "pid": 1336759, "tid": 1336759,
    "ts": 1514658036230.527, "dur": 27.043,
    "args": {
      "External id": 2489761,"Record function id": 0, "Concrete Inputs": ["", "", "1."], "Input type": ["float", "", "Scalar"], "Input Strides": [[], [], []], "Input Dims": [[], [], []], "Ev Idx": 11320
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658036234.954, "dur": 0.330,
    "args": {
      "External id": 2489762,"Record function id": 0, "Concrete Inputs": ["", "6", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], []], "Input Dims": [[], [], [], [], []], "Ev Idx": 11321
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::to", "pid": 1336759, "tid": 1336759,
    "ts": 1514658036275.590, "dur": 0.448,
    "args": {
      "External id": 2489763,"Record function id": 0, "Concrete Inputs": ["", "6", "0", "", "", "False", "False", ""], "Input type": ["float", "Scalar", "Scalar", "", "", "Scalar", "Scalar", ""], "Input Strides": [[], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], []], "Ev Idx": 11322
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_mul_", "pid": 1336759, "tid": 1336759,
    "ts": 1514658036384.932, "dur": 730.416,
    "args": {
      "External id": 2489764,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["TensorList", "float"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11323
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_mul_", "pid": 1336759, "tid": 1336759,
    "ts": 1514658036895.918, "dur": 186.401,
    "args": {
      "External id": 2489765,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["TensorList", "float"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11324
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::isnan", "pid": 1336759, "tid": 1336759,
    "ts": 1514658037164.787, "dur": 30.292,
    "args": {
      "External id": 2489766,"Sequence number": 24740658, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 11325
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::ne", "pid": 1336759, "tid": 1336759,
    "ts": 1514658037169.362, "dur": 25.130,
    "args": {
      "External id": 2489767,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11326
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::is_nonzero", "pid": 1336759, "tid": 1336759,
    "ts": 1514658037199.859, "dur": 1916.465,
    "args": {
      "External id": 2489768,"Sequence number": 24740658, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 11327
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1336759, "tid": 1336759,
    "ts": 1514658037201.479, "dur": 1914.616,
    "args": {
      "External id": 2489769,"Sequence number": 24740658, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 11328
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1336759, "tid": 1336759,
    "ts": 1514658037203.274, "dur": 1911.780,
    "args": {
      "External id": 2489770,"Sequence number": 24740658, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 11329
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::isinf", "pid": 1336759, "tid": 1336759,
    "ts": 1514658039122.743, "dur": 62.492,
    "args": {
      "External id": 2489771,"Sequence number": 24740658, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 11330
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514658039126.156, "dur": 35.976,
    "args": {
      "External id": 2489772,"Record function id": 0, "Concrete Inputs": [""], "Input type": ["float"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 11331
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::empty", "pid": 1336759, "tid": 1336759,
    "ts": 1514658039133.693, "dur": 3.465,
    "args": {
      "External id": 2489773,"Record function id": 0, "Concrete Inputs": ["[0]", "6", "0", "", "", ""], "Input type": ["ScalarList", "Scalar", "Scalar", "", "", ""], "Input Strides": [[], [], [], [], [], []], "Input Dims": [[], [], [], [], [], []], "Ev Idx": 11332
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::abs", "pid": 1336759, "tid": 1336759,
    "ts": 1514658039139.254, "dur": 22.478,
    "args": {
      "External id": 2489774,"Record function id": 0, "Concrete Inputs": ["", ""], "Input type": ["float", "float"], "Input Strides": [[], [1]], "Input Dims": [[], [0]], "Ev Idx": 11333
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::resize_", "pid": 1336759, "tid": 1336759,
    "ts": 1514658039144.396, "dur": 2.925,
    "args": {
      "External id": 2489775,"Record function id": 0, "Concrete Inputs": ["", "[]", ""], "Input type": ["float", "ScalarList", ""], "Input Strides": [[1], [], []], "Input Dims": [[0], [], []], "Ev Idx": 11334
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::eq", "pid": 1336759, "tid": 1336759,
    "ts": 1514658039164.023, "dur": 20.146,
    "args": {
      "External id": 2489776,"Record function id": 0, "Concrete Inputs": ["", "inf"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11335
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::is_nonzero", "pid": 1336759, "tid": 1336759,
    "ts": 1514658039187.495, "dur": 35.624,
    "args": {
      "External id": 2489777,"Sequence number": 24740658, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 11336
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::item", "pid": 1336759, "tid": 1336759,
    "ts": 1514658039190.159, "dur": 32.772,
    "args": {
      "External id": 2489778,"Sequence number": 24740658, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 11337
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_local_scalar_dense", "pid": 1336759, "tid": 1336759,
    "ts": 1514658039191.208, "dur": 31.339,
    "args": {
      "External id": 2489779,"Sequence number": 24740658, "Fwd thread id": 0, "Record function id": 0, "Concrete Inputs": [""], "Input type": ["bool"], "Input Strides": [[]], "Input Dims": [[]], "Ev Idx": 11338
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "Optimizer.step#OptimizersContainer.step", "pid": 1336759, "tid": 1336759,
    "ts": 1514658039257.289, "dur": 5434.111,
    "args": {
      "External id": 2489780,"Record function id": 0, "Ev Idx": 11339
    }
  },
  {
    "ph": "X", "cat": "user_annotation", "name": "Optimizer.step#AdamW.step", "pid": 1336759, "tid": 1336759,
    "ts": 1514658039288.369, "dur": 5381.650,
    "args": {
      "External id": 2489781,"Record function id": 0, "Ev Idx": 11340
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_foreach_add_", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040463.891, "dur": 278.827,
    "args": {
      "External id": 2489782,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["TensorList", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11341
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040485.764, "dur": 1.169,
    "args": {
      "External id": 2489783,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11342
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040488.022, "dur": 0.335,
    "args": {
      "External id": 2489784,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11343
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040488.800, "dur": 0.206,
    "args": {
      "External id": 2489785,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11344
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040489.353, "dur": 0.228,
    "args": {
      "External id": 2489786,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11345
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040489.955, "dur": 0.162,
    "args": {
      "External id": 2489787,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11346
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040490.555, "dur": 0.204,
    "args": {
      "External id": 2489788,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11347
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040491.328, "dur": 0.081,
    "args": {
      "External id": 2489789,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11348
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040491.857, "dur": 0.182,
    "args": {
      "External id": 2489790,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11349
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040492.424, "dur": 0.210,
    "args": {
      "External id": 2489791,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11350
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040492.924, "dur": 0.210,
    "args": {
      "External id": 2489792,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11351
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040493.470, "dur": 0.073,
    "args": {
      "External id": 2489793,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11352
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040493.991, "dur": 0.064,
    "args": {
      "External id": 2489794,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11353
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040494.410, "dur": 0.064,
    "args": {
      "External id": 2489795,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11354
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040494.848, "dur": 0.049,
    "args": {
      "External id": 2489796,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11355
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040495.230, "dur": 0.062,
    "args": {
      "External id": 2489797,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11356
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040495.555, "dur": 0.062,
    "args": {
      "External id": 2489798,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11357
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040496.189, "dur": 0.068,
    "args": {
      "External id": 2489799,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11358
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040496.669, "dur": 0.051,
    "args": {
      "External id": 2489800,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11359
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040497.250, "dur": 0.092,
    "args": {
      "External id": 2489801,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11360
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040497.600, "dur": 0.080,
    "args": {
      "External id": 2489802,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11361
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040498.063, "dur": 0.060,
    "args": {
      "External id": 2489803,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11362
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040498.457, "dur": 0.063,
    "args": {
      "External id": 2489804,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11363
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040499.070, "dur": 0.060,
    "args": {
      "External id": 2489805,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11364
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040499.463, "dur": 0.063,
    "args": {
      "External id": 2489806,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11365
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040499.872, "dur": 0.063,
    "args": {
      "External id": 2489807,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11366
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040500.274, "dur": 0.068,
    "args": {
      "External id": 2489808,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11367
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040500.703, "dur": 0.061,
    "args": {
      "External id": 2489809,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11368
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040501.316, "dur": 0.064,
    "args": {
      "External id": 2489810,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11369
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040501.753, "dur": 0.067,
    "args": {
      "External id": 2489811,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11370
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040502.142, "dur": 0.067,
    "args": {
      "External id": 2489812,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11371
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040502.523, "dur": 0.065,
    "args": {
      "External id": 2489813,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11372
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040502.964, "dur": 0.065,
    "args": {
      "External id": 2489814,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11373
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040503.401, "dur": 0.064,
    "args": {
      "External id": 2489815,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11374
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040503.819, "dur": 0.052,
    "args": {
      "External id": 2489816,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11375
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040504.183, "dur": 0.066,
    "args": {
      "External id": 2489817,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11376
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040504.530, "dur": 0.062,
    "args": {
      "External id": 2489818,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11377
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040504.918, "dur": 0.065,
    "args": {
      "External id": 2489819,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11378
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040505.356, "dur": 0.067,
    "args": {
      "External id": 2489820,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11379
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040505.956, "dur": 0.059,
    "args": {
      "External id": 2489821,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11380
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040506.482, "dur": 0.067,
    "args": {
      "External id": 2489822,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11381
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040506.933, "dur": 0.062,
    "args": {
      "External id": 2489823,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11382
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040507.478, "dur": 0.067,
    "args": {
      "External id": 2489824,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11383
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040507.993, "dur": 0.064,
    "args": {
      "External id": 2489825,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11384
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040508.477, "dur": 0.065,
    "args": {
      "External id": 2489826,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11385
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040508.875, "dur": 0.066,
    "args": {
      "External id": 2489827,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11386
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040509.407, "dur": 0.059,
    "args": {
      "External id": 2489828,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11387
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040509.784, "dur": 0.052,
    "args": {
      "External id": 2489829,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11388
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040510.142, "dur": 0.061,
    "args": {
      "External id": 2489830,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11389
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040510.519, "dur": 0.067,
    "args": {
      "External id": 2489831,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11390
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040511.130, "dur": 0.054,
    "args": {
      "External id": 2489832,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11391
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040511.693, "dur": 0.063,
    "args": {
      "External id": 2489833,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11392
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040512.106, "dur": 0.063,
    "args": {
      "External id": 2489834,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11393
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040512.531, "dur": 0.063,
    "args": {
      "External id": 2489835,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11394
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040512.983, "dur": 0.068,
    "args": {
      "External id": 2489836,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11395
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040513.395, "dur": 0.060,
    "args": {
      "External id": 2489837,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11396
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040513.771, "dur": 0.063,
    "args": {
      "External id": 2489838,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11397
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040514.105, "dur": 0.064,
    "args": {
      "External id": 2489839,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11398
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040514.602, "dur": 0.066,
    "args": {
      "External id": 2489840,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11399
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040514.982, "dur": 0.060,
    "args": {
      "External id": 2489841,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11400
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040515.330, "dur": 0.067,
    "args": {
      "External id": 2489842,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11401
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040515.803, "dur": 0.063,
    "args": {
      "External id": 2489843,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11402
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040516.214, "dur": 0.062,
    "args": {
      "External id": 2489844,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11403
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040516.547, "dur": 0.062,
    "args": {
      "External id": 2489845,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11404
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040516.877, "dur": 0.065,
    "args": {
      "External id": 2489846,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11405
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040517.309, "dur": 0.067,
    "args": {
      "External id": 2489847,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11406
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040517.718, "dur": 0.067,
    "args": {
      "External id": 2489848,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11407
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040518.137, "dur": 0.069,
    "args": {
      "External id": 2489849,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11408
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040518.525, "dur": 0.072,
    "args": {
      "External id": 2489850,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11409
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040518.936, "dur": 0.070,
    "args": {
      "External id": 2489851,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11410
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040519.364, "dur": 0.064,
    "args": {
      "External id": 2489852,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11411
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040519.962, "dur": 0.062,
    "args": {
      "External id": 2489853,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11412
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040520.329, "dur": 0.197,
    "args": {
      "External id": 2489854,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11413
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040520.822, "dur": 0.207,
    "args": {
      "External id": 2489855,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11414
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040521.344, "dur": 0.213,
    "args": {
      "External id": 2489856,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11415
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040555.917, "dur": 0.398,
    "args": {
      "External id": 2489857,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11416
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040557.693, "dur": 0.054,
    "args": {
      "External id": 2489858,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11417
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040558.150, "dur": 0.053,
    "args": {
      "External id": 2489859,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11418
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040558.587, "dur": 0.061,
    "args": {
      "External id": 2489860,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11419
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040558.995, "dur": 0.066,
    "args": {
      "External id": 2489861,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11420
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040559.390, "dur": 0.064,
    "args": {
      "External id": 2489862,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11421
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040559.803, "dur": 0.071,
    "args": {
      "External id": 2489863,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11422
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040560.180, "dur": 0.065,
    "args": {
      "External id": 2489864,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11423
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040560.553, "dur": 0.064,
    "args": {
      "External id": 2489865,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11424
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040560.951, "dur": 0.052,
    "args": {
      "External id": 2489866,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11425
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040561.302, "dur": 0.067,
    "args": {
      "External id": 2489867,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11426
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040561.720, "dur": 0.067,
    "args": {
      "External id": 2489868,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11427
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040562.099, "dur": 0.065,
    "args": {
      "External id": 2489869,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11428
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040562.490, "dur": 0.066,
    "args": {
      "External id": 2489870,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11429
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040562.921, "dur": 0.065,
    "args": {
      "External id": 2489871,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11430
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040563.301, "dur": 0.063,
    "args": {
      "External id": 2489872,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11431
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040563.677, "dur": 0.067,
    "args": {
      "External id": 2489873,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11432
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040564.086, "dur": 0.067,
    "args": {
      "External id": 2489874,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11433
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040564.525, "dur": 0.066,
    "args": {
      "External id": 2489875,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11434
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040564.887, "dur": 0.064,
    "args": {
      "External id": 2489876,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11435
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040565.252, "dur": 0.066,
    "args": {
      "External id": 2489877,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11436
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040565.693, "dur": 0.062,
    "args": {
      "External id": 2489878,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11437
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040566.157, "dur": 0.063,
    "args": {
      "External id": 2489879,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11438
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040566.565, "dur": 0.060,
    "args": {
      "External id": 2489880,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11439
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040566.964, "dur": 0.062,
    "args": {
      "External id": 2489881,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11440
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040567.348, "dur": 0.068,
    "args": {
      "External id": 2489882,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11441
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040567.728, "dur": 0.072,
    "args": {
      "External id": 2489883,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11442
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040568.080, "dur": 0.061,
    "args": {
      "External id": 2489884,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11443
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040568.520, "dur": 0.064,
    "args": {
      "External id": 2489885,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11444
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040568.934, "dur": 0.066,
    "args": {
      "External id": 2489886,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11445
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040569.299, "dur": 0.066,
    "args": {
      "External id": 2489887,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11446
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040569.715, "dur": 0.066,
    "args": {
      "External id": 2489888,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11447
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040570.092, "dur": 0.059,
    "args": {
      "External id": 2489889,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11448
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040570.494, "dur": 0.059,
    "args": {
      "External id": 2489890,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11449
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040570.880, "dur": 0.093,
    "args": {
      "External id": 2489891,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11450
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040571.302, "dur": 0.068,
    "args": {
      "External id": 2489892,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11451
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040571.688, "dur": 0.060,
    "args": {
      "External id": 2489893,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11452
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040572.070, "dur": 0.066,
    "args": {
      "External id": 2489894,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11453
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040572.484, "dur": 0.069,
    "args": {
      "External id": 2489895,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11454
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040572.869, "dur": 0.064,
    "args": {
      "External id": 2489896,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11455
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040573.247, "dur": 0.063,
    "args": {
      "External id": 2489897,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11456
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040573.666, "dur": 0.066,
    "args": {
      "External id": 2489898,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11457
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040574.085, "dur": 0.063,
    "args": {
      "External id": 2489899,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11458
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040574.476, "dur": 0.068,
    "args": {
      "External id": 2489900,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11459
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040574.908, "dur": 0.066,
    "args": {
      "External id": 2489901,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11460
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040575.310, "dur": 0.056,
    "args": {
      "External id": 2489902,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11461
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040575.636, "dur": 0.054,
    "args": {
      "External id": 2489903,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11462
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040576.013, "dur": 0.059,
    "args": {
      "External id": 2489904,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11463
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040576.428, "dur": 0.065,
    "args": {
      "External id": 2489905,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11464
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040576.980, "dur": 0.066,
    "args": {
      "External id": 2489906,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11465
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040577.374, "dur": 0.060,
    "args": {
      "External id": 2489907,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11466
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040577.762, "dur": 0.086,
    "args": {
      "External id": 2489908,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11467
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040578.224, "dur": 0.062,
    "args": {
      "External id": 2489909,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11468
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040585.677, "dur": 0.093,
    "args": {
      "External id": 2489910,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11469
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040587.073, "dur": 0.059,
    "args": {
      "External id": 2489911,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11470
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040587.461, "dur": 0.084,
    "args": {
      "External id": 2489912,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11471
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040587.921, "dur": 0.083,
    "args": {
      "External id": 2489913,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11472
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040588.264, "dur": 0.063,
    "args": {
      "External id": 2489914,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11473
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040588.682, "dur": 0.063,
    "args": {
      "External id": 2489915,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11474
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040589.099, "dur": 0.064,
    "args": {
      "External id": 2489916,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11475
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040589.520, "dur": 0.059,
    "args": {
      "External id": 2489917,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11476
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040589.844, "dur": 0.060,
    "args": {
      "External id": 2489918,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11477
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040590.272, "dur": 0.060,
    "args": {
      "External id": 2489919,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11478
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040590.871, "dur": 0.065,
    "args": {
      "External id": 2489920,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11479
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040591.260, "dur": 0.061,
    "args": {
      "External id": 2489921,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11480
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040591.872, "dur": 0.064,
    "args": {
      "External id": 2489922,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11481
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040592.412, "dur": 0.060,
    "args": {
      "External id": 2489923,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11482
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040592.765, "dur": 0.065,
    "args": {
      "External id": 2489924,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11483
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040593.204, "dur": 0.061,
    "args": {
      "External id": 2489925,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11484
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040593.583, "dur": 0.063,
    "args": {
      "External id": 2489926,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11485
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040594.193, "dur": 0.066,
    "args": {
      "External id": 2489927,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11486
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040594.656, "dur": 0.055,
    "args": {
      "External id": 2489928,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11487
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040595.178, "dur": 0.066,
    "args": {
      "External id": 2489929,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11488
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040595.598, "dur": 0.064,
    "args": {
      "External id": 2489930,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11489
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040596.014, "dur": 0.065,
    "args": {
      "External id": 2489931,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11490
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040596.379, "dur": 0.061,
    "args": {
      "External id": 2489932,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11491
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040596.792, "dur": 0.066,
    "args": {
      "External id": 2489933,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11492
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040597.199, "dur": 0.065,
    "args": {
      "External id": 2489934,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11493
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040597.548, "dur": 0.062,
    "args": {
      "External id": 2489935,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11494
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040597.954, "dur": 0.059,
    "args": {
      "External id": 2489936,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11495
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040598.361, "dur": 0.066,
    "args": {
      "External id": 2489937,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11496
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040598.770, "dur": 0.067,
    "args": {
      "External id": 2489938,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11497
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040599.169, "dur": 0.063,
    "args": {
      "External id": 2489939,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11498
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040599.569, "dur": 0.065,
    "args": {
      "External id": 2489940,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11499
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040599.967, "dur": 0.061,
    "args": {
      "External id": 2489941,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11500
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040600.551, "dur": 0.066,
    "args": {
      "External id": 2489942,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11501
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040600.982, "dur": 0.066,
    "args": {
      "External id": 2489943,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11502
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040601.541, "dur": 0.063,
    "args": {
      "External id": 2489944,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11503
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040602.018, "dur": 0.050,
    "args": {
      "External id": 2489945,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11504
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040602.416, "dur": 0.068,
    "args": {
      "External id": 2489946,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11505
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040602.956, "dur": 0.055,
    "args": {
      "External id": 2489947,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11506
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040603.263, "dur": 0.065,
    "args": {
      "External id": 2489948,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11507
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040603.893, "dur": 0.063,
    "args": {
      "External id": 2489949,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11508
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040604.271, "dur": 0.061,
    "args": {
      "External id": 2489950,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11509
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040604.669, "dur": 0.065,
    "args": {
      "External id": 2489951,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11510
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040605.058, "dur": 0.059,
    "args": {
      "External id": 2489952,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11511
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040605.454, "dur": 0.062,
    "args": {
      "External id": 2489953,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11512
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040605.822, "dur": 0.063,
    "args": {
      "External id": 2489954,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11513
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040606.311, "dur": 0.058,
    "args": {
      "External id": 2489955,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11514
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040606.694, "dur": 0.061,
    "args": {
      "External id": 2489956,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11515
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040607.019, "dur": 0.066,
    "args": {
      "External id": 2489957,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11516
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040607.405, "dur": 0.066,
    "args": {
      "External id": 2489958,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11517
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040607.955, "dur": 0.066,
    "args": {
      "External id": 2489959,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11518
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040608.377, "dur": 0.060,
    "args": {
      "External id": 2489960,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11519
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040608.754, "dur": 0.060,
    "args": {
      "External id": 2489961,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11520
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040609.187, "dur": 0.063,
    "args": {
      "External id": 2489962,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11521
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040609.671, "dur": 0.067,
    "args": {
      "External id": 2489963,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11522
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040610.146, "dur": 0.052,
    "args": {
      "External id": 2489964,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11523
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040610.584, "dur": 0.061,
    "args": {
      "External id": 2489965,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11524
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040610.985, "dur": 0.063,
    "args": {
      "External id": 2489966,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11525
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040611.341, "dur": 0.064,
    "args": {
      "External id": 2489967,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11526
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040611.757, "dur": 0.062,
    "args": {
      "External id": 2489968,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11527
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040612.156, "dur": 0.061,
    "args": {
      "External id": 2489969,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11528
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040612.719, "dur": 0.064,
    "args": {
      "External id": 2489970,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11529
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040613.109, "dur": 0.063,
    "args": {
      "External id": 2489971,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11530
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040613.548, "dur": 0.065,
    "args": {
      "External id": 2489972,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11531
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040613.933, "dur": 0.061,
    "args": {
      "External id": 2489973,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11532
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040614.361, "dur": 0.065,
    "args": {
      "External id": 2489974,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11533
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040614.962, "dur": 0.066,
    "args": {
      "External id": 2489975,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11534
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040615.363, "dur": 0.060,
    "args": {
      "External id": 2489976,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11535
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040615.819, "dur": 0.049,
    "args": {
      "External id": 2489977,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11536
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040616.180, "dur": 0.065,
    "args": {
      "External id": 2489978,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11537
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040616.603, "dur": 0.051,
    "args": {
      "External id": 2489979,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11538
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040616.970, "dur": 0.061,
    "args": {
      "External id": 2489980,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11539
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040617.382, "dur": 0.062,
    "args": {
      "External id": 2489981,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11540
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040617.705, "dur": 0.067,
    "args": {
      "External id": 2489982,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11541
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040618.164, "dur": 0.060,
    "args": {
      "External id": 2489983,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11542
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040618.565, "dur": 0.064,
    "args": {
      "External id": 2489984,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11543
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040618.961, "dur": 0.067,
    "args": {
      "External id": 2489985,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11544
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040619.466, "dur": 0.067,
    "args": {
      "External id": 2489986,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11545
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040619.831, "dur": 0.066,
    "args": {
      "External id": 2489987,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11546
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040620.233, "dur": 0.063,
    "args": {
      "External id": 2489988,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11547
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040620.604, "dur": 0.066,
    "args": {
      "External id": 2489989,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11548
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040620.937, "dur": 0.060,
    "args": {
      "External id": 2489990,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11549
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040621.275, "dur": 0.063,
    "args": {
      "External id": 2489991,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11550
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040621.904, "dur": 0.060,
    "args": {
      "External id": 2489992,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11551
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040622.339, "dur": 0.047,
    "args": {
      "External id": 2489993,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11552
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040622.888, "dur": 0.062,
    "args": {
      "External id": 2489994,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11553
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040623.217, "dur": 0.049,
    "args": {
      "External id": 2489995,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11554
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040623.878, "dur": 0.070,
    "args": {
      "External id": 2489996,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11555
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040624.205, "dur": 0.069,
    "args": {
      "External id": 2489997,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11556
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040624.885, "dur": 0.064,
    "args": {
      "External id": 2489998,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11557
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040625.213, "dur": 0.049,
    "args": {
      "External id": 2489999,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11558
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040625.730, "dur": 0.064,
    "args": {
      "External id": 2490000,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11559
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040626.050, "dur": 0.053,
    "args": {
      "External id": 2490001,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11560
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040626.825, "dur": 0.065,
    "args": {
      "External id": 2490002,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11561
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040627.167, "dur": 0.049,
    "args": {
      "External id": 2490003,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11562
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040627.834, "dur": 0.065,
    "args": {
      "External id": 2490004,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11563
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040628.159, "dur": 0.052,
    "args": {
      "External id": 2490005,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11564
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040628.672, "dur": 0.064,
    "args": {
      "External id": 2490006,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11565
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040628.992, "dur": 0.051,
    "args": {
      "External id": 2490007,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11566
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040629.663, "dur": 0.065,
    "args": {
      "External id": 2490008,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11567
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040630.450, "dur": 0.067,
    "args": {
      "External id": 2490009,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11568
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040631.077, "dur": 0.053,
    "args": {
      "External id": 2490010,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11569
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040631.411, "dur": 0.052,
    "args": {
      "External id": 2490011,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11570
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040632.083, "dur": 0.067,
    "args": {
      "External id": 2490012,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11571
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040632.405, "dur": 0.050,
    "args": {
      "External id": 2490013,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11572
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040632.976, "dur": 0.064,
    "args": {
      "External id": 2490014,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11573
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040633.297, "dur": 0.053,
    "args": {
      "External id": 2490015,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11574
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040633.848, "dur": 0.062,
    "args": {
      "External id": 2490016,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11575
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040634.171, "dur": 0.050,
    "args": {
      "External id": 2490017,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11576
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040634.799, "dur": 0.066,
    "args": {
      "External id": 2490018,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11577
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040635.126, "dur": 0.051,
    "args": {
      "External id": 2490019,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11578
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040635.637, "dur": 0.064,
    "args": {
      "External id": 2490020,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11579
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040635.965, "dur": 0.049,
    "args": {
      "External id": 2490021,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11580
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040636.561, "dur": 0.048,
    "args": {
      "External id": 2490022,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11581
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040636.873, "dur": 0.049,
    "args": {
      "External id": 2490023,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11582
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040637.353, "dur": 0.063,
    "args": {
      "External id": 2490024,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11583
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040637.677, "dur": 0.051,
    "args": {
      "External id": 2490025,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11584
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040638.320, "dur": 0.061,
    "args": {
      "External id": 2490026,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11585
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040638.643, "dur": 0.049,
    "args": {
      "External id": 2490027,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11586
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040639.454, "dur": 0.064,
    "args": {
      "External id": 2490028,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11587
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040639.776, "dur": 0.047,
    "args": {
      "External id": 2490029,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11588
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040640.305, "dur": 0.065,
    "args": {
      "External id": 2490030,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11589
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040640.629, "dur": 0.052,
    "args": {
      "External id": 2490031,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11590
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040641.182, "dur": 0.063,
    "args": {
      "External id": 2490032,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11591
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040641.504, "dur": 0.052,
    "args": {
      "External id": 2490033,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11592
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040642.143, "dur": 0.064,
    "args": {
      "External id": 2490034,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11593
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040642.467, "dur": 0.052,
    "args": {
      "External id": 2490035,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11594
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040643.214, "dur": 0.064,
    "args": {
      "External id": 2490036,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11595
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040643.541, "dur": 0.052,
    "args": {
      "External id": 2490037,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11596
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040644.022, "dur": 0.070,
    "args": {
      "External id": 2490038,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11597
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040644.348, "dur": 0.049,
    "args": {
      "External id": 2490039,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11598
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040644.890, "dur": 0.065,
    "args": {
      "External id": 2490040,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11599
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040645.225, "dur": 0.054,
    "args": {
      "External id": 2490041,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11600
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040645.821, "dur": 0.064,
    "args": {
      "External id": 2490042,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11601
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040646.141, "dur": 0.055,
    "args": {
      "External id": 2490043,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11602
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040646.835, "dur": 0.062,
    "args": {
      "External id": 2490044,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11603
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040647.155, "dur": 0.051,
    "args": {
      "External id": 2490045,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11604
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040647.676, "dur": 0.068,
    "args": {
      "External id": 2490046,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11605
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040648.001, "dur": 0.050,
    "args": {
      "External id": 2490047,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11606
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040648.549, "dur": 0.067,
    "args": {
      "External id": 2490048,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11607
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040648.872, "dur": 0.056,
    "args": {
      "External id": 2490049,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11608
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040649.609, "dur": 0.062,
    "args": {
      "External id": 2490050,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11609
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040649.933, "dur": 0.053,
    "args": {
      "External id": 2490051,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11610
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040650.440, "dur": 0.066,
    "args": {
      "External id": 2490052,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11611
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040650.765, "dur": 0.058,
    "args": {
      "External id": 2490053,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11612
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040651.438, "dur": 0.070,
    "args": {
      "External id": 2490054,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11613
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040651.767, "dur": 0.049,
    "args": {
      "External id": 2490055,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11614
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040652.385, "dur": 0.063,
    "args": {
      "External id": 2490056,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11615
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040652.708, "dur": 0.054,
    "args": {
      "External id": 2490057,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11616
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040653.233, "dur": 0.063,
    "args": {
      "External id": 2490058,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11617
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040653.559, "dur": 0.080,
    "args": {
      "External id": 2490059,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11618
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040654.207, "dur": 0.062,
    "args": {
      "External id": 2490060,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11619
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040654.526, "dur": 0.050,
    "args": {
      "External id": 2490061,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11620
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040655.062, "dur": 0.063,
    "args": {
      "External id": 2490062,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11621
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040655.389, "dur": 0.052,
    "args": {
      "External id": 2490063,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11622
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040656.195, "dur": 0.066,
    "args": {
      "External id": 2490064,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11623
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040656.516, "dur": 0.054,
    "args": {
      "External id": 2490065,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11624
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040657.111, "dur": 0.063,
    "args": {
      "External id": 2490066,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11625
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040657.434, "dur": 0.054,
    "args": {
      "External id": 2490067,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11626
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040658.051, "dur": 0.067,
    "args": {
      "External id": 2490068,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11627
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040658.375, "dur": 0.050,
    "args": {
      "External id": 2490069,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11628
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040658.908, "dur": 0.065,
    "args": {
      "External id": 2490070,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11629
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040659.236, "dur": 0.053,
    "args": {
      "External id": 2490071,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11630
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040659.738, "dur": 0.064,
    "args": {
      "External id": 2490072,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11631
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::result_type", "pid": 1336759, "tid": 1336759,
    "ts": 1514658040660.063, "dur": 0.050,
    "args": {
      "External id": 2490073,"Record function id": 0, "Concrete Inputs": ["", "1"], "Input type": ["float", "Scalar"], "Input Strides": [[], []], "Input Dims": [[], []], "Ev Idx": 11632
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_fused_adamw_", "pid": 1336759, "tid": 1336759,
    "ts": 1514658041248.074, "dur": 3329.904,
    "args": {
      "External id": 2490074,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "0.0001072708166170532", "0.90000000000000002", "0.94999999999999996", "0.10000000000000001", "1.0000000000000001e-15", "False", "False", "", ""], "Input type": ["TensorList", "TensorList", "TensorList", "TensorList", "TensorList", "TensorList", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 11633
    }
  },
  {
    "ph": "X", "cat": "cpu_op", "name": "aten::_fused_adamw_", "pid": 1336759, "tid": 1336759,
    "ts": 1514658044085.120, "dur": 330.957,
    "args": {
      "External id": 2490075,"Record function id": 0, "Concrete Inputs": ["", "", "", "", "", "", "0.0001072708166170532", "0.90000000000000002", "0.94999999999999996", "0.10000000000000001", "1.0000000000000001e-15", "False", "False", "", ""], "Input type": ["TensorList", "TensorList", "TensorList", "TensorList", "TensorList", "TensorList", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "Scalar", "", ""], "Input Strides": [[], [], [], [], [], [], [], [], [], [], [], [], [], [], []], "Input Dims": [[], [], [], [], [], [], [], [], [], [], [], [], [], [], []], "Ev Idx": 11634
    }
  },
  {
    "name": "process_name", "ph": "M", "ts": 1514655936581.954, "pid": 1336759, "tid": 0,
    "args": {
      "name": "python3.12"
    }
  },
  {
    "name": "process_labels", "ph": "M", "ts": 1514655936581.954, "pid": 1336759, "tid": 0,
    "args": {
      "labels": "CPU"
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1514655936581.954, "pid": 1336759, "tid": 0,
    "args": {
      "sort_index": 1336759
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 1514655936581.954, "pid": 1336759, "tid": 1381173,
    "args": {
      "name": "thread 1381173 (pt_autograd_6)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 1514655936581.954, "pid": 1336759, "tid": 1381173,
    "args": {
      "sort_index": 1381173
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 1514655936581.954, "pid": 1336759, "tid": 1336759,
    "args": {
      "name": "thread 1336759 (python3.12)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 1514655936581.954, "pid": 1336759, "tid": 1336759,
    "args": {
      "sort_index": 1336759
    }
  },
  {
    "name": "thread_name", "ph": "M", "ts": 1514655936581.954, "pid": 1336759, "tid": 1381173,
    "args": {
      "name": "thread 1381173 (python3.12)"
    }
  },
  {
    "name": "thread_sort_index", "ph": "M", "ts": 1514655936581.954, "pid": 1336759, "tid": 1381173,
    "args": {
      "sort_index": 1381173
    }
  },
  {
    "ph": "X", "cat": "Trace", "ts": 1514655936499.216, "dur": 2109848.414,
    "pid": "Spans", "tid": "PyTorch Profiler",
    "name": "PyTorch Profiler (0)",
    "args": {
      "Op count": 0
    }
  },
  {
    "name": "process_sort_index", "ph": "M", "ts": 1514655936499.216,
    "pid": "Spans", "tid": 0,
    "args": {
      "sort_index": 536870912
    }
  },
  {
    "name": "Iteration Start: PyTorch Profiler", "ph": "i", "s": "g",
    "pid": "Traces", "tid": "Trace PyTorch Profiler", "ts": 1514655936499.216
  },
  {
    "name": "Record Window End", "ph": "i", "s": "g",
    "pid": "", "tid": "", "ts": 1514658108013.864
  }
  ],
  "traceName": "exp/mtp.1B.batch16.seqlen4096.context4096.warmup2000.update1.steps200000.lr2e-4.cosine/profile_trace/iteration_102912/rank6_trace.json",
  "displayTimeUnit": "ms",
  "baseTimeNanoseconds": 1751410836000000000
}